oxides_calcine.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Inorganic Material Generation using Deep Q-Networks\n",
    "https://github.com/eltonpan/InorganicMaterialRL\n",
    "\n",
    "Elton Pan"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "/home/synthesisproject/anaconda3/envs/dqn_roost_EMD/lib/python3.7/site-packages/sklearn/base.py:315: UserWarning: Trying to unpickle estimator DecisionTreeRegressor from version 0.23.2 when using version 0.24.2. This might lead to breaking code or invalid results. Use at your own risk.\n",
      "  UserWarning)\n",
      "/home/synthesisproject/anaconda3/envs/dqn_roost_EMD/lib/python3.7/site-packages/sklearn/base.py:315: UserWarning: Trying to unpickle estimator RandomForestRegressor from version 0.23.2 when using version 0.24.2. This might lead to breaking code or invalid results. Use at your own risk.\n",
      "  UserWarning)\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "True\n"
     ]
    }
   ],
   "source": [
    "# Env: dqn_roost as of 15/1/2022 (after addition of ROOST)\n",
    "import numpy as np\n",
    "import pickle5 as pickle\n",
    "import torch\n",
    "from torch import nn\n",
    "import collections\n",
    "from model import DQN_pytorch, DCN_pytorch\n",
    "import time\n",
    "from one_hot import featurize_target, onehot_target, element_set, comp_set, one_hot_to_element, element_to_one_hot, one_hot_to_comp, comp_to_one_hot, step_to_one_hot, one_hot_to_step\n",
    "import seaborn as sns\n",
    "import random\n",
    "import matplotlib.pyplot as plt\n",
    "import os\n",
    "from env_constrained import generate_random_act, ConstrainedMaterialEnvironment, extract_data_from_ep, generate_random_ep\n",
    "import collections\n",
    "from sklearn.preprocessing import StandardScaler\n",
    "from tqdm import tqdm \n",
    "from sklearn.manifold import TSNE\n",
    "from numpy.linalg import svd\n",
    "import imageio\n",
    "from pymatgen.core import Composition\n",
    "from constraints.checkers import check_neutrality, check_neutrality, check_neutrality_multiple, check_electronegativity_multiple, check_both_multiple\n",
    "from matminer.featurizers.base import MultipleFeaturizer\n",
    "import matminer.featurizers.composition as cf\n",
    "from sklearn.ensemble import RandomForestRegressor\n",
    "import joblib\n",
    "from roost_models.roost_model import predict_formation_energy, predict_bulk_mod, predict_shear_mod, predict_band_gap\n",
    "import json\n",
    "from configs import configs\n",
    "from utils.metrics import similarity_to_nearest_neighbor\n",
    "os.environ[\"CUDA_VISIBLE_DEVICES\"]=\"1\"\n",
    "print(torch.cuda.is_available())"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Random policy\n",
    "In env.py, we have generated some data using a random policy. Here, we visualize the distribution of final rewards, and then trained an agent using purely data generated by the random policy."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Number of episodes generated by random policy: 1000\n",
      "s_material_inputs: torch.Size([5000, 145])\n",
      "s_step_inputs: torch.Size([5000, 5])\n",
      "a_elem_inputs: torch.Size([5000, 78])\n",
      "a_comp_inputs: torch.Size([5000, 10])\n",
      "Q_targets: torch.Size([5000, 1])\n",
      "en_targets: torch.Size([5000, 1])\n",
      "\n"
     ]
    }
   ],
   "source": [
    "# Load Q_data_random\n",
    "# with open('./data/Q_data_random_RF_constrained.pkl', 'rb') as f:\n",
    "with open('./data/oxides_calcine/Q_data_random.pkl', 'rb') as f:\n",
    "    Q_c_data_random = pickle.load(f)\n",
    "\n",
    "print('Number of episodes generated by random policy:',len(Q_c_data_random))\n",
    "\n",
    "def generate_dqn_dcn_inputs_and_targets_from_Q_c_data(Q_c_data):\n",
    "    '''\n",
    "    Generates inputs and targets for DQN training\n",
    "\n",
    "    Args:\n",
    "    Q_data: List of Q_datapoints for training\n",
    "    \n",
    "    Returns:\n",
    "    s_material_inputs: torch.tensor of Size([num of episodes in Q_data, 145])\n",
    "    s_step_inputs: torch.tensor\n",
    "    a_elem_inputs: torch.tensor\n",
    "    a_comp_inputs: torch.tensor\n",
    "    Q_targets: torch.tensor\n",
    "    en_targets: torch.tensor\n",
    "\n",
    "    '''\n",
    "    s_material_inputs = []\n",
    "    s_step_inputs     = []\n",
    "    a_elem_inputs     = []\n",
    "    a_comp_inputs     = []\n",
    "    Q_targets         = []\n",
    "    en_targets         = []\n",
    "\n",
    "    for episode in Q_c_data:\n",
    "        for step in episode:\n",
    "            # state, action, reward = step\n",
    "            [s_material, s_step], [a_elem, a_comp], Q_target, en_target = step\n",
    "\n",
    "            s_material = torch.tensor(s_material).float()\n",
    "            s_step = torch.tensor(s_step)\n",
    "            a_elem = torch.tensor(a_elem)\n",
    "            a_comp = torch.tensor(a_comp)\n",
    "            Q_target = torch.tensor([Q_target])\n",
    "            en_target = torch.tensor([en_target])\n",
    "\n",
    "            s_material_inputs.append(s_material)\n",
    "            s_step_inputs.append(s_step)\n",
    "            a_elem_inputs.append(a_elem)\n",
    "            a_comp_inputs.append(a_comp)\n",
    "            Q_targets.append(Q_target)\n",
    "            en_targets.append(en_target)\n",
    "\n",
    "    s_material_inputs = torch.stack(s_material_inputs)\n",
    "    s_step_inputs = torch.stack(s_step_inputs)\n",
    "    a_elem_inputs = torch.stack(a_elem_inputs)\n",
    "    a_comp_inputs = torch.stack(a_comp_inputs)\n",
    "    Q_targets = torch.stack(Q_targets)\n",
    "    en_targets = torch.stack(en_targets)\n",
    "\n",
    "    return s_material_inputs, s_step_inputs, a_elem_inputs, a_comp_inputs, Q_targets, en_targets\n",
    "\n",
    "s_material_inputs, s_step_inputs, a_elem_inputs, a_comp_inputs, Q_targets, en_targets = generate_dqn_dcn_inputs_and_targets_from_Q_c_data(Q_c_data = Q_c_data_random)\n",
    "\n",
    "print('s_material_inputs:', s_material_inputs.shape)\n",
    "print('s_step_inputs:', s_step_inputs.shape)\n",
    "print('a_elem_inputs:', a_elem_inputs.shape)\n",
    "print('a_comp_inputs:', a_comp_inputs.shape)\n",
    "print('Q_targets:', Q_targets.shape)\n",
    "print('en_targets:', en_targets.shape)\n",
    "print('')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "/home/synthesisproject/anaconda3/envs/dqn_roost_EMD/lib/python3.7/site-packages/seaborn/distributions.py:2619: FutureWarning: `distplot` is a deprecated function and will be removed in a future version. Please adapt your code to use either `displot` (a figure-level function with similar flexibility) or `histplot` (an axes-level function for histograms).\n",
      "  warnings.warn(msg, FutureWarning)\n"
     ]
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 432x288 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "/home/synthesisproject/anaconda3/envs/dqn_roost_EMD/lib/python3.7/site-packages/seaborn/distributions.py:2619: FutureWarning: `distplot` is a deprecated function and will be removed in a future version. Please adapt your code to use either `displot` (a figure-level function with similar flexibility) or `histplot` (an axes-level function for histograms).\n",
      "  warnings.warn(msg, FutureWarning)\n"
     ]
    },
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAYIAAAEGCAYAAABo25JHAAAAOXRFWHRTb2Z0d2FyZQBNYXRwbG90bGliIHZlcnNpb24zLjQuMiwgaHR0cHM6Ly9tYXRwbG90bGliLm9yZy8rg+JYAAAACXBIWXMAAAsTAAALEwEAmpwYAAA0AUlEQVR4nO3deXQc133g++8PO0CQBLGSIAmQBBdxJ7hJ1GLJkiJRtGzZY8mW7NiW4kTPE2tmnIw9TyfjOE4yc+xJzkuel4xleYnl5cmSZWunrYUjWVwkihu4QCSIhSAAAiAAEiSxUARA/N4fVaBaEJYG2dXV3fX7nNMH3V23qn7oru5f17117xVVxRhjTHAl+R2AMcYYf1kiMMaYgLNEYIwxAWeJwBhjAs4SgTHGBFyK3wFMVH5+vs6ZM8fvMIwxJq7s2bOnQ1ULRloWd4lgzpw57N692+8wjDEmrojI8dGWWdWQMcYEnCUCY4wJOM8SgYhkiMjbIrJfRCpF5O9HKCMi8l0RqRGRAyKy2qt4jDHGjMzLNoILwM2q2i0iqcA2Efm9qr4VUuYOYIF7uxr4gfvXGGNMlHh2RqCObvdhqnsbPrDRXcDP3bJvATkiMsOrmIwxxnyQp20EIpIsIhVAG/CKqu4cVmQm0BjyuMl9zhhjTJR4mghU9aKqrgJmAetFZNmwIjLSasOfEJEHRWS3iOxub2/3IFJjjAmuqFw1pKpngNeBjcMWNQGzQx7PAppHWP9RVV2rqmsLCkbsD2GMMeYyeXnVUIGI5Lj3M4FbgSPDij0HfN69euga4KyqtngVkzHGmA/y8qqhGcBjIpKMk3CeVNUXRORLAKr6CLAZ2ATUAL3AAx7GY4A9e/b4HcIla9as8TsEYwweJgJVPQCUj/D8IyH3FfiyVzEYY4wZn/UsNsaYgLNEYIwxAWeJwBhjAs4SgTHGBJwlAmOMCThLBMYYE3CWCIwxJuAsERhjTMBZIjDGmICzRGCMMQFnicAYYwLOEoExxgScJQJjjAk4SwTGGBNwlgiMMSbgLBEYY0zAWSIwxpiAs0RgjDEBZ4nAGGMCzhKBMcYEnCUCY4wJOEsExhgTcJYIjDEm4CwRGGNMwFkiMMaYgLNEYIwxAedZIhCR2SLymogcFpFKEfkvI5S5SUTOikiFe/uGV/EYY4wZWYqH2x4A/quq7hWRycAeEXlFVd8ZVm6rqt7pYRzGGGPG4NkZgaq2qOpe934XcBiY6dX+jDHGXJ6otBGIyBygHNg5wuINIrJfRH4vIkujEY8xxpj3eFk1BICIZAO/Bb6iqueGLd4LlKpqt4hsAp4BFoywjQeBBwFKSkq8DdgYYwLG0zMCEUnFSQK/UtXfDV+uqudUtdu9vxlIFZH8Eco9qqprVXVtQUGBlyEbY0zgeHnVkAA/AQ6r6r+MUma6Ww4RWe/Gc8qrmIwxxnyQl1VD1wGfAw6KSIX73N8AJQCq+ghwN/AfRWQAOA/cq6rqYUzGGGOG8SwRqOo2QMYp833g+17FYIwxZnzWs9gYYwLOEoExxgScJQJjjAk4SwTGGBNwlgiMMSbgLBEYY0zAWSIwxpiAs0RgjDEBZ4nAGGMCzhKBMcYEnCUCY4wJOEsExhgTcJYIjDEm4CwRGGNMwFkiMMaYgLNEYIwxAWeJwBhjAs4SgTHGBJwlAmOMCThLBMYYE3CWCIwxJuAsERhjTMBZIjDGmICzRGCMMQFnicAYYwLOEoExxgRcilcbFpHZwM+B6cAg8KiqfmdYGQG+A2wCeoH7VXWvVzEZYyZmz549fodwyZo1a/wOIWF5lgiAAeC/qupeEZkM7BGRV1T1nZAydwAL3NvVwA/cv8YYY6LEs6ohVW0Z+nWvql3AYWDmsGJ3AT9Xx1tAjojM8ComY4wxHxSVNgIRmQOUAzuHLZoJNIY8buKDycIYY4yHPE8EIpIN/Bb4iqqeG754hFV0hG08KCK7RWR3e3u7F2EaY0xgeZoIRCQVJwn8SlV/N0KRJmB2yONZQPPwQqr6qKquVdW1BQUF3gRrjDEB5VkicK8I+glwWFX/ZZRizwGfF8c1wFlVbfEqJmOMMR/k5VVD1wGfAw6KSIX73N8AJQCq+giwGefS0Rqcy0cf8DAeY4wxI/AsEajqNkZuAwgto8CXvYrBGGPM+Lw8IzAxQlWpr6/n8OHDtLW1MWnSJEpKSkhJsbffGGOJIOH19PTw1FNPUV9fT2pqKgD9/f1UV1ezfPlyiouLfY7QGOM3SwQJ7OzZs/z7v/87PT09bNq0iVWrVnHgwAHOnDnDoUOH2Lt3L/39/ZSWlvodqjHGRzboXILq7+/n17/+NefPn+eBBx5g3bp1l84IcnJy2LBhA4WFhRw8eBDrm2FMsFkiSFCbN2+mtbWVT37ykyNW/yQnJ7NmzRqys7PZt28f7777rg9RGmNigSWCBHT8+HEqKiq47rrrWLhw4ajlhpLBwMAAhw4dimKExphYYokgwVy8eJEXX3yRnJwcbrzxxnHLT548mfnz59Pa2sqpU6eiEKExJtZYIkgwFRUVtLe3s3HjxkttAuMpKysjMzOTyspKnK4dxpggsUSQQC5evMjWrVuZNWvWmFVCwyUnJ3PVVVdx7tw5WltbPYzQGBOLLBEkkIqKCs6ePcuNN96IM9RT+IqLi8nKyqKmpsbOCowJmLASgYj8VkQ+IiKWOGKUqrJjxw5mzpxJWVnZhNcXEcrKyjh79iwdHR0eRGiMiVXhfrH/APgMUC0i3xaRqzyMyVyGmpoaTp8+zTXXXDPhs4Ehs2bNIj09nbq6ughHZ4yJZWElAlV9VVU/C6wG6oFXRGSHiDzgzjlgfPb222+TnZ3N4sWLL3sbycnJlJSU0N7eTk9PTwSjM8bEsrCrekQkD7gf+HNgH/AdnMTwiieRmbCdOnWKmpoa1qxZQ3Jy8hVtq6SkBICGhoZIhGaMiQPhthH8DtgKZAEfVdWPqeoTqvqfgGwvAzTjq6ioQERYs2bNFW8rMzOToqIiGhsbuXjxYgSiM8bEunDPCH6sqktU9VtDM4iJSDqAqq71LDozrsHBQQ4cOMD8+fOZPHlyRLZZWlpKX18fbW1tEdmeMSa2hZsI/scIz70ZyUDM5Tl27Bjnzp1j5cqVEdtmfn4+6enpnDhxImLbNMbErjGHoRaR6cBMIFNEynlvxrEpONVExmf79+8nIyODRYsWRWybSUlJFBcXU19fT19fH2lpaRHbtjEm9ow3H8HtOA3Es4DQCei7cOYfNj7q6+vjyJEjrFixIuKzjc2aNYtjx47R0tJi8xUYk+DG/PZQ1ceAx0Tkk6r62yjFZMJUXV1Nf38/S5cujfi2p0yZQnZ2NidOnLBEYEyCG69q6E9V9ZfAHBH56+HLVfVfRljNREllZSXZ2dmefFGLCDNmzKC6upoLFy6Qnp4e8X0YY2LDeI3Fk9y/2cDkEW7GJxcuXKC6uprFixeTlOTNyB/Tp08H4OTJk55s3xgTG8arGvqh+/fvoxOOCVd1dTUDAwOeVAsNmTJlCpmZmbS2tl7qaGaMSTzhdij7JxGZIiKpIrJFRDpE5E+9Ds6MrqqqikmTJjF79mzP9iEiTJ8+nY6ODvr7+z3bjzHGX+HWKdymqueAO4EmYCHwNc+iMmO6ePEiNTU1LFiwwLNqoSEzZsxgcHDQOpcZk8DC/RYZGlhuE/C4qp72KB4ThoaGBt59992I9h0YzbRp00hLS7MJa4xJYOEmgudF5AiwFtgiIgXAu96FZcZSVVVFcnIy8+bN83xfQ9VDbW1tNvaQMQkq3GGoHwY2AGtVtR/oAe4aax0R+amItInIoVGW3yQiZ0Wkwr19Y6LBB5GqUlVVxbx586LW43f69OlcvHjRJqwxJkFNpDvqYpz+BKHr/HyM8j8Dvj9Oma2qeucEYgi89vZ2zpw5w/XXXx+1febn55OSkkJraytFRUVR268xJjrCSgQi8gugDKgAhuoHlDG+5FX1DRGZc4XxmWGqqqoAJjQ5/ZVKSkqisLCQtrY2VPWyZ0AzxsSmcM8I1gJLNPKzmm8Qkf1AM/BVVa0cqZCIPAg8CAT+evajR49SXFwcsSGnw1VQUEBzczPnzp1j6tSpUd23McZb4TYWHwKmR3jfe4FSVV0JfA94ZrSCqvqoqq5V1bUFBQURDiN+dHd309TUFNWzgSGFhYUAdhmpMQko3ESQD7wjIi+JyHNDtyvZsaqeU9Vu9/5mIFVE8q9km4muuroaICqXjQ6Xnp7O1KlTLREYk4DCrRr6ZqR37M51cFJVVUTW4ySlU5HeTyKpq6sjOzvbtwbbgoICampqbI4CYxJMWIlAVf8oIqXAAlV9VUSygDFnSReRx4GbgHwRaQL+Drdjmqo+AtwN/EcRGQDOA/d60AaRMFSVuro6FixY4FtjbWFhITU1NXR0dFBcXOxLDMaYyAv3qqG/wGmszcW5emgm8Ahwy2jrqOp9Y21TVb+Pc3mpCUNrayu9vb1R6UQ2mpycHFJTU2lra7NEYEwCCbeN4MvAdcA5AFWtBgq9Csp8UG1tLYCviSApKYn8/Hza29uxkzdjEke4ieCCqvYNPXA7ldk3QRTV1dVRVFREdna2r3EUFhZy4cIFzp0752scxpjICTcR/FFE/gZnEvs/AX4DPO9dWCZUf38/DQ0Nvp4NDBm6fNeuHjImcYSbCB4G2oGDwP8FbAa+7lVQ5v2OHz/OxYsXKSsr8zsUMjIymDJlCu3t7X6HYoyJkHCvGhoUkWeAZ1TVvgGirLa2luTk5JjpVV1YWEhtbS39/f2kpqaOv4IxJqaNeUYgjm+KSAdwBKgSkXYbKTS66urqKC0tjZkv3fz8fFSVU6es24fx1uDgIO3t7Rw7dow333yT+vp6Gw7dA+OdEXwF52qhdap6DEBE5gE/EJG/UtV/9Ti+wOvq6qKtrY0VK1b4Hcol06ZNIzk5mfb29ksT3BsTSapKY2MjVVVVXLhwAYDKSmcosuzsbG6++WZWrVplAyBGyHiJ4PPAn6jqpYHoVbXOna/4ZcASgcfq6uoAYqJ9YEhycjJ5eXk2P4HxxMDAAHv37qWtrY3c3FyWLVtGbm4uK1eupKGhgR07dvDcc89RXV3NJz7xiZg5U45n4yWC1NAkMERV20XEXv0oqKurIysrK+bmAcjPz6etrY3e3l6ysrL8DsckiP7+ft5++206OztZunQpc+bMufSrf9KkSSxevJirrrqKt956i5dffplz587x+c9/3oY8uULjXTXUd5nLTASoKrW1tZSVlcXcKfDQZaR29ZCJlMHBQfbu3cuZM2dYs2YNc+fOHfG4FxE2bNjApz71KZqbm3nqqacYHBz0IeLEMV4iWCki50a4dQHLoxFgkLW1tdHT0xMT/QeGy87OJiMjw6qHTMQcPnyY9vZ2li9fzowZM8Ytv3jxYjZt2kR1dTVbtmyJQoSJa8yqIVUdc2A5461YGFZiNCJCQUEBra2tNmuZuWJtbW0cO3aMOXPmTOgy6bVr19La2sqOHTsoKyuLyc9KPAi3Q5nxQV1dHQUFBUyZMsXvUEaUn59Pf38/Z86c8TsUE8f6+vrYv38/kydPZvHixRNe//bbbyc/P5+nn36ad99914MIE58lghg1MDDA8ePHY/oXjrUTmEg4cuQIfX19rFq1iuTkiVdCpKam8vGPf5zu7m5ee+01DyJMfJYIYlRDQwMDAwMxddnocGlpaUydOtXaCcxlO3PmDA0NDcyZM+eK5sKeOXMma9euZdeuXbS0tEQwwmCwRBCjamtrSUpKorS01O9QxlRQUEBnZyf9/f1+h2LijKpSWVlJWlpaRObhvuWWW8jMzOTll1+2YdInyBJBjKqrq6OkpCTmr4+24SbM5Tp58iSdnZ0sWrQoIp3CMjIyuPHGG6mvr6empiYCEQaHJYIY1N3dTWtra0y3DwwJHW7CmHCpKlVVVUyaNInZs2dHbLtr1qxh2rRpvPrqq3ZWMAGWCGLQsWPHgNgaVmI0NtyEuRzNzc10dXWxaNEikpIi9zWUnJzMTTfdRFtbG4cPH47YdhOdJYIYVFtbS2ZmZtwM6Jafn09PTw+9vb1+h2LigKpSU1NDdnZ2WB3HJmrZsmXk5eXxxhtv2FlBmCwRxBhVpa6ujnnz5kX0l5KX7DJSMxHt7e10dXV5NnRKUlIS119/PSdPnuTo0aMR334iio9vmgDp6Oigq6srLtoHhthwE2YiamtrSU9PZ+bMmZ7tY8WKFeTk5LBjxw7P9pFILBHEmFgeVmI0Q8NNdHR02Km4GdOZM2c4deqU52e8SUlJrF+/noaGBpqbmz3bT6KwRBBj6urqyMvLIycnx+9QJqSgoMCGmzDjqq2tJSUlJSrTrpaXl5OWlsbOnTs931e8s0QQQwYGBqivr4+rs4Eh+fn5gLUTmNH19PTQ0tIStWlXMzIyKC8v59ChQ3R1dXm+v3hmiSCGNDU10d/fHxeXjQ5nw02Y8dTX1yMizJ07N2r7XL9+PYODg+zatStq+4xHniUCEfmpiLSJyKFRlouIfFdEakTkgIis9iqWeDE0rMScOXP8DuWy2HATZjQXL16kqamJ6dOnk5GREbX95ubmctVVV7F79247Lsfg5RnBz4CNYyy/A1jg3h4EfuBhLHGhrq6OWbNmkZ6e7ncol8WGmzCjaWlpob+/35exs66++mrOnz/PgQMHor7veOFZIlDVN4DTYxS5C/i5Ot4CckQk8r1L4kRvby/Nzc1x2T4wxIabMKNpaGggKyuLvLy8qO+7tLSU6dOns2vXLruqbRR+thHMBBpDHje5z32AiDwoIrtFZHeifsnE07ASo7HhJsxIuru7OX36NCUlJb7MZCcirFmzhpMnT9qlpKPwMxGMdESMmK5V9VFVXauqa4d6sSaaoU42xcXFfodyRQoKCmy4CfM+DQ0NiAizZs3yLYbly5eTmprK7t27fYshlvmZCJqA0GEHZwGBTNdDY6+UlZXFzbASo7HLSE2owcFBmpqaKCwsjGoj8XDp6eksW7aMyspKm85yBH5+6zwHfN69euga4KyqBnJqodCxV+KdDTdhQrW2ttLX1xcTEyytXbuW/v5+Dh486HcoMcfLy0cfB94EFolIk4h8UUS+JCJfcotsBuqAGuBHwF96FUusG5pEY/78+T5HcuVsuAkTqqGhgczMTGKhSnfGjBlMnz6dPXv22LE5TIpXG1bV+8ZZrsCXvdp/PKmtraWgoIApU6b4HUpEFBQU0NjYyJkzZ5g2bZrf4Rif9Pb20tHRwcKFC31pJB5uqNH4xRdf5MSJE762WcSa+K6QTgB9fX0cP348Ic4Ghlg7gQHnbACI6AxkV2r58uWkpaWxZ88ev0OJKZYIfHb8+HEuXryYEO0DQ2y4CTM4OEhjYyOFhYVkZmb6Hc4l1mg8MksEPqupqSElJSUmGtMiyYabCLa2tjYuXLgQlVFGJ2r16tX09/dz6NCIo98EkiUCn9XW1jJ37lxSUjxrrvGFDTcRbA0NDaSnp1NYWOh3KB9QXFxMUVERe/fu9TuUmGGJwEednZ2cOnUqoaqFhthwE8F1/vx52tramD17dkz2ixERVq9eTUtLCy0tgbxi/QNi710KkES6bHQ4G24iuBobnZFjYqmReLjly5eTkpJiZwUuSwQ+qqmpIScnh9zcXL9D8YQNNxE8qkpjYyP5+flMmjTJ73BGlZmZyZIlSzh48KC1Y2GJwDf9/f3U1dXFzDXWXhjqRGTVQ8HR3t7O+fPnY7KReLjVq1dz4cIFKisr/Q7Fd5YIfFJXV8fAwACLFi3yOxTPTJo0iaysLE6ePOl3KCZKGhoaSEtLY/r06X6HMq6SkhLy8vKseghLBL45evQo6enpCXfZaCgRoaioiI6ODgYGBvwOx3js3Xff5eTJk8yaNSsmG4mHExHKy8tpbGwM/Flr7L9bCUhVOXr0KGVlZSQnJ/sdjqeKiooYHBy0RuMAaGpqQlXjolpoyMqVK0lKSgr8WYElAh80NzfT3d3NwoUL/Q7Fc7m5uaSkpFj1UIJTVRoaGsjNzSU7O9vvcMKWnZ3NokWL2L9/f6DPWi0R+ODo0aOICAsWLPA7FM8lJSVRWFjIyZMnbcTHBHbq1Cl6e3vj6mxgyOrVqzl//jxVVVV+h+IbSwQ+OHr0KLNnzyYrK8vvUKJi+vTp9PX10dnZ6XcoxiMNDQ2kpqYyY0b8TTs+b948pk6dGujqIUsEUXb27FlaW1sDUS00pKCgABGx6qEE1dfXR2trKzNnzozLNq+kpCTKy8upq6sL7I8VSwRRdvToUYCEvmx0uNTUVPLy8iwRJKimpiYGBwfjslpoyKpVqxAR9u3b53covrBEEGVVVVXk5uaSl5fndyhRVVRURHd3N93d3X6HYiJoqJE4JycnridWmjp1KvPnz6eiooLBwUG/w4k6SwRR1NvbS11dHYsXL07Y3sSjKSoqArCzggTT2dlJd3d3XJ8NDFm9ejVdXV1UV1f7HUrUWSKIoiNHjqCqLF261O9Qoi4rK4upU6fS3Nzsdygmgurr60lJSaG4uNjvUK7YggULyM7ODmSjsSWCKKqsrCQ3Nzcuut97obi4mLNnz9LT0+N3KCYCLly4QGtrK7NmzUqI+TSSk5NZuXIl1dXVnDt3zu9wosoSQZT09PRw7NgxlixZErhqoSFDlxbaGPCJobGxkcHBwYQaJmX16tWoKhUVFX6HElWWCKIkyNVCQ7Kyspg2bZpVDyWA0J7EkydP9juciMnNzWXu3Lns27cvUB0gLRFESWVlJXl5eZcaTYNqxowZnDt3zq4einPt7e309vYm1NnAkPLycs6cOUNdXZ3foUSNJYIo6Onpob6+PtDVQkOGGhXtrCC+HT9+PG6Gm56oxYsXk5mZGahGY0sEUXD48OHAVwsNycjIIDc3l+bm5kCdeieS0DmJ47En8XhSUlJYuXIlR44cCUyjsSWCKDh06BB5eXkUFhb6HUpMKC4upru7m7a2Nr9DMZehvr4eVU3IaqEh69evR1XZtWuX36FEhaeJQEQ2ikiViNSIyMMjLL9JRM6KSIV7+4aX8fjh9OnTHD9+nJUrVwa+WmjI0NVDBw8e9DkSM1EDAwM0NDQwffr0hB40cdq0aSxatIg9e/YEYk5jzxKBiCQD/wbcASwB7hORJSMU3aqqq9zbP3gVj18qKioQEVauXOl3KDEjPT2dwsJCDhw4EMju/PGsqamJ/v5+5s2b53conrv66qs5f/48Bw4c8DsUz3l5RrAeqFHVOlXtA34N3OXh/mLO4OAgFRUVlJWVxfU4LF6YPXs2XV1d1NTU+B2KCZOqcuzYMaZOncq0adP8DsdzpaWlTJ8+nZ07dyZ8e5aXiWAm0BjyuMl9brgNIrJfRH4vIiO2porIgyKyW0R2x9PconV1dXR1dVFeXu53KDGnqKiISZMmBerKjHjX1tZGT08P8+bNC0Q1p4hw9dVX097ezrFjx/wOx1NeJoKRjpThaXUvUKqqK4HvAc+MtCFVfVRV16rq2oKCgshG6aF9+/aRmZkZqLkHwpWUlMSqVas4evQoXV1dfodjwnDs2DEyMjLicvKZy7Vs2TImTZrEW2+95XconvIyETQBs0MezwLed/G4qp5T1W73/mYgVUTyPYwpanp7ezly5AgrVqxIiHFYvFBeXh7I7vzx6Ny5c3R0dDBnzhySkoJzsWFKSgrr1q2juro6oUfO9fId3QUsEJG5IpIG3As8F1pARKaLe44pIuvdeE55GFPUDDWEWrXQ6PLy8igtLQ1cd/54VF1dTXJyckIMNz1R69evJy0tja1bt/odimc8SwSqOgA8BLwEHAaeVNVKEfmSiHzJLXY3cEhE9gPfBe7VBPhGUFX27dvHjBkzAj+kxHhWr15NZ2cn9fX1fodiRtHV1UVLSwtz584lLS3N73CiLjMzk3Xr1lFZWUlHR4ff4XjC03M8Vd2sqgtVtUxV/6f73COq+oh7//uqulRVV6rqNaq6w8t4oqW+vp62tjbWrVvndygxb/HixWRkZLB7926/QzGjGDobCMIlo6PZsGEDKSkpbNu2ze9QPBGcyr4o2rlzJ1lZWSxfvtzvUGJeamoq5eXlHD58mLNnz/odjhmmu7ub5uZm5syZE8izgSGTJk1izZo1HDhwICEnuLdEEGGdnZ1UVVWxZs0aayQO09VXXw04CdTElurqapKSkgJ9NjDk2muvJSkpie3bt/sdSsRZIoiwHTt2kJSUxNq1a/0OJW5MnTqVJUuWsHfvXi5cuOB3OMZ1+vRpTpw4wZw5c0hPT/c7HN9NmTKF8vJy9u3bx+nTp/0OJ6IsEURQd3c3FRUVrFy50noST9CGDRu4cOGCtRXEkC1btgS+bWC4D33oQyQnJ7Nlyxa/Q4koSwQRtHPnTgYGBrjuuuv8DiXuzJw5k3nz5vHmm28GYpCvWNfQ0MA777xDWVkZGRkZfocTMyZPnsy1117LO++8Q2Nj4/grxAlLBBHS29vL22+/zZIlS8jLy/M7nLh0ww030NPTY8NO+ExVeemll5g8ebKdDYzg2muvJTs7m5dffjlh+r9YIoiQHTt20NfXx4033uh3KHGrtLSUkpIStm3bZmcFPjpw4ADNzc3ccsstdsHDCNLS0vjwhz9MU1MThw8f9juciLBEEAHd3d3s3LmT5cuX2+QzV0BEuOWWW+ju7k74sV1iVV9fH1u2bKG4uJgVK1b4HU7MWrVqFYWFhbz66qsJ8aPFEkEEvPbaawwODnLTTTf5HUrcKykpYeHChWzfvp3e3l6/wwmcbdu20dXVxe233x6IEUYvV1JSEhs3bqSzs5PXX3/d73CumCWCK9Ta2sq+fftYt24dubm5foeTEG655Rb6+vp47bXX/A4lUE6ePMn27dtZuXJlIMcUmqi5c+dSXl7Om2++SXNz8/grxDBLBFdgqFEtIyPD2gYiqLCwkLVr17Jnzx5aW1v9DicQBgcHef7558nIyOC2227zO5y4cdttt5Gdnc2zzz7LxYsX/Q7nslkiuAL79++nvr6em2++mczMTL/DSSgf/vCHyczMZPPmzTadZRRs3bqVEydOcMcddyT0XMSRlpGRwUc+8hHa2triehwiSwSXqbu7m5deeomSkhLWrFnjdzgJJzMzk9tuu43Gxkbefvttv8NJaCdOnOCPf/wjy5cvZ9myZX6HE3cWLVrEsmXLeOONN+K2isgSwWVQVV544QX6+/v56Ec/ao1qHlmxYgULFixgy5YtCdelP1acP3+e3/zmN0yZMoVNmzb5HU7cuuOOO8jOzuY3v/kN7777rt/hTJglgsuwZ88eqqqquPXWW8nPT4gJ1WKSiHDnnXeSkpLCU089xcDAgN8hJZTBwUGefvppurq6uOeee6wH8RXIysri7rvv5uzZszzzzDNx19HMEsEEtbS08NJLL1FWVnZp1EzjnSlTpnDXXXfR0tLCK6+84nc4CeXVV1+lurqajRs3MnPmTL/DiXuzZ8/m9ttvp6qqKu7GIrJEMAE9PT088cQTZGVl8YlPfMKqhKLkqquu4uqrr+btt99m3759foeTEHbt2sWbb77JunXrbAKlCFq/fj2rV69m+/btcTWAovUfD1NfXx+PP/44PT09PPDAA0yaNMnvkALltttuo729nRdeeIGcnBzmzp3rd0hx68CBA2zevJmFCxeyceNGv8NJKCLCpk2b6Orq4sUXXyQjIyMuGuDtjCAMAwMDPPnkkzQ3N/PJT36S4uJiv0MKnKSkJO6++27y8vJ4/PHHaWpq8jukuFRRUcEzzzzD3Llzueeee0hKsq+ASEtOTuaee+6hpKSE3/3udxw4cMDvkMZlR8E4+vv7efzxx6mtreXOO+/kqquu8jukwMrMzORzn/sc2dnZ/PKXv+T48eN+hxQ3VJUdO3bw7LPPMnfuXO69914bUM5DqampfPazn6W0tJSnn36aN998M6YbkC0RjKG7u5vHHnuMY8eOcdddd7F69Wq/Qwq8yZMn84UvfIHJkyfzy1/+MmFGf/TSwMAAzz//PK+88gpLlizhvvvuC/T8w9GSlpbGZz7zGRYvXszLL7/MCy+8ELNXvlkiGEVDQwM/+tGPaGtr41Of+hSrVq3yOyTjmjp1Kvfffz9FRUU8+eSTbNmyxXofj6K9vZ0f//jH7Nu3jxtuuIG7777bzgSiKDU1lXvuuYfrrruOvXv38pOf/IS2tja/w/oAOyKGGRgY4I033mDbtm3k5OTwwAMPMGPGDL/DMsNMmjSJ+++/n82bN7Nt27ZLZ20FBQV+hxYT+vv72bFjB1u3biU9PZ3PfOYzLFiwwO+wAklEuPXWW5k9ezbPPvssP/zhD9mwYQM33ngjqampfocHWCK4ZHBwkEOHDvH666/T2dnJypUrueOOO2zS7hiWkpLCxz72MebNm8eLL77II488wvr167nuuuvIzs72OzxfXLx4kX379vHGG2/Q1dXF0qVL2bhxY2Bfj1iyaNEiHnroIV555RW2b99OZWUlN954I8uXLyc5OdnX2CSWGzBGsnbtWo3k9bkDAwNUVlaydetWTp06RWFhIbfffnvCTtG3Z88ev0O4JJJjNPX09PDqq69SUVFBSkoK5eXlXHvtteTk5ERsH7Gsu7ub/fv3s3v3bs6cOcPs2bO5+eabmTNnzhVtN1GPF7/V19fzhz/8gZMnTzJlyhSuueYaVq1a5englSKyR1XXjrgsiIng4sWLHD9+nIMHD3L48GEuXLhAUVERH/rQh1i8eHFCdxRL9A/2qVOn2L59O/v372dwcJDS0lKWLl3K4sWLE+5XcVdXF7W1tRw5coSjR4+iqsyePZsbbriB+fPnR+Q4TvTjxU+qSm1tLdu2beP48eMkJSVRVlbG0qVLmT9/fsT7KvmWCERkI/AdIBn4sap+e9hycZdvAnqB+1V1zJnLLzcRnD59mkOHDnH8+HEaGhoYGBggLS2NxYsXs3z5cubNm5fQCWBIUD7Y586dY+/evVRWVtLR0QFAQUEBJSUlzJ49m6KiIvLy8mKmjnY8fX19dHR00NLSQnNzM01NTZcaHbOzs1mxYgXl5eURH/sqKMeL31paWjh06BCVlZWcPXsWeO94LS4uprCwkMLCwiu62suXRCAiycBR4E+AJmAXcJ+qvhNSZhPwn3ASwdXAd1R1zAF8LjcRHDlyhCeeeIKioiJKS0uZO3cuZWVlcfNFEClB+2CrKu3t7VRVVdHQ0EBjYyMXLly4tDwnJ4cpU6aQnZ1NdnY2kydPJisri7S0tA/cUlJSSEpKQkTG/Ds4OIiqoqqj3u/v7//Ara+vj97eXnp7e+np6aG3t5euri46Ozvp6em5FHNGRgbFxcXMmzePsrIyioqKPPsRE7TjxW+qyokTJ6ivr7/0o7Wvr+/S8uuvv55bbrnlsrY9ViLwsrF4PVCjqnVuEL8G7gLeCSlzF/BzdbLRWyKSIyIzVLUl0sGUlZXxta99zSbdCBgRufRrCpyLAjo6Omhvb6e9vZ1Tp07R1dXFyZMnqamped+Hzk9ZWVlkZWWRnZ3NwoULyc3NJTc3lxkzZpCTkxOIs9cgEhFmzZrFrFmzuP7661FVOjs7aWtro62tzbNRDbxMBDOBxpDHTTi/+scrMxN4XyIQkQeBB92H3SJSFdlQr1g+0OF3EGGKp1jB4vVSPMUKFu+VKh1tgZeJYKSfLMProcIpg6o+CjwaiaC8ICK7RzvlijXxFCtYvF6Kp1jB4vWSlz2Lm4DZIY9nAcPncQunjDHGGA95mQh2AQtEZK6IpAH3As8NK/Mc8HlxXAOc9aJ9wBhjzOg8qxpS1QEReQh4Cefy0Z+qaqWIfMld/giwGeeKoRqcy0cf8Coej8VstdUI4ilWsHi9FE+xgsXrmbjrUGaMMSaybPRRY4wJOEsExhgTcJYIwiAiuSLyiohUu3+njVBmtoi8JiKHRaRSRP5LyLJvisgJEalwb5s8inOjiFSJSI2IPDzCchGR77rLD4jI6nDX9SHWz7oxHhCRHSKyMmRZvYgcdF/LqMwQHka8N4nI2ZD3+BvhrutTvF8LifWQiFwUkVx3WVRfXxH5qYi0icihUZbHzHEbZrwxdeyGZajLu91GvwH/BDzs3n8Y+F8jlJkBrHbvT8YZXmOJ+/ibwFc9jjEZqAXmAWnA/qH9h5TZBPwep//GNcDOcNf1IdZrgWnu/TuGYnUf1wP5UXz/w4n3JuCFy1nXj3iHlf8o8H98fH0/BKwGDo2yPCaO2wnEGzPHbrg3OyMIz13AY+79x4CPDy+gqi3qDpinql3AYZxe0tFyaUgPVe0Dhob0CHVpSA9VfQvIEZEZYa4b1VhVdYeqdroP38LpY+KXK3l9ov3aXs4+7wMe9zimUanqG8DpMYrEynEbVrwxduyGxRJBeIrU7d/g/i0cq7CIzAHKgZ0hTz/knir+dKSqpQgYbbiOcMqEs24kTXR/X8T5RThEgZdFZI87/IjXwo13g4jsF5Hfi8jSCa4bSWHvU0SygI3Ab0OejvbrO55YOW4vh9/HblhshjKXiLwKTB9h0X+f4HaycT5UX1HVc+7TPwD+Eecg+Efg/wH+7PKjHXnXIzwX7pAeYQ31EUFh709EPozzYbo+5OnrVLVZRAqBV0TkiPsrzSvhxLsXKFXVbrcN6BlgQZjrRtpE9vlRYLuqhv7CjfbrO55YOW4nJEaO3bBYInCp6q2jLRORk+KOiuqeko44+7SIpOIkgV+p6u9Ctn0ypMyPgBciF/klVzKkR1oY60ZSWEOLiMgK4MfAHap6auh5VW12/7aJyNM4VQRefpjGjTck6aOqm0Xkf4tIfjjremAi+7yXYdVCPry+44mV4zZsMXTshsfvRop4uAH/zPsbi/9phDIC/Bz4f0dYNiPk/l8Bv/YgxhSgDpjLew1nS4eV+Qjvb3R7O9x1fYi1BKfH+bXDnp8ETA65vwPY6PH7H06803mvg+Z6oMF9naP62k7k/QSm4tR1T/Lz9XX3NYfRG19j4ridQLwxc+yG/f/4HUA83IA8YAtQ7f7NdZ8vBja796/HOS09AFS4t03usl8AB91lzxGSGCIc5yacq5Vqgf/uPvcl4EvufQH+zV1+EFg71roev6bjxfpjoDPktdztPj/P/cDvByqjEWuY8T7kxrMfp4Hw2rHW9Tte9/H9DPtR4sfri3NG0gL04/z6/2KsHrdhxhtTx244NxtiwhhjAs6uGjLGmICzRGCMMQFnicAYYwLOEoExxgScJQJjjAk4SwRxxB0hsiLk9rD7/OsiMuFJskVklXg0EuqVEpGPi8iSkMf/ICKjdvpzy2wWkRz39peXsc8ZIvKCez9PnNFku0Xk+2OsM+7ItJcRh4jI191tHnXjWBqyvN7trIaIrBGRYyJSPs42v+Bur1pEvjBKmXQRecIdyXOnO1RKVInIHBH5zBjLl4rI/3Ffl2oR+VsREXfZ/UPvlYgkichj7pAuIiKvejS0S0KwRBBfzqvqqpDbt69we6twrsP+ABHxu9f5x4FLiUBVv6Gqr461gqpuUtUzQA4w4UQA/DXwI/f+u8DfAl8dZ52HgS2qugCnj0kkhkL+Ms4IlitVdSHwLeA5EckILeT2Xn0K+LSq7httY+IML/13wNU4nd3+bpQvxS8Cnao6H/hX4H9F4H+ZqDnAiIlARDJx+uF8231dVuK8Tn85rJwAjwCpwJ+rc438L4aXMyH87shgt/BvQPcoz7+O28kGuA14E2fsm98A2e7z63B6Mu4H3sbpVdoAtON0evk0znDZjwIvA/8fUIrz5XbA/VvibutnwHfd7dUBd4fE8jVgl7vO34c8/7fAEeAVnA45X3Wf/wu3/H6c4TmycD7cp4Fjbmxl7j7vxhnW98mQ7d4EPO/erwfycUahPO+u+884XwJ3hazzK+BjI7yOdUD6sOfuB74/xntShdtBEGco8qoIvM+NQNmw534BfDHk/7zBjff6MLZ3H/DDkMc/BO4bodxLwAb3fgrQgdtb+gr+lxGPFZxOYv8MHMLpJPZp9/m3gLPue/dXw7b1RZxRSEOfKwMaQ98rd3+/BVJCyk1jlJ7AdlMbayjOZIpIRcjjb6nqE0MP3OqCrwO3qmqPiPzfwF+LyLeBJ3A+bLtEZArQC3wDJ4E85K7/TWANzpfLeRF5HueD95iI/BnOB+zj7u5m4PSmvgrnV9pTInIbzkBr63E+6M+JyIfcfX0SZ0TWFJwktcfdzu9U9Ufu/v8Hzpfd90TkOZzx/Z9ylw39m68APxSRSarag5PALr0GroeBZaq6yl33RpyhPZ4Vkak4ieZ91SMiMhfn1/CF0V/+Eb1vZFp3MLH3EZHJwNZR1v+Mqr4TUnYKzpAPtcPK7QaWhjx+FvhTVd0WRozhjtJ5qZyqDojIWZxe9R2hhURkK86cG8N9VUc+a/vAsQL8B5wz0pU4yXuXiLyB8959VVXvHGE7S3nvuMGNs1ZEst3XDZyzicPATao6EFKu0636ytOQsX+MwxJBfDk/9OU2imtwqlO2u1+caThnB4uAFlXdBe8NkBby5RrqOVU9797fgPOBBecX6T+FlHtGVQeBd0SkyH3uNvc2VE2RjZMYJgPPDm3XTTBDlrkJIMct/9IY/9/QF9QfgI+KyFM449D8t3HW+aOI/Jv7Jf0fgN+Gfkm4ZuCcHUWcOvNTrLrCzQjvH1nzVeDPReQlVb0YxrofCOtyy6nqDePsb7iRjpXrgcfd2E+KyB9xzlrPjbYRPvgajBTnXpyEsx7YPqxMG86wMJYIhrE2gsQiwCv6XhvCElX9ImN/gIbrGWNZ6DZCfzlLyN9vhex/vqr+hJG/YIb8DHhIVZcDfw9kjFF2yBPAp4CbgV3uF+14fgF8FngA+PcRlp8Pc9/DnRRnRFpklJFpRWTysEb+0NuS0LJuku4RkXnDNrMaeCfk8UPu3/8dRozhjkZ6qZzbRjQ0KN3w/2frKP/LaI35ox0rE1UJvO+iCPd16g45Bo7gHBtPhDawuzJw3mczjCWCxPIWcJ2IzAdn0hERWYjz4SgWkXXu85PdD3oXI5/iD9mBM0wxOF+i41VDvAT8mThzMiAiM91f4dtwfsFnuMs+ErLOZKBFnCG8Pxvy/FixvY7zxfgXfLBaaLR1fwZ8BUBVK0dY5yhOQ+VEPcd71UxfwKmyeR9V7dL3N/KH3t4ZXh6n7vy7buMo7hfs9TjtNkMGcer+F4nIP7jl1ovIz0fY3kvAbSIyzW0kvo2Rz7xC/5e7caavHPGMYJT/ZczG/GHeAD4tIskiUoAz/ePbjP2+/wq4fijhuK/Pd3n/mSqqugNnELgXRaTELSs4I8TWTyDGwLCqofgyvI3gD6p66SoVVW0XkfuBx0Uk3X3666p6VEQ+DXzP/fCcB24FXgMedrf5rRH295+Bn4rI13CqTR4YKzhVfVlEFgNvutVO3Tj12LvcOv/9wHGc+u6z7mp/izOT23GcRsOhL4FfAz8Skf+M86UUup+L4lzmeT/D6vrd5adEZLs4k4v/XlW/pqonReQwzoQxI8XeIyK1IjJfVWvAuUwTmAKkicjHgdtU9R0R+THwiKruBr4NPCkiX8RpfL9nrNcoTN/Dadw8KCIXgVacxu73/ZpV1QsichfwRxE5CZxkhF+8qnpaRP4Rp1Ee4B/UnYjGTSK7VfU54CfAL0SkBudM4N7h24qgp3GqHvfjnGn+N1VtFZFTwICI7Ad+pqr/GvJ/nHf/3++JyL/hzFn8C5wG4vdR1RfcBPMHEbkBZ6jqt0aoEjRgo4+a6BCRbHVm78rC+TX4oLpzPEdp/1k4iWa1qp4dpcwngDWq+vVoxRVJIvLPwC9U9YDfscQaEfkOTvvXFr9jiUV2RmCi5VG3PjwDeCzKSeBW4KfAv4yWBABU9WkRyYtWXJGmql/zO4YYdsiSwOjsjMAYYwLOGouNMSbgLBEYY0zAWSIwxpiAs0RgjDEBZ4nAGGMC7v8HEu5+hk6DAkwAAAAASUVORK5CYII=\n",
      "text/plain": [
       "<Figure size 432x288 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "# Plot distribution of final rewards by random policy\n",
    "plt.figure()\n",
    "final_rewards_random = Q_targets[0::5].reshape(-1) # find final rewards\n",
    "sns.distplot(final_rewards_random, color= 'grey')\n",
    "plt.xlabel('Rewards')\n",
    "plt.show()\n",
    "\n",
    "# Plot distribution of en by random policy\n",
    "plt.figure()\n",
    "en_random = en_targets[0::5].reshape(-1) # find final rewards\n",
    "sns.distplot(en_random, color= 'grey')\n",
    "plt.xlabel('Electronegativity (1.0 = OK, 0.0 = not OK)')\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "before scaling: tensor([[3.0000e+00, 6.7823e-01, 6.2488e-01,  ..., 0.0000e+00, 5.5973e-02,\n",
      "         9.8478e-03],\n",
      "        [3.0000e+00, 6.7823e-01, 6.2488e-01,  ..., 0.0000e+00, 5.5973e-02,\n",
      "         9.8478e-03],\n",
      "        [2.0000e+00, 8.6897e-01, 8.5846e-01,  ..., 0.0000e+00, 6.2480e-04,\n",
      "         7.6507e-05],\n",
      "        ...,\n",
      "        [2.0000e+00, 7.1429e-01, 6.4256e-01,  ..., 0.0000e+00, 6.3796e-03,\n",
      "         1.5623e-03],\n",
      "        [1.0000e+00, 1.0000e+00, 1.0000e+00,  ..., 1.0000e+00, 0.0000e+00,\n",
      "         0.0000e+00],\n",
      "        [0.0000e+00, 0.0000e+00, 0.0000e+00,  ..., 0.0000e+00, 0.0000e+00,\n",
      "         0.0000e+00]])\n",
      "after scaling: tensor([[ 0.9094,  0.2304,  0.1828,  ..., -0.5626, -0.3777, -0.3565],\n",
      "        [ 0.9094,  0.2304,  0.1828,  ..., -0.5626, -0.3777, -0.3565],\n",
      "        [ 0.1566,  0.7751,  0.8540,  ..., -0.5626, -0.6569, -0.6213],\n",
      "        ...,\n",
      "        [ 0.1566,  0.3333,  0.2336,  ..., -0.5626, -0.6279, -0.5811],\n",
      "        [-0.5963,  1.1493,  1.2607,  ...,  1.7776, -0.6601, -0.6234],\n",
      "        [-1.3491, -1.7066, -1.6129,  ..., -0.5626, -0.6601, -0.6234]])\n"
     ]
    }
   ],
   "source": [
    "# Standardizing s_material data\n",
    "scaler = StandardScaler() # Fit scaler\n",
    "print('before scaling:', s_material_inputs)\n",
    "scaler.fit(s_material_inputs)\n",
    "s_material_inputs_scaled = torch.tensor(scaler.transform(s_material_inputs)).float()\n",
    "print('after scaling:', s_material_inputs_scaled)\n",
    "\n",
    "# # Save standard scaler\n",
    "# joblib.dump(scaler, 'data/oxides_calcine/std_scaler.bin', compress=True)\n",
    "\n",
    "# Load standard scaler\n",
    "scaler = joblib.load('data/oxides_calcine/std_scaler.bin')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "TRAIN SIZES:\n",
      "s_material_inputs_train: torch.Size([4000, 145])\n",
      "s_step_inputs_train: torch.Size([4000, 5])\n",
      "a_elem_inputs_train: torch.Size([4000, 78])\n",
      "a_comp_inputs_train: torch.Size([4000, 10])\n",
      "Q_targets_train: torch.Size([4000, 1])\n",
      "en_targets_train: torch.Size([4000, 1])\n",
      "\n",
      "VALIDATION SIZES:\n",
      "s_material_inputs_valid: torch.Size([1000, 145])\n",
      "s_step_inputs_valid: torch.Size([1000, 5])\n",
      "a_elem_inputs_valid: torch.Size([1000, 78])\n",
      "a_comp_inputs_valid: torch.Size([1000, 10])\n",
      "Q_targets_valid: torch.Size([1000, 1])\n",
      "en_targets_valid: torch.Size([1000, 1])\n"
     ]
    }
   ],
   "source": [
    "# Train-validation split\n",
    "def train_validation_split(s_material_inputs_scaled, s_step_inputs, a_elem_inputs, a_comp_inputs, Q_targets, en_targets, valid_size = 0.2):\n",
    "    '''\n",
    "    Args:\n",
    "    s_material_inputs_scaled\n",
    "    s_step_inputs\n",
    "    a_elem_inputs\n",
    "    a_comp_inputs\n",
    "    Q_targets\n",
    "    en_targets\n",
    "    valid_size: Float (0-1). Size of validation set\n",
    "\n",
    "    Returns:\n",
    "    s_material_inputs_train \n",
    "    s_step_inputs_train     \n",
    "    a_elem_inputs_train     \n",
    "    a_comp_inputs_train     \n",
    "    Q_targets_train   \n",
    "    en_targets_train   \n",
    "    s_material_inputs_valid \n",
    "    s_step_inputs_valid \n",
    "    a_elem_inputs_valid\n",
    "    a_comp_inputs_valid \n",
    "    Q_targets_valid\n",
    "    en_targets_valid\n",
    "  \n",
    "    '''\n",
    "    max_idx = s_material_inputs_scaled.shape[0] # Find size of dataset\n",
    "    idx = [x for x in range(max_idx)] # Generate indices\n",
    "    random.shuffle(idx) # Shuffle order idx \n",
    "    split = int(round(max_idx*(1-valid_size)))\n",
    "    train_idx, valid_idx = idx[:split], idx[split:] # Split indices\n",
    "\n",
    "    s_material_inputs_train = s_material_inputs_scaled[train_idx] # Use scaled s_material\n",
    "    s_step_inputs_train     = s_step_inputs[train_idx]\n",
    "    a_elem_inputs_train     = a_elem_inputs[train_idx]\n",
    "    a_comp_inputs_train     = a_comp_inputs[train_idx]\n",
    "    Q_targets_train         = Q_targets[train_idx]\n",
    "    en_targets_train         = en_targets[train_idx]\n",
    "\n",
    "    s_material_inputs_valid = s_material_inputs_scaled[valid_idx] # Use scaled s_material\n",
    "    s_step_inputs_valid     = s_step_inputs[valid_idx]\n",
    "    a_elem_inputs_valid     = a_elem_inputs[valid_idx]\n",
    "    a_comp_inputs_valid     = a_comp_inputs[valid_idx]\n",
    "    Q_targets_valid         = Q_targets[valid_idx]\n",
    "    en_targets_valid         = en_targets[valid_idx]\n",
    "\n",
    "    return s_material_inputs_train, s_step_inputs_train, a_elem_inputs_train, a_comp_inputs_train, Q_targets_train, en_targets_train,s_material_inputs_valid, s_step_inputs_valid, a_elem_inputs_valid, a_comp_inputs_valid, Q_targets_valid, en_targets_valid\n",
    "\n",
    "s_material_inputs_train, s_step_inputs_train, a_elem_inputs_train, a_comp_inputs_train, Q_targets_train, en_targets_train, s_material_inputs_valid, s_step_inputs_valid, a_elem_inputs_valid, a_comp_inputs_valid, Q_targets_valid, en_targets_valid \\\n",
    "= train_validation_split(s_material_inputs_scaled = s_material_inputs_scaled.to('cuda'), \n",
    "                                    s_step_inputs = s_step_inputs.to('cuda'), \n",
    "                                    a_elem_inputs = a_elem_inputs.to('cuda'),\n",
    "                                    a_comp_inputs = a_comp_inputs.to('cuda'),\n",
    "                                    Q_targets  = Q_targets.to('cuda'),\n",
    "                                    en_targets = en_targets.to('cuda'),\n",
    "                                    valid_size = 0.2)\n",
    "\n",
    "print('TRAIN SIZES:')\n",
    "print('s_material_inputs_train:', s_material_inputs_train.shape)\n",
    "print('s_step_inputs_train:', s_step_inputs_train.shape)\n",
    "print('a_elem_inputs_train:', a_elem_inputs_train.shape)\n",
    "print('a_comp_inputs_train:', a_comp_inputs_train.shape)\n",
    "print('Q_targets_train:', Q_targets_train.shape)\n",
    "print('en_targets_train:', en_targets_train.shape)\n",
    "print('')\n",
    "print('VALIDATION SIZES:')\n",
    "print('s_material_inputs_valid:', s_material_inputs_valid.shape)\n",
    "print('s_step_inputs_valid:', s_step_inputs_valid.shape)\n",
    "print('a_elem_inputs_valid:', a_elem_inputs_valid.shape)\n",
    "print('a_comp_inputs_valid:', a_comp_inputs_valid.shape)\n",
    "print('Q_targets_valid:', Q_targets_valid.shape)\n",
    "print('en_targets_valid:', en_targets_valid.shape)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Here we train DQN using ONLY data generated by random policy."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# DQN"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "CUDA for model:  True\n",
      "Epoch =  0 Train_loss = 825.0890 Valid_loss = 817.6148\n",
      "Epoch =  1 Train_loss = 823.8765 Valid_loss = 815.4438\n",
      "Epoch =  2 Train_loss = 821.6919 Valid_loss = 811.7571\n",
      "Epoch =  3 Train_loss = 817.9802 Valid_loss = 806.1318\n",
      "Epoch =  4 Train_loss = 812.3134 Valid_loss = 798.1215\n",
      "Epoch =  5 Train_loss = 804.2396 Valid_loss = 787.2304\n",
      "Epoch =  6 Train_loss = 793.2580 Valid_loss = 772.9222\n",
      "Epoch =  7 Train_loss = 778.8276 Valid_loss = 754.6052\n",
      "Epoch =  8 Train_loss = 760.3483 Valid_loss = 731.6474\n",
      "Epoch =  9 Train_loss = 737.1770 Valid_loss = 703.3208\n",
      "Epoch =  10 Train_loss = 708.5998 Valid_loss = 668.9036\n",
      "Epoch =  11 Train_loss = 673.8574 Valid_loss = 627.5472\n",
      "Epoch =  12 Train_loss = 632.1080 Valid_loss = 578.8220\n",
      "Epoch =  13 Train_loss = 582.8416 Valid_loss = 533.1862\n",
      "Epoch =  14 Train_loss = 537.7017 Valid_loss = 513.4312\n",
      "Epoch =  15 Train_loss = 516.5355 Valid_loss = 493.6201\n",
      "Epoch =  16 Train_loss = 496.5858 Valid_loss = 463.9693\n",
      "Epoch =  17 Train_loss = 466.6065 Valid_loss = 424.0165\n",
      "Epoch =  18 Train_loss = 426.0808 Valid_loss = 375.9909\n",
      "Epoch =  19 Train_loss = 376.9592 Valid_loss = 326.1525\n",
      "Epoch =  20 Train_loss = 327.5884 Valid_loss = 284.6291\n",
      "Epoch =  21 Train_loss = 286.8219 Valid_loss = 261.7994\n",
      "Epoch =  22 Train_loss = 263.5315 Valid_loss = 253.0753\n",
      "Epoch =  23 Train_loss = 253.3926 Valid_loss = 247.5631\n",
      "Epoch =  24 Train_loss = 247.3575 Valid_loss = 241.4476\n",
      "Epoch =  25 Train_loss = 239.6775 Valid_loss = 232.8673\n",
      "Epoch =  26 Train_loss = 230.7878 Valid_loss = 226.9247\n",
      "Epoch =  27 Train_loss = 225.5182 Valid_loss = 224.0103\n",
      "Epoch =  28 Train_loss = 221.4930 Valid_loss = 217.3847\n",
      "Epoch =  29 Train_loss = 214.5577 Valid_loss = 207.6956\n",
      "Epoch =  30 Train_loss = 204.7758 Valid_loss = 196.9189\n",
      "Epoch =  31 Train_loss = 193.4143 Valid_loss = 186.8561\n",
      "Epoch =  32 Train_loss = 182.5956 Valid_loss = 178.1063\n",
      "Epoch =  33 Train_loss = 172.9515 Valid_loss = 170.3381\n",
      "Epoch =  34 Train_loss = 164.6296 Valid_loss = 164.0151\n",
      "Epoch =  35 Train_loss = 158.1772 Valid_loss = 159.8447\n",
      "Epoch =  36 Train_loss = 153.9064 Valid_loss = 158.8218\n",
      "Epoch =  37 Train_loss = 152.3357 Valid_loss = 159.4003\n",
      "Epoch =  38 Train_loss = 152.6544 Valid_loss = 158.6725\n",
      "Epoch =  39 Train_loss = 152.1607 Valid_loss = 155.2411\n",
      "Epoch =  40 Train_loss = 149.0918 Valid_loss = 148.9254\n",
      "Epoch =  41 Train_loss = 143.2165 Valid_loss = 140.9071\n",
      "Epoch =  42 Train_loss = 135.7235 Valid_loss = 132.7792\n",
      "Epoch =  43 Train_loss = 128.0376 Valid_loss = 125.7618\n",
      "Epoch =  44 Train_loss = 121.6852 Valid_loss = 120.4525\n",
      "Epoch =  45 Train_loss = 117.0151 Valid_loss = 116.8585\n",
      "Epoch =  46 Train_loss = 113.7359 Valid_loss = 114.3852\n",
      "Epoch =  47 Train_loss = 111.4763 Valid_loss = 112.9326\n",
      "Epoch =  48 Train_loss = 110.0800 Valid_loss = 112.5011\n",
      "Epoch =  49 Train_loss = 109.4826 Valid_loss = 112.5108\n",
      "Epoch =  50 Train_loss = 109.3575 Valid_loss = 112.0371\n",
      "Epoch =  51 Train_loss = 108.9120 Valid_loss = 110.2241\n",
      "Epoch =  52 Train_loss = 107.2674 Valid_loss = 107.1160\n",
      "Epoch =  53 Train_loss = 104.4419 Valid_loss = 103.5925\n",
      "Epoch =  54 Train_loss = 101.3027 Valid_loss = 100.4514\n",
      "Epoch =  55 Train_loss = 98.5575 Valid_loss = 97.8827\n",
      "Epoch =  56 Train_loss = 96.1585 Valid_loss = 95.6324\n",
      "Epoch =  57 Train_loss = 93.8497 Valid_loss = 93.6211\n",
      "Epoch =  58 Train_loss = 91.7569 Valid_loss = 92.1296\n",
      "Epoch =  59 Train_loss = 90.2253 Valid_loss = 91.1683\n",
      "Epoch =  60 Train_loss = 89.1677 Valid_loss = 90.3844\n",
      "Epoch =  61 Train_loss = 88.3506 Valid_loss = 89.4024\n",
      "Epoch =  62 Train_loss = 87.3215 Valid_loss = 88.2498\n",
      "Epoch =  63 Train_loss = 86.0557 Valid_loss = 87.1560\n",
      "Epoch =  64 Train_loss = 84.8295 Valid_loss = 86.0864\n",
      "Epoch =  65 Train_loss = 83.7029 Valid_loss = 84.7868\n",
      "Epoch =  66 Train_loss = 82.4778 Valid_loss = 83.2340\n",
      "Epoch =  67 Train_loss = 81.1367 Valid_loss = 81.8158\n",
      "Epoch =  68 Train_loss = 79.9292 Valid_loss = 80.6783\n",
      "Epoch =  69 Train_loss = 79.0696 Valid_loss = 79.7375\n",
      "Epoch =  70 Train_loss = 78.3252 Valid_loss = 78.7092\n",
      "Epoch =  71 Train_loss = 77.4891 Valid_loss = 77.6218\n",
      "Epoch =  72 Train_loss = 76.6153 Valid_loss = 76.6902\n",
      "Epoch =  73 Train_loss = 75.8520 Valid_loss = 75.8673\n",
      "Epoch =  74 Train_loss = 75.1283 Valid_loss = 74.9467\n",
      "Epoch =  75 Train_loss = 74.2782 Valid_loss = 73.9557\n",
      "Epoch =  76 Train_loss = 73.3938 Valid_loss = 73.1524\n",
      "Epoch =  77 Train_loss = 72.7062 Valid_loss = 72.4559\n",
      "Epoch =  78 Train_loss = 72.1154 Valid_loss = 71.7863\n",
      "Epoch =  79 Train_loss = 71.5077 Valid_loss = 71.2147\n",
      "Epoch =  80 Train_loss = 70.9771 Valid_loss = 70.7357\n",
      "Epoch =  81 Train_loss = 70.5179 Valid_loss = 70.2529\n",
      "Epoch =  82 Train_loss = 70.0443 Valid_loss = 69.6774\n",
      "Epoch =  83 Train_loss = 69.5161 Valid_loss = 69.0894\n",
      "Epoch =  84 Train_loss = 69.0043 Valid_loss = 68.5516\n",
      "Epoch =  85 Train_loss = 68.5324 Valid_loss = 68.0653\n",
      "Epoch =  86 Train_loss = 68.0519 Valid_loss = 67.6382\n",
      "Epoch =  87 Train_loss = 67.5815 Valid_loss = 67.2856\n",
      "Epoch =  88 Train_loss = 67.1714 Valid_loss = 66.9635\n",
      "Epoch =  89 Train_loss = 66.8206 Valid_loss = 66.5691\n",
      "Epoch =  90 Train_loss = 66.4715 Valid_loss = 66.1976\n",
      "Epoch =  91 Train_loss = 66.1495 Valid_loss = 65.8847\n",
      "Epoch =  92 Train_loss = 65.8677 Valid_loss = 65.6457\n",
      "Epoch =  93 Train_loss = 65.5814 Valid_loss = 65.4754\n",
      "Epoch =  94 Train_loss = 65.2945 Valid_loss = 65.3321\n",
      "Epoch =  95 Train_loss = 65.0507 Valid_loss = 65.1109\n",
      "Epoch =  96 Train_loss = 64.8062 Valid_loss = 64.8519\n",
      "Epoch =  97 Train_loss = 64.5663 Valid_loss = 64.6235\n",
      "Epoch =  98 Train_loss = 64.3629 Valid_loss = 64.4464\n",
      "Epoch =  99 Train_loss = 64.1635 Valid_loss = 64.3349\n"
     ]
    }
   ],
   "source": [
    "# Initialize Q network\n",
    "dqn = DQN_pytorch().cuda()\n",
    "print('CUDA for model: ', next(dqn.parameters()).is_cuda)\n",
    "\n",
    "def train_DQN(dqn, num_epoch, s_material_inputs_train, s_step_inputs_train, a_elem_inputs_train, a_comp_inputs_train, Q_targets_train ,s_material_inputs_valid ,s_step_inputs_valid ,a_elem_inputs_valid ,a_comp_inputs_valid ,Q_targets_valid):\n",
    "    '''\n",
    "    Trains DQN and prints: 1. training loss 2. validation loss\n",
    "\n",
    "    Args:\n",
    "    dqn: DQN\n",
    "    num_epoch: Int. Number of training epochs \n",
    "    s_material_inputs_train\n",
    "    s_step_inputs_train\n",
    "    a_elem_inputs_train\n",
    "    a_comp_inputs_train\n",
    "    Q_targets_train\n",
    "    s_material_inputs_valid\n",
    "    s_step_inputs_valid\n",
    "    a_elem_inputs_valid\n",
    "    a_comp_inputs_valid\n",
    "    Q_targets_valid\n",
    "\n",
    "    Returns:\n",
    "    Updated dqn\n",
    "    train_loss_list\n",
    "    valid_loss_list\n",
    "    '''\n",
    "    train_loss_list = []\n",
    "    valid_loss_list = []\n",
    "\n",
    "    # Train Q network\n",
    "    optimizer = torch.optim.Adam(dqn.parameters(), lr=1e-2)\n",
    "    loss_func = torch.nn.SmoothL1Loss().float()\n",
    "\n",
    "    # start = time.time()\n",
    "\n",
    "    for epoch in range(num_epoch):\n",
    "        # Train loss\n",
    "        prediction_train = dqn(s_material = s_material_inputs_train, # torch.Size([batch_size, 145]) and must be .float() !!!\n",
    "                        s_step      = s_step_inputs_train,     # torch.Size([batch_size,5])\n",
    "                        a_elem      = a_elem_inputs_train,     # torch.Size([batch_size,80])\n",
    "                        a_comp      = a_comp_inputs_train,  # torch.Size([batch_size,10])\n",
    "                        )                    # Input x and predict based on x\n",
    "        train_loss       = loss_func(prediction_train.float(), Q_targets_train.float()).float()   # Must be (1. nn output, 2. target)\n",
    "        optimizer.zero_grad()   # Clear gradients for next train\n",
    "        train_loss.backward()         # Backpropagation, compute gradients\n",
    "        optimizer.step()        # Apply gradients\n",
    "\n",
    "        # Validation loss\n",
    "        prediction_valid = dqn(s_material = s_material_inputs_valid, # torch.Size([batch_size, 145]) and must be .float() !!!\n",
    "                        s_step      = s_step_inputs_valid,     # torch.Size([batch_size,5])\n",
    "                        a_elem      = a_elem_inputs_valid,     # torch.Size([batch_size,80])\n",
    "                        a_comp      = a_comp_inputs_valid,  # torch.Size([batch_size,10]))\n",
    "                        )\n",
    "        valid_loss = loss_func(prediction_valid.float(), Q_targets_valid.float()).float()   # Must be (1. nn output, 2. target)\n",
    "        train_loss_list.append(train_loss.cpu().detach().numpy())\n",
    "        valid_loss_list.append(valid_loss.cpu().detach().numpy())\n",
    "\n",
    "        print('Epoch = ', epoch, 'Train_loss = %.4f' % train_loss.cpu().data.numpy(), 'Valid_loss = %.4f' % valid_loss.cpu().data.numpy())\n",
    "    # end = time.time()\n",
    "    # print('time taken for batch:', end - start)\n",
    "    \n",
    "    return dqn, train_loss_list, valid_loss_list\n",
    "dqn, train_loss_list, valid_loss_list = train_DQN(dqn =  dqn, num_epoch = 100, s_material_inputs_train = s_material_inputs_train, s_step_inputs_train = s_step_inputs_train,\n",
    " a_elem_inputs_train = a_elem_inputs_train, a_comp_inputs_train = a_comp_inputs_train, Q_targets_train = Q_targets_train,\n",
    " s_material_inputs_valid = s_material_inputs_valid, s_step_inputs_valid = s_step_inputs_valid, a_elem_inputs_valid = a_elem_inputs_valid,\n",
    " a_comp_inputs_valid = a_comp_inputs_valid, Q_targets_valid = Q_targets_valid)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# DCN"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Epoch =  0 Train_loss = 0.2458 Valid_loss = 0.2515\n",
      "Epoch =  1 Train_loss = 0.2462 Valid_loss = 0.2417\n",
      "Epoch =  2 Train_loss = 0.2349 Valid_loss = 0.2408\n",
      "Epoch =  3 Train_loss = 0.2326 Valid_loss = 0.2399\n",
      "Epoch =  4 Train_loss = 0.2304 Valid_loss = 0.2375\n",
      "Epoch =  5 Train_loss = 0.2280 Valid_loss = 0.2325\n",
      "Epoch =  6 Train_loss = 0.2254 Valid_loss = 0.2282\n",
      "Epoch =  7 Train_loss = 0.2231 Valid_loss = 0.2276\n",
      "Epoch =  8 Train_loss = 0.2211 Valid_loss = 0.2256\n",
      "Epoch =  9 Train_loss = 0.2190 Valid_loss = 0.2234\n",
      "Epoch =  10 Train_loss = 0.2165 Valid_loss = 0.2246\n",
      "Epoch =  11 Train_loss = 0.2136 Valid_loss = 0.2208\n",
      "Epoch =  12 Train_loss = 0.2109 Valid_loss = 0.2232\n",
      "Epoch =  13 Train_loss = 0.2088 Valid_loss = 0.2201\n",
      "Epoch =  14 Train_loss = 0.2066 Valid_loss = 0.2211\n",
      "Epoch =  15 Train_loss = 0.2040 Valid_loss = 0.2183\n",
      "Epoch =  16 Train_loss = 0.2015 Valid_loss = 0.2194\n",
      "Epoch =  17 Train_loss = 0.1996 Valid_loss = 0.2172\n",
      "Epoch =  18 Train_loss = 0.1980 Valid_loss = 0.2231\n",
      "Epoch =  19 Train_loss = 0.1980 Valid_loss = 0.2180\n",
      "Epoch =  20 Train_loss = 0.1983 Valid_loss = 0.2229\n",
      "Epoch =  21 Train_loss = 0.1947 Valid_loss = 0.2214\n",
      "Epoch =  22 Train_loss = 0.1927 Valid_loss = 0.2171\n",
      "Epoch =  23 Train_loss = 0.1931 Valid_loss = 0.2192\n",
      "Epoch =  24 Train_loss = 0.1892 Valid_loss = 0.2219\n",
      "Epoch =  25 Train_loss = 0.1891 Valid_loss = 0.2164\n",
      "Epoch =  26 Train_loss = 0.1874 Valid_loss = 0.2166\n",
      "Epoch =  27 Train_loss = 0.1849 Valid_loss = 0.2214\n",
      "Epoch =  28 Train_loss = 0.1849 Valid_loss = 0.2167\n",
      "Epoch =  29 Train_loss = 0.1819 Valid_loss = 0.2167\n",
      "Epoch =  30 Train_loss = 0.1803 Valid_loss = 0.2220\n",
      "Epoch =  31 Train_loss = 0.1799 Valid_loss = 0.2174\n",
      "Epoch =  32 Train_loss = 0.1772 Valid_loss = 0.2198\n",
      "Epoch =  33 Train_loss = 0.1747 Valid_loss = 0.2213\n",
      "Epoch =  34 Train_loss = 0.1734 Valid_loss = 0.2181\n",
      "Epoch =  35 Train_loss = 0.1725 Valid_loss = 0.2260\n",
      "Epoch =  36 Train_loss = 0.1718 Valid_loss = 0.2191\n",
      "Epoch =  37 Train_loss = 0.1695 Valid_loss = 0.2223\n",
      "Epoch =  38 Train_loss = 0.1662 Valid_loss = 0.2220\n",
      "Epoch =  39 Train_loss = 0.1636 Valid_loss = 0.2210\n",
      "Epoch =  40 Train_loss = 0.1620 Valid_loss = 0.2264\n",
      "Epoch =  41 Train_loss = 0.1621 Valid_loss = 0.2237\n",
      "Epoch =  42 Train_loss = 0.1639 Valid_loss = 0.2384\n",
      "Epoch =  43 Train_loss = 0.1683 Valid_loss = 0.2246\n",
      "Epoch =  44 Train_loss = 0.1568 Valid_loss = 0.2262\n",
      "Epoch =  45 Train_loss = 0.1600 Valid_loss = 0.2382\n",
      "Epoch =  46 Train_loss = 0.1615 Valid_loss = 0.2280\n",
      "Epoch =  47 Train_loss = 0.1531 Valid_loss = 0.2292\n",
      "Epoch =  48 Train_loss = 0.1565 Valid_loss = 0.2369\n",
      "Epoch =  49 Train_loss = 0.1539 Valid_loss = 0.2314\n",
      "Epoch =  50 Train_loss = 0.1502 Valid_loss = 0.2286\n",
      "Epoch =  51 Train_loss = 0.1520 Valid_loss = 0.2348\n",
      "Epoch =  52 Train_loss = 0.1481 Valid_loss = 0.2361\n",
      "Epoch =  53 Train_loss = 0.1472 Valid_loss = 0.2297\n",
      "Epoch =  54 Train_loss = 0.1466 Valid_loss = 0.2310\n",
      "Epoch =  55 Train_loss = 0.1419 Valid_loss = 0.2384\n",
      "Epoch =  56 Train_loss = 0.1434 Valid_loss = 0.2335\n",
      "Epoch =  57 Train_loss = 0.1412 Valid_loss = 0.2322\n",
      "Epoch =  58 Train_loss = 0.1377 Valid_loss = 0.2378\n",
      "Epoch =  59 Train_loss = 0.1399 Valid_loss = 0.2332\n",
      "Epoch =  60 Train_loss = 0.1363 Valid_loss = 0.2357\n",
      "Epoch =  61 Train_loss = 0.1347 Valid_loss = 0.2370\n",
      "Epoch =  62 Train_loss = 0.1330 Valid_loss = 0.2343\n",
      "Epoch =  63 Train_loss = 0.1343 Valid_loss = 0.2393\n",
      "Epoch =  64 Train_loss = 0.1312 Valid_loss = 0.2357\n",
      "Epoch =  65 Train_loss = 0.1297 Valid_loss = 0.2357\n",
      "Epoch =  66 Train_loss = 0.1265 Valid_loss = 0.2387\n",
      "Epoch =  67 Train_loss = 0.1270 Valid_loss = 0.2357\n",
      "Epoch =  68 Train_loss = 0.1263 Valid_loss = 0.2447\n",
      "Epoch =  69 Train_loss = 0.1270 Valid_loss = 0.2360\n",
      "Epoch =  70 Train_loss = 0.1254 Valid_loss = 0.2431\n",
      "Epoch =  71 Train_loss = 0.1238 Valid_loss = 0.2375\n",
      "Epoch =  72 Train_loss = 0.1212 Valid_loss = 0.2393\n",
      "Epoch =  73 Train_loss = 0.1174 Valid_loss = 0.2393\n",
      "Epoch =  74 Train_loss = 0.1168 Valid_loss = 0.2394\n",
      "Epoch =  75 Train_loss = 0.1158 Valid_loss = 0.2432\n",
      "Epoch =  76 Train_loss = 0.1156 Valid_loss = 0.2397\n",
      "Epoch =  77 Train_loss = 0.1159 Valid_loss = 0.2482\n",
      "Epoch =  78 Train_loss = 0.1160 Valid_loss = 0.2406\n",
      "Epoch =  79 Train_loss = 0.1179 Valid_loss = 0.2611\n",
      "Epoch =  80 Train_loss = 0.1266 Valid_loss = 0.2419\n",
      "Epoch =  81 Train_loss = 0.1214 Valid_loss = 0.2453\n",
      "Epoch =  82 Train_loss = 0.1102 Valid_loss = 0.2486\n",
      "Epoch =  83 Train_loss = 0.1119 Valid_loss = 0.2419\n",
      "Epoch =  84 Train_loss = 0.1103 Valid_loss = 0.2482\n",
      "Epoch =  85 Train_loss = 0.1115 Valid_loss = 0.2465\n",
      "Epoch =  86 Train_loss = 0.1058 Valid_loss = 0.2459\n",
      "Epoch =  87 Train_loss = 0.1056 Valid_loss = 0.2508\n",
      "Epoch =  88 Train_loss = 0.1094 Valid_loss = 0.2473\n",
      "Epoch =  89 Train_loss = 0.1016 Valid_loss = 0.2504\n",
      "Epoch =  90 Train_loss = 0.1047 Valid_loss = 0.2535\n",
      "Epoch =  91 Train_loss = 0.1048 Valid_loss = 0.2469\n",
      "Epoch =  92 Train_loss = 0.1010 Valid_loss = 0.2518\n",
      "Epoch =  93 Train_loss = 0.1011 Valid_loss = 0.2520\n",
      "Epoch =  94 Train_loss = 0.0991 Valid_loss = 0.2480\n",
      "Epoch =  95 Train_loss = 0.0995 Valid_loss = 0.2565\n",
      "Epoch =  96 Train_loss = 0.0987 Valid_loss = 0.2514\n",
      "Epoch =  97 Train_loss = 0.0951 Valid_loss = 0.2498\n",
      "Epoch =  98 Train_loss = 0.0963 Valid_loss = 0.2560\n",
      "Epoch =  99 Train_loss = 0.0949 Valid_loss = 0.2535\n"
     ]
    }
   ],
   "source": [
    "# Initialize constraint network\n",
    "dcn = DCN_pytorch().cuda()\n",
    "\n",
    "def train_DCN(dcn, num_epoch, s_material_inputs_train, s_step_inputs_train, a_elem_inputs_train, a_comp_inputs_train, Q_targets_train, en_targets_train, s_material_inputs_valid ,s_step_inputs_valid ,a_elem_inputs_valid ,a_comp_inputs_valid , Q_targets_valid, en_targets_valid):\n",
    "    '''\n",
    "    Trains DCN and prints: 1. training loss 2. validation loss\n",
    "\n",
    "    Args:\n",
    "    dcn: DCN\n",
    "    num_epoch: Int. Number of training epochs \n",
    "    s_material_inputs_train\n",
    "    s_step_inputs_train\n",
    "    a_elem_inputs_train\n",
    "    a_comp_inputs_train\n",
    "    Q_targets_train\n",
    "    en_targets_train\n",
    "    s_material_inputs_valid\n",
    "    s_step_inputs_valid\n",
    "    a_elem_inputs_valid\n",
    "    a_comp_inputs_valid\n",
    "    Q_targets_valid\n",
    "    en_targets_valid\n",
    "\n",
    "    Returns:\n",
    "    Updated dcn\n",
    "    train_loss_list\n",
    "    valid_loss_list\n",
    "    '''\n",
    "    train_loss_list = []\n",
    "    valid_loss_list = []\n",
    "\n",
    "    # Train Q network\n",
    "    optimizer = torch.optim.Adam(dcn.parameters(), lr=1e-2)\n",
    "    # loss_func = torch.nn.SmoothL1Loss().float()\n",
    "    loss_func = torch.nn.MSELoss().float() # Works better than L1Loss\n",
    "    # loss_func = torch.nn.L1Loss().float() # Trained model gives pred close to 0\n",
    "\n",
    "    # start = time.time()\n",
    "\n",
    "    for epoch in range(num_epoch):\n",
    "        # Train loss\n",
    "        prediction_train = dcn(s_material = s_material_inputs_train, # torch.Size([batch_size, 145]) and must be .float() !!!\n",
    "                        s_step      = s_step_inputs_train,     # torch.Size([batch_size,5])\n",
    "                        a_elem      = a_elem_inputs_train,     # torch.Size([batch_size,80])\n",
    "                        a_comp      = a_comp_inputs_train,  # torch.Size([batch_size,10])\n",
    "                        )                    # Input x and predict based on x\n",
    "        train_loss       = loss_func(prediction_train.float(), en_targets_train.float()).float()   # Must be (1. nn output, 2. target)\n",
    "        optimizer.zero_grad()   # Clear gradients for next train\n",
    "        train_loss.backward()         # Backpropagation, compute gradients\n",
    "        optimizer.step()        # Apply gradients\n",
    "\n",
    "        # Validation loss\n",
    "        prediction_valid = dcn(s_material = s_material_inputs_valid, # torch.Size([batch_size, 145]) and must be .float() !!!\n",
    "                        s_step      = s_step_inputs_valid,     # torch.Size([batch_size,5])\n",
    "                        a_elem      = a_elem_inputs_valid,     # torch.Size([batch_size,80])\n",
    "                        a_comp      = a_comp_inputs_valid,  # torch.Size([batch_size,10]))\n",
    "                        )\n",
    "        valid_loss = loss_func(prediction_valid.float(), en_targets_valid.float()).float()   # Must be (1. nn output, 2. target)\n",
    "        train_loss_list.append(train_loss.cpu().detach().numpy())\n",
    "        valid_loss_list.append(valid_loss.cpu().detach().numpy())\n",
    "\n",
    "        print('Epoch = ', epoch, 'Train_loss = %.4f' % train_loss.data.cpu().numpy(), 'Valid_loss = %.4f' % valid_loss.data.cpu().numpy())\n",
    "    # end = time.time()\n",
    "    # print('time taken for batch:', end - start)\n",
    "    \n",
    "    return dcn, train_loss_list, valid_loss_list\n",
    "dcn, train_loss_list, valid_loss_list = train_DCN(dcn =  dcn, num_epoch = 100, s_material_inputs_train = s_material_inputs_train, s_step_inputs_train = s_step_inputs_train,\n",
    " a_elem_inputs_train = a_elem_inputs_train, a_comp_inputs_train = a_comp_inputs_train, Q_targets_train = Q_targets_train, en_targets_train = en_targets_train,\n",
    " s_material_inputs_valid = s_material_inputs_valid, s_step_inputs_valid = s_step_inputs_valid, a_elem_inputs_valid = a_elem_inputs_valid,\n",
    " a_comp_inputs_valid = a_comp_inputs_valid, Q_targets_valid = Q_targets_valid, en_targets_valid = en_targets_valid)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 432x288 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "# Plot train and valid losses\n",
    "plt.figure()\n",
    "plt.plot(range(0,len(train_loss_list)), train_loss_list, label = 'Train')\n",
    "plt.plot(range(0,len(valid_loss_list)), valid_loss_list, label = 'Valid')\n",
    "plt.legend()\n",
    "# plt.ylim(130,170)\n",
    "plt.xlabel('Training epoch')\n",
    "plt.ylabel('Loss')\n",
    "# plt.savefig('plots/train_valid_loss.png', dpi = 300)\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Material: BaTiO3\n",
      "Q value of max action:  -877.0697631835938\n",
      "9\n",
      "\n",
      "Material: SrTiO3\n",
      "Q value of max action:  -858.019775390625\n"
     ]
    }
   ],
   "source": [
    "def ranked_max_actions(state, dqn, n_actions = len(element_set)*len(comp_set), oxide = False):\n",
    "    \"\"\"\n",
    "    Returns a list of ranked actions based on Q-network prodictions\n",
    "\n",
    "    Args:\n",
    "    state: List [s_material, s_step] where s_material is the string representation of material and s_step is the integer step number\n",
    "    dqn: Q-network for Q-value prediction\n",
    "    oxide: Choose oxygen only \n",
    "\n",
    "    Returns:\n",
    "    ranked_max_actions: List [[s_material, s_step, Q-value]\n",
    "                              ...\n",
    "                              [s_material, s_step, Q-value]\n",
    "                              ]\n",
    "    where s_material is the one-hot torch.tensor.float() of material and\n",
    "    s_step is the  one-hot torch.tensor.float()\n",
    "    0th index corresponds to highest Q-value\n",
    "    \"\"\"\n",
    "\n",
    "    s_material, s_step = state\n",
    "    # s_material = torch.tensor(onehot_target(s_material)).float() # Get one-hot rep of s_material\n",
    "    s_material = torch.tensor(featurize_target(s_material)).reshape(1, -1) # Get one-hot rep of s_material\n",
    "    s_material = torch.tensor(scaler.transform(s_material)).reshape(-1).float()  # Standardize s_material\n",
    "    s_step = torch.tensor(step_to_one_hot([s_step])[0]).float() # Get one-hot rep of s_step\n",
    "\n",
    "    if oxide: # If oxygen only, we only need smaller action space\n",
    "        n_actions = len(comp_set) - 1 # since adding 0 oxygen is NOT an option\n",
    "    s_material_input = torch.stack([s_material]*n_actions).to('cuda')\n",
    "    s_step_input = torch.stack([s_step]*n_actions).to('cuda')\n",
    "\n",
    "    a_elem_input = []\n",
    "    a_comp_input = []\n",
    "\n",
    "    if oxide: # Consider oxygen only\n",
    "        a_elem = torch.tensor(element_to_one_hot(['O'])[0]).float()\n",
    "        # for comp in comp_set:\n",
    "        for comp in comp_set[1:]: # consider only non-zero compositions for adding oxygen\n",
    "            a_comp = torch.tensor(comp_to_one_hot([comp])[0]).float()\n",
    "\n",
    "            a_elem_input.append(a_elem)\n",
    "            a_comp_input.append(a_comp)\n",
    "\n",
    "    else:\n",
    "        for elem in element_set:\n",
    "            a_elem = torch.tensor(element_to_one_hot([elem])[0]).float()\n",
    "            for comp in comp_set:\n",
    "                a_comp = torch.tensor(comp_to_one_hot([comp])[0]).float()\n",
    "\n",
    "                a_elem_input.append(a_elem)\n",
    "                a_comp_input.append(a_comp)\n",
    "    \n",
    "    a_elem_input = torch.stack(a_elem_input).float().to('cuda')\n",
    "    a_comp_input = torch.stack(a_comp_input).float().to('cuda')\n",
    "\n",
    "    \n",
    "    Q_pred = dqn(s_material = s_material_input, # torch.Size([batch_size, 40, 115]) and must be .float() !!!\n",
    "                    s_step      = s_step_input,     # torch.Size([batch_size,5])\n",
    "                    a_elem      = a_elem_input,     # torch.Size([batch_size,80])\n",
    "                    a_comp      = a_comp_input,  # torch.Size([batch_size,10])\n",
    "\n",
    "                )   \n",
    "    # Rank according to Q-values\n",
    "    Q_pred = Q_pred.cpu().detach().numpy()\n",
    "    Q_pred = Q_pred.reshape(n_actions)\n",
    "    order = Q_pred.argsort()\n",
    "    ranks = order.argsort() # ranks of Q-values, 0th index = lowest Q, last index = highest Q\n",
    "    \n",
    "    \n",
    "    a_elem_input_ranked = torch.zeros([n_actions, len(element_set)])\n",
    "    a_comp_input_ranked = torch.zeros([n_actions, len(comp_set)])\n",
    "    Q_pred_ranked = np.zeros(n_actions)\n",
    "\n",
    "    for rank in reversed(range(n_actions)): # Start with most valuable\n",
    "        rank_idx = ranks[rank] # rank_idx = final position\n",
    "        rank_idx = (n_actions - 1) - rank_idx # Reverse rank_idx i.e. Want 0th index to correspond to max Q\n",
    "\n",
    "        # Sort Q-values\n",
    "        Q = Q_pred[rank]\n",
    "        Q_pred_ranked[rank_idx] = Q\n",
    "\n",
    "        # Sort a_elem according to Q-values\n",
    "        a_elem = a_elem_input[rank]\n",
    "        a_elem_input_ranked[rank_idx] = a_elem\n",
    "\n",
    "        # Sort a_comp according to Q-values\n",
    "        a_comp = a_comp_input[rank]\n",
    "        a_comp_input_ranked[rank_idx] = a_comp\n",
    "\n",
    "    # print('Max Q: ', max(Q_pred_ranked))\n",
    "    # print('Min Q: ', min(Q_pred_ranked))\n",
    "\n",
    "    return a_elem_input_ranked, a_comp_input_ranked, Q_pred_ranked\n",
    "\n",
    "mat_1 = 'BaTiO3'\n",
    "mat_2 = 'SrTiO3'\n",
    "\n",
    "print('Material: {}'.format(mat_1))\n",
    "a_elem_input_ranked, a_comp_input_ranked, Q_pred_ranked = ranked_max_actions([mat_1, 1], dqn = dqn, oxide = True)\n",
    "print('Q value of max action: ', Q_pred_ranked[0])\n",
    "print(len(Q_pred_ranked))\n",
    "# print('max_a_elem: ', tuple(a_elem_input_ranked[0].tolist()))\n",
    "# print('max_a_comp: ', tuple(a_comp_input_ranked[0].tolist()))\n",
    "# print('')\n",
    "print('')\n",
    "\n",
    "print('Material: {}'.format(mat_2))\n",
    "a_elem_input_ranked, a_comp_input_ranked, Q_pred_ranked = ranked_max_actions([mat_2, 1], dqn = dqn, oxide = True)\n",
    "print('Q value of max action: ', Q_pred_ranked[0])\n",
    "# print('max_a_elem: ', tuple(a_elem_input_ranked[0].tolist()))\n",
    "# print('max_a_comp: ', tuple(a_comp_input_ranked[0].tolist()))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(tensor(0.0028, device='cuda:0', grad_fn=<MeanBackward0>),\n",
       " tensor(-0.0285, device='cuda:0', grad_fn=<MeanBackward0>),\n",
       " tensor(-0.0056, device='cuda:0', grad_fn=<MeanBackward0>),\n",
       " tensor(-0.0444, device='cuda:0', grad_fn=<MeanBackward0>))"
      ]
     },
     "execution_count": 10,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# Investigate weights of dqn\n",
    "# dqn.fc1.weight.mean(), dqn.fc2.weight.mean(), dqn.fc3.weight.mean(), dqn.fc4.weight.mean(),\n",
    "dcn.fc1.weight.mean(), dcn.fc2.weight.mean(), dcn.fc3.weight.mean(), dcn.fc4.weight.mean(),"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "tensor([[0.9994]], device='cuda:0', grad_fn=<SigmoidBackward>)\n"
     ]
    }
   ],
   "source": [
    "# For testing the DCN function\n",
    "# s_material = torch.tensor(onehot_target('BaTiO3'))\n",
    "s_material = torch.tensor(scaler.transform(torch.tensor(featurize_target('BaTiO3')).reshape(1,-1))).float().to('cuda')\n",
    "# s_material = s_material.reshape(1, s_material.shape[0])\n",
    "# print(s_material.shape)\n",
    "\n",
    "s_step = torch.zeros(5)\n",
    "s_step[3] = 1.\n",
    "s_step = s_step.reshape(1, s_step.shape[0]).to('cuda')\n",
    "# print(s_step.shape)\n",
    "\n",
    "a_elem = torch.zeros(78)\n",
    "a_elem[2] = 1.\n",
    "a_elem = a_elem.reshape(1, a_elem.shape[0]).to('cuda')\n",
    "# print(a_elem.shape)\n",
    "\n",
    "a_comp = torch.zeros(10)\n",
    "a_comp[1] = 1.\n",
    "a_comp = a_comp.reshape(1, a_comp.shape[0]).to('cuda')\n",
    "# print(a_comp.shape)\n",
    "\n",
    "output = dcn(s_material, s_step, a_elem, a_comp)\n",
    "\n",
    "print(output)\n",
    "# print(output.shape)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "'As2Bi7Ir2Cs2O1'"
      ]
     },
     "execution_count": 12,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "tasks = configs['tasks']\n",
    "env = ConstrainedMaterialEnvironment(element_set = element_set,\n",
    "                          comp_set =  comp_set,\n",
    "                          tasks = tasks)\n",
    "\n",
    "def generate_max_act(state, dqn, stochastic_top_frac = 0, oxide = False):\n",
    "    '''\n",
    "    Generates max action.\n",
    "    Args:\n",
    "    state: state\n",
    "    dqn: Q-network\n",
    "    stochastic_top_frac: Int (0-1). action choosen from the top n fraction of ranked actions. Default = 0. \n",
    "                         e.g. stochastic_top_frac = 0.1 corresponds to choosing from top 10% of action\n",
    "\n",
    "    Returns: max_a_elem, max_a_comp\n",
    "    '''\n",
    "    a_elem_input_ranked, a_comp_input_ranked, _ = ranked_max_actions(state = state, dqn = dqn, n_actions = len(element_set)*len(comp_set),\n",
    "                                                                    oxide = oxide)\n",
    "    \n",
    "    tot_num_actions = len(a_elem_input_ranked)\n",
    "    if stochastic_top_frac != 0: # if stochastic\n",
    "        act_idx = stochastic_top_frac*tot_num_actions # cut-off index\n",
    "        rank = np.random.randint(0,round(act_idx)) # take top stochastic_top_frac fraction of ranked actions\n",
    "    else: # deterministic\n",
    "        rank = 0 # take top action (index 0)\n",
    "    max_a_elem = a_elem_input_ranked[rank] # 0th index corresponds to max\n",
    "    max_a_comp = a_comp_input_ranked[rank] # 0th index corresponds to max\n",
    "    max_a_elem = tuple(max_a_elem.tolist())\n",
    "    max_a_comp = tuple(max_a_comp.tolist())\n",
    "    action = [max_a_elem, max_a_comp]\n",
    "\n",
    "    return action\n",
    "\n",
    "def generate_constrained_max_act(state, dqn, dcn, en_threshold, stochastic_top_frac = 0, oxide = False):\n",
    "    '''\n",
    "    Generates max action.\n",
    "    Args:\n",
    "    state: state\n",
    "    dqn: Q-network\n",
    "    en_threshold: Float (0.0 - 1.0). Determines how tight the constraint is. 1.0 being the tightest. \n",
    "    stochastic_top_frac: Int (0-1). action choosen from the top n fraction of ranked actions. Default = 0. \n",
    "                         e.g. stochastic_top_frac = 0.1 corresponds to choosing from top 10% of action\n",
    "\n",
    "    Returns: max_a_elem, max_a_comp\n",
    "    '''\n",
    "    a_elem_input_ranked, a_comp_input_ranked, _ = ranked_max_actions(state = state, dqn = dqn, n_actions = len(element_set)*len(comp_set),\n",
    "                                                                    oxide = oxide)\n",
    "    \n",
    "    tot_num_actions = len(a_elem_input_ranked)\n",
    "    en = 0\n",
    "\n",
    "    count = 0\n",
    "    while en <= en_threshold: # Keep iterating until predicted en goes above 0.5\n",
    "        if stochastic_top_frac != 0: # if stochastic\n",
    "            act_idx = stochastic_top_frac*tot_num_actions # cut-off index\n",
    "            rank = np.random.randint(0,round(act_idx)) # take top stochastic_top_frac fraction of ranked actions\n",
    "        else: # deterministic\n",
    "            rank = 0 # take top action (index 0)\n",
    "        max_a_elem = a_elem_input_ranked[rank] # 0th index corresponds to max\n",
    "        max_a_comp = a_comp_input_ranked[rank] # 0th index corresponds to max\n",
    "\n",
    "        # print('state:', state)\n",
    "        s_material, s_step = state # need to featurize/encode s_material and s_step\n",
    "        # s_material = torch.tensor(onehot_target(s_material)).float() # Get one-hot rep of s_material\n",
    "        s_material = torch.tensor(featurize_target(s_material)).reshape(1, -1) # Get one-hot rep of s_material\n",
    "        s_material_inputs = torch.tensor(scaler.transform(s_material)).float().to('cuda')  # Standardize s_material\n",
    "        s_step_inputs = torch.tensor(step_to_one_hot([s_step])[0]).reshape(1, -1).float().to('cuda') # Get one-hot rep of s_step\n",
    "        max_a_elem = max_a_elem.reshape(1, -1).to('cuda')\n",
    "        max_a_comp = max_a_comp.reshape(1, -1).to('cuda')\n",
    "        # print('state:', state)\n",
    "        # print('s_material_inputs:', s_material_inputs)\n",
    "        # print('s_step_inputs:', s_step_inputs)\n",
    "        # print('max_a_elem:', max_a_elem)\n",
    "        # print('max_a_comp:', max_a_comp)\n",
    "        en = dcn(s_material = s_material_inputs, # torch.Size([batch_size, 145]) and must be .float() !!!\n",
    "                        s_step      = s_step_inputs,     # torch.Size([batch_size,5])\n",
    "                        a_elem      = max_a_elem,     # torch.Size([batch_size,80])\n",
    "                        a_comp      = max_a_comp,  # torch.Size([batch_size,10])\n",
    "                        )  # Update en with prediction using DCN\n",
    "        # print('en pred:', en[0])\n",
    "        count += 1\n",
    "        if count >= 100: # Tried more than 100 times\n",
    "            print('Could not satisfy en constraint after %.0f actions' % count)\n",
    "            break\n",
    "    max_a_elem = max_a_elem[0]\n",
    "    max_a_comp = max_a_comp[0]\n",
    "    max_a_elem = tuple(max_a_elem.tolist())\n",
    "    max_a_comp = tuple(max_a_comp.tolist())\n",
    "    action = [max_a_elem, max_a_comp]\n",
    "\n",
    "    return action\n",
    "\n",
    "def generate_ep_with_dqn_dcn(dqn, dcn, en_threshold, max_steps = 5, epsilon = 0, stochastic_top_frac = 0, oxide = False):\n",
    "    '''\n",
    "    Generates an episode with trained DQN\n",
    "    \n",
    "    Args:\n",
    "    dqn: Q-network for Q-value prediction\n",
    "    dcn: Constraint-network for constraint value prediction\n",
    "    en_threshold: Float (0.0 - 1.0). Determines how tight the constraint is. 1.0 being the tightest.\n",
    "    max_steps: Int\n",
    "    epsilon: Float. Probability of taking random action \n",
    "    stochastic_top_frac: Int (0-1). action choosen from the top n fraction of ranked actions. Default = 0. \n",
    "                         e.g. stochastic_top_frac = 0.1 corresponds to choosing from top 10% of action\n",
    "\n",
    "    Returns: \n",
    "    env.path (an episode): List of SAR data in the form of [[material, step], [element, composition], reward]\n",
    "    final_compound: Str.\n",
    "    \n",
    "    '''\n",
    "    env.initialize()\n",
    "\n",
    "    for i in range(max_steps):\n",
    "\n",
    "        if np.random.uniform(0,1) > epsilon:\n",
    "            if i == max_steps-1: # If last step, choose oxygen only\n",
    "                action = generate_constrained_max_act(state = [env.state, env.counter+1], dqn = dqn, dcn = dcn, en_threshold = en_threshold, stochastic_top_frac = stochastic_top_frac,\n",
    "                                                oxide = oxide)\n",
    "                # action = generate_random_act(oxide = oxide)\n",
    "            else:\n",
    "                action = generate_constrained_max_act(state = [env.state, env.counter+1], dqn = dqn, dcn = dcn, en_threshold = en_threshold, stochastic_top_frac = stochastic_top_frac,\n",
    "                                                oxide = False)\n",
    "        else:\n",
    "            if i == max_steps-1: # If last step, choose oxygen only\n",
    "                action = generate_random_act(oxide = oxide)\n",
    "            else:\n",
    "                action = generate_random_act(oxide = False)\n",
    "        # print(action)\n",
    "        # Take step with action\n",
    "        env.step(action)\n",
    "        final_compound = env.state\n",
    "        # print('step:', env.counter)\n",
    "        # print('state:',final_compound)\n",
    "        # print('reward:',env.reward())\n",
    "        # print(env.num_steps_taken)\n",
    "        # print('')\n",
    "    return env.path, final_compound\n",
    "ep, final_compound = generate_ep_with_dqn_dcn(dqn = dqn, dcn = dcn, stochastic_top_frac = 0.15, en_threshold = 0., \n",
    "                                                oxide = True)\n",
    "\n",
    "# compounds = []\n",
    "# for i in tqdm(range(100)):\n",
    "#     _, final_compound = generate_ep_with_dqn_dcn(dqn = dqn, dcn = dcn, en_threshold = 0.0, stochastic_top_frac = 0.20)\n",
    "#     compounds.append(final_compound)\n",
    "\n",
    "# print(final_compound)\n",
    "# print(check_electronegativity(final_compound))\n",
    "# len(compounds)\n",
    "final_compound"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "metadata": {},
   "outputs": [],
   "source": [
    "# ========= FOR TRAINED POLICY ===========\n",
    "def generate_Q_c_datapoints_and_evaluate_policy(dqn, dcn, num_eps, epsilon, en_threshold = 0.4, stochastic_top_frac = 0.20, oxide = False):\n",
    "    '''\n",
    "    Generates 1) Q datapoints for training 2) List of final rewards using a DQN.\n",
    "\n",
    "    Args:\n",
    "    dqn: Q-netowrk\n",
    "    num_eps: Int. Number of episodes to run.\n",
    "    epsilon: Float. Probability of taking a random action\n",
    "    en_threshold: Float (0.0 - 1.0). Determines how tight the constraint is. 1.0 being the tightest. Default 0.4\n",
    "    stochastic_top_frac: Int (0-1). action choosen from the top n fraction of ranked actions. Default = 0. \n",
    "                         e.g. stochastic_top_frac = 0.1 corresponds to choosing from top 10% of action\n",
    "\n",
    "    Returns:\n",
    "    Q_c_data_trained: List of Q datapoints for training\n",
    "    en_targets_trained: List of en datapoints for training\n",
    "    final_rewards_trained: List of corresponding final rewards\n",
    "    final_compounds: List (of Str) of final compounds\n",
    "    '''\n",
    "    # start = time.time()\n",
    "\n",
    "    # Generate episodes with DQN\n",
    "    episodes = []\n",
    "    final_compounds = []\n",
    "    for j in range(num_eps):\n",
    "        episode, final_compound = generate_ep_with_dqn_dcn(dqn = dqn, dcn = dcn, epsilon = epsilon, en_threshold = en_threshold, stochastic_top_frac = stochastic_top_frac,\n",
    "                                                        oxide = oxide)\n",
    "        episodes.append(episode)\n",
    "        final_compounds.append(final_compound)\n",
    "\n",
    "    Q_c_data_trained = []\n",
    "    # Extract Q_data from episodes\n",
    "    for episode in episodes:\n",
    "        Q_c_data = extract_data_from_ep(episode)\n",
    "        Q_c_data_trained.append(Q_c_data)\n",
    "\n",
    "    # end = time.time()\n",
    "    # print('time taken:', end - start)\n",
    "\n",
    "    # Save Q_data\n",
    "    # with open('./data/Q_data_trained_RF.pkl', 'wb') as f:\n",
    "    #     pickle.dump(Q_data_trained, f, pickle.HIGHEST_PROTOCOL)\n",
    "\n",
    "    # # Load Q_data_random\n",
    "    # with open('./data/Q_data_trained_RF.pkl', 'rb') as f:\n",
    "    #     Q_data_trained = pickle.load(f)\n",
    "\n",
    "    # print('Number of episodes generated by agent trained on random policy data:', len(Q_data_trained))\n",
    "\n",
    "    # Extract final rewards from episode data\n",
    "    Q_targets_trained         = []\n",
    "    en_targets_trained         = []\n",
    "\n",
    "    for episode in Q_c_data_trained:\n",
    "        for step in episode:\n",
    "            # state, action, reward = step\n",
    "            _, _, Q_target, en_target = step\n",
    "            Q_target = torch.tensor([Q_target])\n",
    "            Q_targets_trained.append(Q_target)\n",
    "            en_target = torch.tensor([en_target])\n",
    "            en_targets_trained.append(en_target)\n",
    "\n",
    "    Q_targets_trained = torch.stack(Q_targets_trained)\n",
    "    en_targets_trained = torch.stack(en_targets_trained)\n",
    "    # print('Q_targets:', Q_targets_trained.shape)\n",
    "    final_rewards_trained = Q_targets_trained[0::5].reshape(-1) # Find final rewards, index of 5 since length of each episode is 5\n",
    "    final_en_trained = en_targets_trained[0::5].reshape(-1)\n",
    "\n",
    "    return Q_c_data_trained, en_targets_trained, final_rewards_trained, final_compounds\n",
    "\n",
    "# Q_c_data_trained, en_targets_trained, final_rewards_trained, final_compounds = \\\n",
    "# generate_Q_c_datapoints_and_evaluate_policy(dqn = dqn, dcn = dcn, num_eps = 50, epsilon = 0, en_threshold = 0.4, \n",
    "#                                         stochastic_top_frac = 0.15,\n",
    "#                                         oxide = True)\n",
    "# print(len(Q_c_data_trained))\n",
    "# print(len(en_targets_trained))\n",
    "# print(len(final_rewards_trained))\n",
    "# print(final_compounds)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "metadata": {},
   "outputs": [],
   "source": [
    "# count = 0\n",
    "# for target in en_targets_trained[::5]:\n",
    "#     if target[0] == 1.:\n",
    "#         count += 1\n",
    "# print(count)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "As shown in the plot below, an agent trained based on just data generated by a random policy can improve the distribution of rewards."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "metadata": {},
   "outputs": [],
   "source": [
    "# # Plot distribution of final rewards by random vs trained policy\n",
    "# plt.figure(figsize=(10,7))\n",
    "# sns.distplot(final_rewards_random, hist = True, norm_hist = True, label = 'Random')\n",
    "# sns.distplot(final_rewards_trained, hist = True, norm_hist = True, label = 'Trained')\n",
    "# plt.ylabel('Probability density')\n",
    "# plt.xlabel('Reward')\n",
    "# # plt.ylim(0,0.0025)\n",
    "# plt.xlim(-2000,0)\n",
    "# plt.legend()\n",
    "# # plt.savefig(fname = 'plots/reward_dist.png', dpi = 300)\n",
    "# plt.show()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 16,
   "metadata": {},
   "outputs": [],
   "source": [
    "# # Plot distribution of en by random policy\n",
    "# sns.distplot(en_targets_trained, color= 'grey')\n",
    "# plt.xlabel('Electronegativity (1.0 = OK, 0.0 = not OK)')\n",
    "# plt.show()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Deep Reinforcement Learning\n",
    "We now close the RL loop by training the agent on data generated by increasingly greedy policies. "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 17,
   "metadata": {},
   "outputs": [],
   "source": [
    "# replay_buffer = collections.deque(maxlen = 50000)\n",
    "# Q_c_data_trained, en_targets_trained, final_rewards_trained, final_compounds = generate_Q_c_datapoints_and_evaluate_policy(dqn = dqn, \n",
    "#                                                                                                                            dcn = dcn,\n",
    "#                                                                                                     num_eps = 100, \n",
    "#                                                                                                     epsilon = 0, \n",
    "#                                                                                                     en_threshold = 0.4,\n",
    "#                                                                                                     stochastic_top_frac = 0.2)\n",
    "# s_material_inputs, s_step_inputs, a_elem_inputs, a_comp_inputs, Q_targets, en_targets = generate_dqn_dcn_inputs_and_targets_from_Q_c_data(Q_c_data = Q_c_data_trained)\n",
    "\n",
    "def update_replay_buffer(replay_buffer, s_material_inputs, s_step_inputs, a_elem_inputs, a_comp_inputs, Q_targets, en_targets):\n",
    "    '''\n",
    "    Adds inputs into replay buffer\n",
    "    Args:\n",
    "    replay_buffer: collections.deque\n",
    "    \n",
    "    Returns: \n",
    "    updated replay buffer\n",
    "    \n",
    "    '''\n",
    "    for s_material, s_step, a_elem, a_comp, Q_target, en_target in zip(s_material_inputs, s_step_inputs, a_elem_inputs, a_comp_inputs, Q_targets, en_targets):\n",
    "        Q_c_datapoint = [s_material, s_step, a_elem, a_comp, Q_target, en_target]\n",
    "        replay_buffer.extend([Q_c_datapoint])\n",
    "    \n",
    "    return replay_buffer\n",
    "\n",
    "# replay_buffer = update_replay_buffer(replay_buffer = replay_buffer, s_material_inputs = s_material_inputs, \n",
    "#                                                             s_step_inputs = s_step_inputs, \n",
    "#                                                             a_elem_inputs = a_elem_inputs, \n",
    "#                                                             a_comp_inputs = a_comp_inputs, \n",
    "#                                                                 Q_targets = Q_targets,\n",
    "#                                                                 en_targets = en_targets)\n",
    "# print('Number of Q datapoints in replay buffer (5 per episode):', len(replay_buffer))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 18,
   "metadata": {},
   "outputs": [],
   "source": [
    "def sample_replay_buffer(replay_buffer, n = 100):\n",
    "    '''\n",
    "    Samples datapoints from replay buffer for DQN training\n",
    "\n",
    "    Args:\n",
    "    replay_buffer: collections.deque\n",
    "    n: Int. Number of datapoints to sample\n",
    "\n",
    "    Returns:\n",
    "    s_material_inputs: torch.tensor of Size([num of episodes in Q_data, 145])\n",
    "    s_step_inputs: torch.tensor\n",
    "    a_elem_inputs: torch.tensor\n",
    "    a_comp_inputs: torch.tensor\n",
    "    Q_targets: torch.tensor\n",
    "    em_targets: torch.tensor\n",
    "\n",
    "    '''\n",
    "\n",
    "    s_material_inputs = []\n",
    "    s_step_inputs     = []\n",
    "    a_elem_inputs     = []\n",
    "    a_comp_inputs     = []\n",
    "    Q_targets         = []\n",
    "    en_targets         = []\n",
    "\n",
    "    for Q_c_datapoint in random.sample(replay_buffer, n):\n",
    "        s_material, s_step, a_elem, a_comp, Q_target, en_target = Q_c_datapoint\n",
    "\n",
    "        s_material = torch.tensor(s_material).float()\n",
    "        s_step = torch.tensor(s_step)\n",
    "        a_elem = torch.tensor(a_elem)\n",
    "        a_comp = torch.tensor(a_comp)\n",
    "        Q_target = torch.tensor([Q_target])\n",
    "        en_target = torch.tensor([en_target])\n",
    "\n",
    "        s_material_inputs.append(s_material)\n",
    "        s_step_inputs.append(s_step)\n",
    "        a_elem_inputs.append(a_elem)\n",
    "        a_comp_inputs.append(a_comp)\n",
    "        Q_targets.append(Q_target)\n",
    "        en_targets.append(en_target)\n",
    "\n",
    "    s_material_inputs = torch.stack(s_material_inputs).to('cuda')\n",
    "    s_step_inputs = torch.stack(s_step_inputs).to('cuda')\n",
    "    a_elem_inputs = torch.stack(a_elem_inputs).to('cuda')\n",
    "    a_comp_inputs = torch.stack(a_comp_inputs).to('cuda')\n",
    "    Q_targets = torch.stack(Q_targets).to('cuda')\n",
    "    en_targets = torch.stack(en_targets).to('cuda')\n",
    "\n",
    "    return s_material_inputs, s_step_inputs, a_elem_inputs, a_comp_inputs, Q_targets, en_targets\n",
    "\n",
    "# s_material_inputs, s_step_inputs, a_elem_inputs, a_comp_inputs, Q_targets, en_targets = sample_replay_buffer(replay_buffer = replay_buffer, n = 100)\n",
    "\n",
    "# print('s_material_inputs:', s_material_inputs.shape)\n",
    "# print('s_step_inputs:', s_step_inputs.shape)\n",
    "# print('a_elem_inputs:', a_elem_inputs.shape)\n",
    "# print('a_comp_inputs:', a_comp_inputs.shape)\n",
    "# print('Q_targets:', Q_targets.shape)\n",
    "# print('en_targets:', en_targets.shape)\n",
    "    "
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Training the DQN"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 19,
   "metadata": {},
   "outputs": [],
   "source": [
    "# num_iter = 500 # No. of training iterations\n",
    "# epsilon  = 0.99   # Initial epsilon\n",
    "\n",
    "# en_threshold =  0 # Because we don't want to train DCN online\n",
    "# # en_threshold_ub = 0.5 # upper bound for en_threshold\n",
    "# # iteration = np.arange(num_iter)\n",
    "# # en_threshold_vs_iter = -en_threshold_ub*(np.e**(-iteration/(num_iter*0.3))) + en_threshold_ub\n",
    "\n",
    "# replay_buffer = collections.deque(maxlen = 50000) # Initialize replay buffer\n",
    "# dqn_save_freq = 10 # Save DQN model every n iterations\n",
    "# dqn = DQN_pytorch().cuda() # Initialize DQN\n",
    "# # dcn = DCN_pytorch().cuda() # Initialize DCN\n",
    "# env = ConstrainedMaterialEnvironment(element_set = element_set, comp_set = comp_set,) # Initialize env\n",
    "\n",
    "# # Meta-data to store\n",
    "# final_rewards_vs_iter   = []\n",
    "# en_vs_iter              = []\n",
    "# compounds_vs_iter       = []\n",
    "# train_loss_list_vs_iter = []\n",
    "# valid_loss_list_vs_iter = []\n",
    "# epsilon_vs_iter         = []\n",
    "\n",
    "# for iteration in tqdm(range(num_iter)):\n",
    "#     print('=========== TRAINING ITERATION %.0f ===========' % iteration)\n",
    "# #     en_threshold = en_threshold_vs_iter[iteration]\n",
    "#     print('Current epsilon = ', epsilon)\n",
    "#     print('Current en threshold = ', en_threshold)\n",
    "#     epsilon_vs_iter.append(epsilon)\n",
    "#     # 1. GENERATE Q DATAPOINTS & ADD TO REPLAY BUFFER\n",
    "#     print('A) GENERATING DATAPOINTS')\n",
    "#     Q_c_data_trained, en_targets_trained, final_rewards_trained, final_compounds = \\\n",
    "#         generate_Q_c_datapoints_and_evaluate_policy(dqn = dqn, dcn = dcn, num_eps = 100, epsilon = epsilon, en_threshold = en_threshold,\n",
    "#                                                                                                     stochastic_top_frac = 0.15,\n",
    "#                                                                                                     oxide = True) # Generate episodes and score them\n",
    "#     final_rewards_vs_iter.append(final_rewards_trained) # Save list of final reward\n",
    "#     en_vs_iter.append(en_targets_trained)\n",
    "#     compounds_vs_iter.append(final_compounds) # Save list of compounds generated at the end of each episode\n",
    "#     print('Average final reward = ', np.mean(np.array(final_rewards_trained)))\n",
    "#     print('Compounds generated:', final_compounds)\n",
    "#     s_material_inputs, s_step_inputs, a_elem_inputs, a_comp_inputs, Q_targets, en_targets = generate_dqn_dcn_inputs_and_targets_from_Q_c_data(Q_c_data = Q_c_data_trained) # Processing to give inputs and targets for DQN\n",
    "#     replay_buffer = update_replay_buffer(replay_buffer = replay_buffer, s_material_inputs = s_material_inputs, \n",
    "#                                                                 s_step_inputs = s_step_inputs, \n",
    "#                                                                 a_elem_inputs = a_elem_inputs, \n",
    "#                                                                 a_comp_inputs = a_comp_inputs, \n",
    "#                                                                     Q_targets = Q_targets,\n",
    "#                                                                    en_targets = en_targets)\n",
    "#     # 2. SAMPLE FROM REPLAY BUFFER, STANDARDIZE, TRAIN-VALIDATION SPLIT, TRAIN Q-NETWORK\n",
    "#     print('B) TRAINING Q-NETWORK and C-NETWORK')\n",
    "#     s_material_inputs, s_step_inputs, a_elem_inputs, a_comp_inputs, Q_targets, en_targets = sample_replay_buffer(replay_buffer = replay_buffer, n = 100)\n",
    "#     s_material_inputs_scaled = torch.tensor(scaler.transform(s_material_inputs.cpu())).float().to('cuda') # Standardize s_material_inputs\n",
    "#     s_material_inputs_train, s_step_inputs_train, a_elem_inputs_train, a_comp_inputs_train, Q_targets_train, en_targets_train, s_material_inputs_valid, s_step_inputs_valid, a_elem_inputs_valid, a_comp_inputs_valid, Q_targets_valid, en_targets_valid \\\n",
    "#     = train_validation_split(s_material_inputs_scaled = s_material_inputs_scaled, \n",
    "#                                         s_step_inputs = s_step_inputs, \n",
    "#                                         a_elem_inputs = a_elem_inputs,\n",
    "#                                         a_comp_inputs = a_comp_inputs,\n",
    "#                                         Q_targets = Q_targets,\n",
    "#                                         en_targets = en_targets,\n",
    "#                                         valid_size = 0.2) # Train-validation split of datapoints sampled from replay buffer\n",
    "#     dqn, train_loss_list, valid_loss_list = train_DQN(dqn =  dqn, num_epoch = 100, s_material_inputs_train = s_material_inputs_train, s_step_inputs_train = s_step_inputs_train,\n",
    "#     a_elem_inputs_train = a_elem_inputs_train, a_comp_inputs_train = a_comp_inputs_train, Q_targets_train = Q_targets_train,\n",
    "#     s_material_inputs_valid = s_material_inputs_valid, s_step_inputs_valid = s_step_inputs_valid, a_elem_inputs_valid = a_elem_inputs_valid,\n",
    "#     a_comp_inputs_valid = a_comp_inputs_valid, Q_targets_valid = Q_targets_valid) # Train DQN\n",
    "    \n",
    "# #     dcn, train_loss_list, valid_loss_list = train_DCN(dcn =  dcn, num_epoch = 100, s_material_inputs_train = s_material_inputs_train, s_step_inputs_train = s_step_inputs_train,\n",
    "# #     a_elem_inputs_train = a_elem_inputs_train, a_comp_inputs_train = a_comp_inputs_train, Q_targets_train = Q_targets_train, en_targets_train = en_targets_train,\n",
    "# #     s_material_inputs_valid = s_material_inputs_valid, s_step_inputs_valid = s_step_inputs_valid, a_elem_inputs_valid = a_elem_inputs_valid,\n",
    "# #     a_comp_inputs_valid = a_comp_inputs_valid, Q_targets_valid = Q_targets_valid, en_targets_valid = en_targets_valid)\n",
    "# #     train_loss_list_vs_iter.append(train_loss_list) # Save list of training loss vs training epoch for this iter\n",
    "# #     valid_loss_list_vs_iter.append(valid_loss_list) # Save list of validation loss vs training epoch for this iter\n",
    "\n",
    "#     if iteration in range(0, num_iter+1, dqn_save_freq): # Save DQN model every n iterations\n",
    "#         torch.save(dqn, './dqn_models/oxides_calcine/dqn_test_%.0fiter_RF_constrained' % iteration)\n",
    "\n",
    "# #     if iteration in range(0, num_iter+1, dqn_save_freq): # Save DCN model every n iterations\n",
    "# #         torch.save(dcn, './dcn_models/constrained_DQN_cuda/dcn_test_%.0fiter_RF_constrained' % iteration)\n",
    "\n",
    "#     epsilon *= 0.99 # Decay epsilon\n",
    "\n",
    "# with open('./training_data/oxides_calcine/final_rewards_vs_iter_RF_constrained.pkl', 'wb') as f:\n",
    "#         pickle.dump(final_rewards_vs_iter, f, pickle.HIGHEST_PROTOCOL)\n",
    "\n",
    "# with open('./training_data/oxides_calcine/en_vs_iter_RF_constrained.pkl', 'wb') as f:\n",
    "#         pickle.dump(en_vs_iter, f, pickle.HIGHEST_PROTOCOL)\n",
    "\n",
    "# with open('./training_data/oxides_calcine/compounds_vs_iter_RF_constrained.pkl', 'wb') as f:\n",
    "#         pickle.dump(compounds_vs_iter, f, pickle.HIGHEST_PROTOCOL)\n",
    "\n",
    "# with open('./training_data/oxides_calcine/train_loss_list_vs_iter_RF_constrained.pkl', 'wb') as f:\n",
    "#         pickle.dump(train_loss_list_vs_iter, f, pickle.HIGHEST_PROTOCOL)\n",
    "\n",
    "# with open('./training_data/oxides_calcine/valid_loss_list_vs_iter_RF_constrained.pkl', 'wb') as f:\n",
    "#         pickle.dump(valid_loss_list_vs_iter, f, pickle.HIGHEST_PROTOCOL)\n",
    "\n",
    "# with open('./training_data/oxides_calcine/epsilon_vs_iter_RF_constrained.pkl', 'wb') as f:\n",
    "#         pickle.dump(epsilon_vs_iter, f, pickle.HIGHEST_PROTOCOL)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 20,
   "metadata": {},
   "outputs": [],
   "source": [
    "# torch.save(dqn, './dqn_models/oxides_calcine/dqn_test_500iter_RF_constrained')"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Applying constraints: Train DCN on random data\n",
    "After DQN has been trained, we now train a separate neural network (DCN) to serve as a mask on top of DQN actions. This biases the actions toward those that are likely to give valid compounds (eg. correct electronegativity). \n",
    "\n",
    "We do this instead of training the DCN within the RL as this method performs better (higher validity) and, more importantly, is much faster (approx. 5x less run time)."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 21,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Number of episodes generated by random policy: 1000\n",
      "before scaling: tensor([[3.0000e+00, 6.7823e-01, 6.2488e-01,  ..., 0.0000e+00, 5.5973e-02,\n",
      "         9.8478e-03],\n",
      "        [3.0000e+00, 6.7823e-01, 6.2488e-01,  ..., 0.0000e+00, 5.5973e-02,\n",
      "         9.8478e-03],\n",
      "        [2.0000e+00, 8.6897e-01, 8.5846e-01,  ..., 0.0000e+00, 6.2480e-04,\n",
      "         7.6507e-05],\n",
      "        ...,\n",
      "        [2.0000e+00, 7.1429e-01, 6.4256e-01,  ..., 0.0000e+00, 6.3796e-03,\n",
      "         1.5623e-03],\n",
      "        [1.0000e+00, 1.0000e+00, 1.0000e+00,  ..., 1.0000e+00, 0.0000e+00,\n",
      "         0.0000e+00],\n",
      "        [0.0000e+00, 0.0000e+00, 0.0000e+00,  ..., 0.0000e+00, 0.0000e+00,\n",
      "         0.0000e+00]])\n",
      "after scaling: tensor([[ 0.9094,  0.2304,  0.1828,  ..., -0.5626, -0.3777, -0.3565],\n",
      "        [ 0.9094,  0.2304,  0.1828,  ..., -0.5626, -0.3777, -0.3565],\n",
      "        [ 0.1566,  0.7751,  0.8540,  ..., -0.5626, -0.6569, -0.6213],\n",
      "        ...,\n",
      "        [ 0.1566,  0.3333,  0.2336,  ..., -0.5626, -0.6279, -0.5811],\n",
      "        [-0.5963,  1.1493,  1.2607,  ...,  1.7776, -0.6601, -0.6234],\n",
      "        [-1.3491, -1.7066, -1.6129,  ..., -0.5626, -0.6601, -0.6234]],\n",
      "       device='cuda:0')\n",
      "TRAIN SIZES:\n",
      "s_material_inputs_train: torch.Size([4000, 145])\n",
      "s_step_inputs_train: torch.Size([4000, 5])\n",
      "a_elem_inputs_train: torch.Size([4000, 78])\n",
      "a_comp_inputs_train: torch.Size([4000, 10])\n",
      "Q_targets_train: torch.Size([4000, 1])\n",
      "en_targets_train: torch.Size([4000, 1])\n",
      "\n",
      "VALIDATION SIZES:\n",
      "s_material_inputs_valid: torch.Size([1000, 145])\n",
      "s_step_inputs_valid: torch.Size([1000, 5])\n",
      "a_elem_inputs_valid: torch.Size([1000, 78])\n",
      "a_comp_inputs_valid: torch.Size([1000, 10])\n",
      "Q_targets_valid: torch.Size([1000, 1])\n",
      "en_targets_valid: torch.Size([1000, 1])\n",
      "Epoch =  0 Train_loss = 0.2479 Valid_loss = 0.2594\n",
      "Epoch =  1 Train_loss = 0.2569 Valid_loss = 0.2388\n",
      "Epoch =  2 Train_loss = 0.2369 Valid_loss = 0.2381\n",
      "Epoch =  3 Train_loss = 0.2368 Valid_loss = 0.2360\n",
      "Epoch =  4 Train_loss = 0.2347 Valid_loss = 0.2324\n",
      "Epoch =  5 Train_loss = 0.2305 Valid_loss = 0.2317\n",
      "Epoch =  6 Train_loss = 0.2287 Valid_loss = 0.2323\n",
      "Epoch =  7 Train_loss = 0.2282 Valid_loss = 0.2311\n",
      "Epoch =  8 Train_loss = 0.2256 Valid_loss = 0.2296\n",
      "Epoch =  9 Train_loss = 0.2229 Valid_loss = 0.2281\n",
      "Epoch =  10 Train_loss = 0.2209 Valid_loss = 0.2269\n",
      "Epoch =  11 Train_loss = 0.2193 Valid_loss = 0.2267\n",
      "Epoch =  12 Train_loss = 0.2177 Valid_loss = 0.2267\n",
      "Epoch =  13 Train_loss = 0.2152 Valid_loss = 0.2265\n",
      "Epoch =  14 Train_loss = 0.2122 Valid_loss = 0.2254\n",
      "Epoch =  15 Train_loss = 0.2092 Valid_loss = 0.2242\n",
      "Epoch =  16 Train_loss = 0.2066 Valid_loss = 0.2238\n",
      "Epoch =  17 Train_loss = 0.2037 Valid_loss = 0.2234\n",
      "Epoch =  18 Train_loss = 0.2011 Valid_loss = 0.2250\n",
      "Epoch =  19 Train_loss = 0.1993 Valid_loss = 0.2268\n"
     ]
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 432x288 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "# Load Q_data_random\n",
    "with open('./data/oxides_calcine/Q_data_random.pkl', 'rb') as f:\n",
    "    Q_c_data_random = pickle.load(f)\n",
    "\n",
    "print('Number of episodes generated by random policy:',len(Q_c_data_random))\n",
    "\n",
    "s_material_inputs, s_step_inputs, a_elem_inputs, a_comp_inputs, Q_targets, en_targets = generate_dqn_dcn_inputs_and_targets_from_Q_c_data(Q_c_data = Q_c_data_random)\n",
    "\n",
    "# Standardizing s_material data\n",
    "scaler = StandardScaler() # Fit scaler\n",
    "print('before scaling:', s_material_inputs)\n",
    "scaler.fit(s_material_inputs)\n",
    "s_material_inputs_scaled = torch.tensor(scaler.transform(s_material_inputs)).float().to('cuda')\n",
    "print('after scaling:', s_material_inputs_scaled)\n",
    "\n",
    "# Train-validation split\n",
    "s_material_inputs_train, s_step_inputs_train, a_elem_inputs_train, a_comp_inputs_train, Q_targets_train, en_targets_train, s_material_inputs_valid, s_step_inputs_valid, a_elem_inputs_valid, a_comp_inputs_valid, Q_targets_valid, en_targets_valid \\\n",
    "= train_validation_split(s_material_inputs_scaled = s_material_inputs_scaled.to('cuda'), \n",
    "                                    s_step_inputs = s_step_inputs.to('cuda'), \n",
    "                                    a_elem_inputs = a_elem_inputs.to('cuda'),\n",
    "                                    a_comp_inputs = a_comp_inputs.to('cuda'),\n",
    "                                    Q_targets  = Q_targets.to('cuda'),\n",
    "                                    en_targets = en_targets.to('cuda'),\n",
    "                                    valid_size = 0.2)\n",
    "\n",
    "print('TRAIN SIZES:')\n",
    "print('s_material_inputs_train:', s_material_inputs_train.shape)\n",
    "print('s_step_inputs_train:', s_step_inputs_train.shape)\n",
    "print('a_elem_inputs_train:', a_elem_inputs_train.shape)\n",
    "print('a_comp_inputs_train:', a_comp_inputs_train.shape)\n",
    "print('Q_targets_train:', Q_targets_train.shape)\n",
    "print('en_targets_train:', en_targets_train.shape)\n",
    "print('')\n",
    "print('VALIDATION SIZES:')\n",
    "print('s_material_inputs_valid:', s_material_inputs_valid.shape)\n",
    "print('s_step_inputs_valid:', s_step_inputs_valid.shape)\n",
    "print('a_elem_inputs_valid:', a_elem_inputs_valid.shape)\n",
    "print('a_comp_inputs_valid:', a_comp_inputs_valid.shape)\n",
    "print('Q_targets_valid:', Q_targets_valid.shape)\n",
    "print('en_targets_valid:', en_targets_valid.shape)\n",
    "\n",
    "# Initialize constraint network\n",
    "dcn = DCN_pytorch().cuda()\n",
    "\n",
    "# Train DCN on random data\n",
    "dcn, train_loss_list, valid_loss_list = train_DCN(dcn =  dcn, num_epoch = 20, s_material_inputs_train = s_material_inputs_train, s_step_inputs_train = s_step_inputs_train,\n",
    " a_elem_inputs_train = a_elem_inputs_train, a_comp_inputs_train = a_comp_inputs_train, Q_targets_train = Q_targets_train, en_targets_train = en_targets_train,\n",
    " s_material_inputs_valid = s_material_inputs_valid, s_step_inputs_valid = s_step_inputs_valid, a_elem_inputs_valid = a_elem_inputs_valid,\n",
    " a_comp_inputs_valid = a_comp_inputs_valid, Q_targets_valid = Q_targets_valid, en_targets_valid = en_targets_valid)\n",
    "\n",
    "# Plot train and valid losses\n",
    "plt.figure()\n",
    "plt.plot(range(0,len(train_loss_list)), train_loss_list, label = 'Train')\n",
    "plt.plot(range(0,len(valid_loss_list)), valid_loss_list, label = 'Valid')\n",
    "plt.legend()\n",
    "# plt.ylim(130,170)\n",
    "plt.xlabel('Training epoch')\n",
    "plt.ylabel('Loss')\n",
    "# plt.savefig('plots/train_valid_loss.png', dpi = 300)\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 22,
   "metadata": {},
   "outputs": [],
   "source": [
    "torch.save(dcn, './dcn_models/oxides_calcine/dcn_random_roost_constrained')"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Visualization of results"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Final reward of compounds generated from constrained model (DQN + DCN)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 23,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(tensor(0.0038, device='cuda:0', grad_fn=<MeanBackward0>),\n",
       " tensor(-0.0054, device='cuda:0', grad_fn=<MeanBackward0>),\n",
       " tensor(-0.0098, device='cuda:0', grad_fn=<MeanBackward0>),\n",
       " tensor(-0.0311, device='cuda:0', grad_fn=<MeanBackward0>))"
      ]
     },
     "execution_count": 23,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# Earlier models\n",
    "# dqn = torch.load('./dqn_models/dqn_test_500iter_RF_constrained') # Trained in constrained RL\n",
    "# dqn = torch.load('./dqn_models/dqn_test_490iter_RF') # Trained in unconstrained RL\n",
    "# dcn = torch.load('./dcn_models/dcn_test_random_RF_constrained') # Trained on random data\n",
    "\n",
    "# Final constrained models\n",
    "dqn = torch.load('./dqn_models/oxides_calcine/dqn_490iter_roost_constrained') # Trained in constrained RL\n",
    "dcn = torch.load('./dcn_models/oxides_calcine/dcn_random_roost_constrained') # Trained on random data\n",
    "\n",
    "# dqn.fc1.weight.mean(), dqn.fc2.weight.mean(), dqn.fc3.weight.mean(), dqn.fc4.weight.mean(),\n",
    "dcn.fc1.weight.mean(), dcn.fc2.weight.mean(), dcn.fc3.weight.mean(), dcn.fc4.weight.mean(),"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 24,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "Could not satisfy en constraint after 100 actions\n",
      "['Pt7P9Ag2O2', 'O2As8As4O2', 'Te2Fe2N8Ag7O7', 'Pt2P1N7P2O7', 'Sn5S9Cl4Au8O2', 'Li2Ni2C1Se6O2', 'K6Ni2Se1O6', 'P8Ge8S4P8O7', 'Pd2S2Ni2Cs4O2', 'Ag8Se1Ge5O7', 'P8Sn1S7O2', 'N1Te1Ir3O2', 'Se1Pt8Si2P9O8', 'Sn8Ir6P9O8', 'Sn9Na2P6O6', 'As9Ir4O2', 'F7Cl2Cl7O9', 'Hg1Br7Ir4Na8O7', 'Si1Pt9K1O6', 'As7Tl7Ni2P2O8', 'In2Pt2Mn2Ir9O2', 'Bi1Se2As5O2', 'Ir9Si9Pt6O6', 'Sn2Ir7S7Cl7O8', 'Se1Au2Ge7S2O8', 'Bi9Ir8P8Dy2O8', 'As9S4As1Sn6O7', 'F8As6Sb7Hg8O2', 'Ir4Fe2Tl8Hg5O2', 'Pb2Ir2Pt2Ir2O8', 'As3As1Sn5Ir2O8', 'Ag9Ir2Pb3Si7O6', 'Sn3Ir4Tl1Pb1O2', 'Sn4S8O9Pt8O8', 'Co2Nb9Sn4Ir8O2', 'Pd2As1Cl7O7', 'K4S7Ir2Pt9O8', 'Hg1As6Pt2Ir5O8', 'Bi1Ir8Ir4Na8O2', 'Tl1C2Se9O8', 'K4Ir7N7As2O2', 'F9W7O9', 'Ag1Ge8P8Pb4O2', 'As6P2P7O2', 'Pb1Ir9As5N7O2', 'Au8Sn9Ir9Ir8O8', 'Pd1S2Pb7As1O2', 'Pb9Ir9Pt2Sn2O2', 'Sn6Ni2Na2O7', 'Pd9P1Ir1Sn6O6', 'N8Pt9O2', 'In2Se2Si2Sn7O8', 'N1O9S7Tl7O8', 'Na8S2C2O2', 'As6P2Pb6Pb9O2', 'Pt8Se4Pb2O2', 'Rb2Se2Pt4Pb2O2', 'As7Pb2Ir6As6O2', 'Cl8S8Pt2O2', 'Ir1Fe2Cs4Mn7O2', 'Bi5S2As6Cl8O2', 'Te8As9S7O2', 'Bi4Ir2As2O2', 'Bi5Ir2N8O2', 'Sn3Ir8Pb1Pt7O2', 'N4Sb2Br7F9O8', 'Pb6Ir7Pt7O8', 'Pd1Ir6Pb8Ir1O8', 'Mn2Ni2Ir7Sn8O8', 'Sn9P2Tl8Pb7O2', 'B8As9As1Sb8O2', 'Pd1Ir2W2O8', 'Ag8Ir1Ir7O7', 'In2Ir7Mn2Mn9O6', 'Rb2As1N8Pt1O6', 'Rb2Pt2N4F8O2', 'Bi1Ir8Sn1Se7O2', 'Bi3Ir2Pb8Ir9O8', 'P8P2Sn3O2', 'Cl1Ir8Mn8Pt4O7', 'Au8Pb4As1O2', 'Ir6Si9Se7O8', 'Dy2Ge7Pt7Ir8O8', 'Sn3As5Ni2Pt8O2', 'Si8Sn2Pb2O7', 'Bi9Se2Cs9Bi4O8', 'F9As8Cs7Pb7O8', 'Co2Ce1In1Ir8O1', 'As6S8F3Cl1O2', 'Bi9Ir9Pt2Se2O2', 'Cs9Ni2N2O2', 'Sn1Ir8O7', 'Cs7Ir6S2Pb5O2', 'Hg8K5Cs7O7', 'Sn9Pt2Pt2P2O8', 'As9Sn3S6O2', 'N1As6Pb6Cl5O2', 'Pd1As1K5Ir6O7', 'Na8Ir2O6', 'Pd9Pb1Ir1O2', 'Cl9Ni8Cs7O2', 'Dy2N2Pt8O2', 'Tl8Ir5P7O2', 'K7Ir2Tb2O7', 'F9Ge7Te2Rb7O8', 'Hg2Sn7Cl9S2O2', 'Hg8Li2Ir9Na2O2', 'Bi8S2O2', 'K2Ir2As1Ir8O7', 'Cs1Ni2Hg8Ge2O7', 'Li2Pb1As8As7O6', 'Sn7As3As1P2O2', 'Pb2S2Na2Pt9O8', 'Cl8S2Pb9O2', 'Mg2Lu2Pd7O2', 'Sn9Se8Pt6As1O8', 'Tl1Ir5As5Se8O2', 'As3As9Na2Tl7O2', 'As6As6As7Ir8O8', 'Pb4Ir9P7O8', 'Pb3S2Pb6S6O2', 'F6Ir6Ge8O8', 'Sn1As1Ni8Pb9O2', 'Cs1Ni2O7', 'Ir8Ir4Pb2As5O8', 'Sn2Ni2Ir7Si2O8', 'Br8Ir2Pt7O8', 'Sn4S4Se8F7O8', 'Cl8Fe2Na6O2', 'K9Ni8Ir2Ir2O6', 'Ir8N2Pb3Pd2O2', 'Ir8Se4O8', 'Ag2Ir7Si8Hg4O2', 'Si7Pb3O2', 'Pb8Se2Rh2As1O8', 'O8S4O7S9O2', 'Hg2Bi8Ir1Pb1O2', 'Ir8Se8As3Pb5O2', 'Tl8Ir2Se9Tl2O8', 'F6As3K7Ir1O2', 'Bi8Ir8S7Pt9O2', 'Rb8Ir6Se8Pt4O7', 'Sn9Se2Pb1O2', 'In8S7Ir4F1O8', 'Si2Tl7Ir2Pb3O2', 'Cs9Ir9Pb1Pt9O7', 'S2Ir7As1As5O8', 'K4Mn2Ir2As9O7', 'Cs9Ir2F1N2O2', 'Pb5Ir2Pb1O2', 'Ir2Ge8Ir9O8', 'Co2Sn2Sn5Pt2O2', 'Bi3Ir2Ir1Se8O8', 'Li8Ir8Ni2Tl7O2', 'In8Si2P9F8O8', 'O8S2Pb4Pb9O2', 'Cs2Na2Pb9Sn1O2', 'Mn8Ge2P6Sn2O7', 'Sn2P7Cl5O2O2', 'N1As6Sn7S4O8', 'Mg8Ho8As9Ir5O8', 'Tl2Ir2P7F4O2', 'Bi3Ir8Ir4Pb1O8', 'In2Re2N4Sn1O8', 'Tl2Ir5Tl9Pb5O8', 'Cs1Ir6Tl1W8O2', 'F6Tl8Sn2Sn3O8', 'Sn2Mn2Ir2O7', 'P1Au7Pb1Sn7O7', 'O2Sn8Sn7S9O2', 'Pb4S2Tl2As8O2', 'Ag9P1Na9N7O7', 'Cl7Ni9As7O7', 'Sn3Ir2Si7O7', 'Cs7Si2Se9Si8O2', 'Ag1Ge7Tl1Na7O2', 'Dy8Ir5Sn9Tl7O8', 'K8Mn2P7C1O8', 'Sn7Au2Ir6Pb3O2', 'Rb8Se2Tl7S2O2', 'Pb8Ir9Re2N2O8', 'Pb4Ir7Sn1Ir6O2', 'Mn2Ni2Se1Se2O2', 'As4Pt8P7N7O2', 'Te2As3As1O7', 'As5Au7Ir2O2', 'Au2Pb7Tl1Se2O2', 'Bi2Se2Si2Cl7O2', 'Pd1Ir8K7Ge4O8', 'Ir7Ir8Pb5Na2O7', 'Sn9Se8Pt8Si2O8', 'Hg8Ir3Ir2Ir5O2', 'N1S8O7', 'P7As5O7', 'Bi4Ge2Ir9Ir4O8', 'Bi5Ir8Ir8Ir7O2', 'Cl9Sn9O7', 'Se2Pt9Na8O2', 'Na2S2K5O8', 'Cl8As7Hg7O8', 'F2Au8P8As1O7', 'As1S8As4Cl5O2', 'F6P7As4S2O2', 'Hg2Se7Pt2O2', 'Tl8Ir5Mn2Se9O2', 'Dy2Ge7Pb5O2', 'O2Pt7Se7O2', 'Cs4Se2Ir7Pt9O7', 'K9N2O2', 'Er2Ge7Si4P7O7', 'Bi6Ir2S2O8', 'Sn2Ir4S7Hg9O2', 'Tl1Ge2Ir6Ir8O8', 'Se8As4Si2As5O2', 'Se1Au2O8', 'N4As2Sb6O9O7', 'As3Sn5Ir2Na2O2', 'Mn8Ni2Ir5Pt2O7', 'Pd8P1Cs1O7', 'Au2Pb4As7Pb5O2', 'Se2Pb7Ir9Se7O8', 'As5As9Pt7Si7O2', 'Cl9As1Sn5W2O2', 'N1Ni2Li1Pb5O8', 'Sn5Pt9Tl7Ir5O2', 'K8Pb1P8Li8O7', 'P2Sn8S7Pb3O8', 'P1Pb2Tl2Si8O2', 'Hg9Rb8As3Ir1O6', 'Sn7Mn2Se4Pb8O8', 'Au2N8Ir1F4O8', 'P8Pb2S2O8', 'Pb9Ni2Ir7Se9O2', 'Br8P8Pb6S8O2', 'Sn9Sn1As5Ir2O8', 'Hg2S7O7O2', 'Rb2Mn2Cs6Na8O2', 'Ni2Si7Ni8O9', 'K6Ni2Pt7O7', 'Er2Ni2Ge2Se2O2', 'Hg2Ir2Na2Tl1O7', 'K2Ni8P1Ir9O2', 'F8Pb6As1Cs6O8', 'Bi2Ir9Tl1Se2O8', 'Pd1Ir5C8Sn1O2', 'Ir8P4O9', 'Bi8F3O2', 'Tl1Ir1S8Sn6O8', 'Pb1Ir2Mn2Pb1O2', 'Er2Ge8Si2O6', 'As6As6S2As2O2', 'Tl9Ir5Pb5Pt8O8', 'Ir4Tl2Re2Pt2O2', 'Dy2O7', 'Ir9N4As6O2', 'Ni8Pt4Pb7Pt7O8', 'Sn4Ir7Ir8N2O2', 'Sn2S9S9O2', 'Rb8Ir7W2P4O2', 'Ag9P1Pt7Pt6O7', 'Ag2Ir5Pb7O8', 'Ir2Ir2Si7Cs1O6', 'As9Na2As9Pt8O7', 'Au2Pb8N2Ir6O8', 'Li2N2Ir1Fe2O5', 'Tl8Ir1As1O8', 'Bi4Ir7Pb2Si1O2', 'Cl1O7', 'B2Sn9Pb8Cl8O8', 'Bi7Ni2Pb2Sn4O2', 'As9Ir2P2O8', 'F9Sn3Hg6Bi9O2', 'Mn8Se1O6', 'As7Pb2Pt2Tl2O8', 'Si8As1Pt8O7', 'Pb9Br8C2O2', 'Bi4Ni2Pb2Se2O2', 'B8Ge7Re2Cs7O8', 'Mg8K8Pt8Hg8O7', 'Dy2Ir2N7Te9O8', 'Bi3Se2Ir3S2O8', 'As5Ir8Ir8N2O2', 'Pd1S7Sn9Sn2O7', 'Hg8Te6Li8As1O7', 'K7Mn2Ir6Ir5O6', 'P2As7Se2As7O2', 'Cl9Pt2Hg9As5O2', 'Rb2Ir2O7', 'Rb8Ir7Ge8Ir7O7', 'Cl9As2Mn2P9O7', 'Bi1Ir9Ge8Pt2O2', 'Ag1Ir2As2Na7O7', 'Ag9Ge8Pt1N4O7', 'Cs2Ge2Ir2N9O8', 'Sn8P2Pb5O2', 'Rb8Mn2Tl8Ni2O7', 'Ir6Si9O7', 'Pb7Ir7Pt7Pb1O2', 'Au8Se7Ir7As5O2', 'Sn7Pt6Ir8Ir6O8', 'Mn8Se1O6', 'K2Ir2Si2O7', 'Br8Pt8Ir5Ir9O8', 'Bi9S2Tl2W2O8', 'Na2Ir3S9As8O8', 'K6O2', 'Bi1Ir2Se7O2', 'F5Tl7As2O2', 'Ag9Ir7Pt8Sn2O7', 'Tl9Ir2N4O2', 'Bi1Ni2Sn7O8', 'Rb2Na2Cl3O2', 'Na2Fe2Nd1Te4O8', 'Er2Ge7As5Ir8O2', 'Hg2P7Sn7O2', 'Ir1Si9Cl8P1O2', 'K7Ir9Pb7O2', 'P8Na2Sn2Si2O6', 'Ir7As3N7O6', 'N2Ru8F7Sb2O2', 'Sn3Re2Si9O7', 'Na8Au1Pb8O6', 'Tl1Ir2Si1Ir6O8', 'Ir4Na2Si2O2', 'Ag8Ir2Pb6As4O2', 'Cl9As6Cl7S6O2', 'Sn7Se8S2Sn6O7', 'As9Tl7P2As2O2', 'As4Sn3S4F4O8', 'Tl8K5Ge7Na7O2', 'Si7Ge9Ni8O7', 'Cl2As9Pt7O2', 'K9Ir2Sn3O7', 'Sn6S2O7', 'Cs2Si2Sn9Pt7O2', 'Sn4Ir7Ir8O7', 'Se2Pb8Cs8O6', 'Br8P7As3P2O2', 'F6O7', 'Ni2Pb8N4Se9O8', 'Tl1Ir2Si5Ir7O8', 'F6Ni2F7Sn4O2', 'Hg2Cs1K8O6', 'N8Ge7P2Bi2O2', 'K8Ir2S8Na9O7', 'Au2As4Na7Tl4O6', 'Pb8Ir2Ir4O2', 'N4As2Cl2Sn1O7', 'Ni2As3Pt5Si9O7', 'Rb8Pt2Tl8Pb5O7', 'Hg8Rb2F5Ge7O8', 'As5Pt7Tl2Sn2O2', 'In2Si2S2Sb4O8', 'Pb2Ir7As8Pb3O2', 'Sn5Ni8N2O2', 'As2S7O7Sn2O7', 'Au8Na8O7', 'K9Ir7S2Pb4O8', 'Pb7Ir7S7O8', 'Dy8Ge7As5O2', 'F4Tl7O2', 'Hg7Au2Ir4Pt8O2', 'Ir9N9O6', 'F2As4Bi1O2', 'In2S2As3Te2O8', 'Mg2Er1Ag2Tl8O2', 'Pb9Ir9Ir6Ir4O8', 'As4Ir7O8', 'Se1As3Ir8C8O2', 'Sn2Se8Pb9Pb7O2', 'Ag1Ge2Se7Na2O7', 'Sn4Ir9Ir9O2', 'B8Sn9Au8As5O8', 'Hg2Pb4Ir1Ir2O8', 'F2As4As4Ir6O2', 'Pd8Ir1Sn1Te7O2', 'Bi7Se2Cl7Sn1O8', 'Re2Si7P2O2', 'S2Ge4O6', 'As8S7Pb8O2', 'Co2P5Ir5Pb8O8', 'Hg8O2Cs7O6', 'Se2Ir7O9', 'P8Ir2S9S6O7', 'Pb6Ni2Pb3Pt4O8', 'Pt2Se9As1Na7O8', 'Rb2Cs6Pb4O2', 'Ir2Pb7Pt2Sn3O2', 'As7As5S7O7', 'Bi7Ir7Pb7O2', 'Mn2Ir9Ni8Sn1O2', 'Mn2N7Na8Si1O9', 'As2Pb3Pt9O7', 'Pd8Pt2Ir5Si4O2', 'Pb5Ir3N2Bi4O2', 'P2Pb6O2', 'Pb1Ir8As7Ir6O2', 'Se2Fe2S5Si2O8', 'Sn1Au2N9Ge7O7', 'Ir7Ge2N7S9O8', 'Bi4Ir2Pb8O8', 'Rb2Ir2S9O7', 'Ir7S5Si2O2', 'Sn6P2Ir2As6O2', 'Cs9Pt7As5Pb6O8', 'Cs4Ir4Ge8Sn9O2', 'Hg1Cs7Ir9Si2O7', 'As8S8Cl8Cl4O7', 'Sn8Ni8Se9O8', 'Ag8Ge2Ni2Na2O1', 'Li8Mn2Mn2Ir1O1', 'Hg9Sb2Si9O7', 'Pb6Ir3Tl5P2O2', 'N9Te9Sn7Cs2O2', 'As2Pb8Pb7O8', 'F8Mn8Se9O8', 'F2Au9Mn2Pt9O2', 'Ir8As1C8O2', 'Sn2Pb2Ni2Pb5O2', 'Dy2Ge2Ir2Ir6O8', 'Tl8Ge2Pd8Si8O7', 'Hg8F2Ni2As2O8', 'Co2P9Cs2P2O7', 'F9Sn9As7O6', 'Pb2Ir7As5O8', 'N2Pb8O8', 'Ag1Ge2S8Ir5O8', 'Pb7Ir3N7Ir1O8', 'Pb6Ge2Pb2O8', 'P2Sn5P8Tl7O8', 'Sn7Na2N8O7', 'As7P7S7Sb2O2', 'N4Bi1Ir6Cl8O2', 'Cs1P7As1Pb1O2', 'Bi3Ni2Pb7Si9O8', 'Pt8Pb7Sn4O8', 'Ir7S8N2Li9O2', 'Pb3Ir9Sn5Ir5O8', 'Pd1P1Ir2N1O8', 'Pd2S2Ge2Pb1O6', 'Mg2Pd9Se9O7', 'N2Pt2Sb2O6', 'Mg8Mo2Ni2O1', 'Bi3Se2Ir2Pb1O8', 'F6Sn6Na8Ir2O2', 'Bi4Ir7As9O8', 'Sn7As5Se2Ir8O2', 'Pb2Ni2Sn1Ir1O2', 'Cl8Ru8Pb3O6', 'Rb8S7Pt7O7', 'Hg8Bi2Cl2Ir9O2', 'Bi6O6O2', 'Sn9Ir7Se4Pb8O8', 'Re2Co2K1Pb1O2', 'Cs4P7Se7Pb6O2', 'P8Ge8Pt4Ir2O2', 'Pd2Ir9Cs4O6', 'Li2As1Si2Ir5O8', 'As4Ir8P2P2O8', 'Rb2Ir6Tl4P9O8', 'Hg9Cl4Se2Re2O9', 'Tl1Hg2Pt4O2', 'Dy2Ir2Pt2Ir2O2', 'Pb6Ir9Re2O8', 'Mn8P1Tl5O7', 'Cs1As1Ir5O6', 'As9Pb4Ir6Sn8O2', 'Mg2As5Ge7O9', 'Ag2P1Na8Ir7O6', 'Pd9Ge2Ir1Pt9O2', 'Pb7Ir7Ir2O8', 'Cs7Mn2Re2Sn1O2', 'Rb8Ir2As5Tl1O7', 'Li2Sn1Pb5Tl7O8', 'In8Re2Mn2N7O2', 'Mg2Ni9N8Sb2O2', 'Re2Sn5As9As6O2', 'F6W2Eu8F1O2', 'Ag1N2Pb2Pb3O2', 'Bi6Se2Br2Ir2O8', 'B8Tl2O7', 'Pb9Ir2S7Se1O8', 'K4Au1Re7Pb2O2', 'Ni8Pb1Na7Ir6O8', 'Bi7Ir7As7Pb2O8', 'Se7Ir7O8', 'Pb6S2Pb2S6O8', 'K2Ir2Pt9O7', 'Co2Co1Dy2Os2O1', 'O2Sn9Ir2O2', 'As9Ir6Si7Pb9O8', 'Se1Ir2N9Sn5O8', 'Hg7Hg6Cl1Br2O6', 'Cl2Au9O7', 'Pd8Ir5K7Ir4O6', 'Se2Ir7Se2Ir2O2', 'Ag1Ge8Ir8Pt4O2', 'Pd1Se3Ni2O2', 'Pd1Ge2Ir1Ir5O8', 'Bi8Ir8Pb1P8O2', 'Pb4Ir9Tl2Ir2O8', 'Hg2Li9Ir4Sn5O7', 'Sn1P7Pt7Se4O7', 'Pb1O2', 'Hg2As2Ir4O2', 'Er2Ge7Pt2O7', 'Pb4Ni2As3O8', 'Co2Mn1Ge2Se1O1', 'Rb8S7K8Tl4O7', 'Tl1Ir5O2', 'Sn2Au2Ge2Ir5O8', 'Bi8S2Si2Ir2O2', 'Ir9Tl2As2Ir2O2', 'Pd1Ir5Si8Se8O8', 'Ir8As7Pt2O2', 'Bi9Ge2Ni2Ir8O2', 'Cs7Re2Te7Pb3O2', 'Pb4S2Pt8As1O2', 'Bi6Ni2Pb7O8', 'Sn2Ir9N7Si9O2', 'Hg9Bi4K1Ir7O2', 'Pd9S7Cs8Hg6O7', 'Cl2Pb9As9Tl8O2', 'Ir2P9Tl7O8', 'Hg1Cs8S1N1O7', 'K2Au1Sn9O6', 'Se8Pt4Pt8O2', 'Cs4Mn2Pb1Na8O7', 'Ag9Ge7Se2Ir5O2', 'Pb4Ge2Ni2Ge2O2', 'Au2Si2O7', 'In2Ir7P7Sn7O8', 'Pt8Sn8S7O7', 'Pd2S2F8O2', 'Cl1Ir7Mn8S2O7', 'Ni8Au2Si7O7', 'As8As5Pb3Si2O2', 'Sn6Pt2O7', 'As9Au8As1P1O7', 'Ni8P7Pb5F8O2', 'Hg1Hg6Ir1O6', 'Au2Tl8Si2Sn5O8', 'As4Ir7N2O8', 'Bi7Ir8Pb2As7O2', 'N1P8S2O7', 'Te8Ge5Ni8O7', 'Te8Ni8Na8Se1O7', 'As3Ni2N7Hg5O2', 'Cl2Ca7Ge7O6', 'Ag2Ge7Mn2N8O7', 'Cl9O7O4Ir7O8', 'F6Sb2Bi2O8', 'As8As1Ir9O6', 'N8O7Pd7Ir6O2', 'Dy2Se2N4Pd8O2', 'Sn1S2Sn6S9O7', 'Na2Ir9As5O2', 'F1As4S2O8', 'K4N2Si7O8', 'K8Ni2S6O8', 'Bi2Ir9Pb3Ir4O8', 'Pb2Ir2Pt9O2', 'Te8Pb8S2Pb6O8', 'K8Au1Na8O6', 'F4Cl7Ir8Rb1O8', 'S2Tl2Sn4O2', 'Co2Tl9Sn7O2', 'Rb2Na2Ir4Si7O6', 'Ir8Se7Si8As2O2', 'Si8Sn7Sb6O7', 'Hg7Sn1Sn8F6O2', 'P7Ni2O7', 'Hg2Ir7Re2Pb2O2', 'Pb1Ge2P2F8O2', 'Ir2Pb3N8Mn2O8', 'Re2Pb7Se9As2O2', 'Pb2Ir7Si5Pt8O2', 'As9Au8Se2Ge2O8', 'Cs1Ni2K4Ge7O2', 'Pd1Ir1Sn5O8', 'Ag8N7Se9Bi1O2', 'Tl2Ir1Pb1O2', 'Pb5Ir7Tl1O2', 'Pt8As8S7O7', 'Cl2Ge8Se8Ir9O2', 'Mg8Lu8O7', 'Sn4S9K5Tl9O2', 'Sn6Ir9O8', 'Cl9Ir2Se2O2', 'Au2Pb9N8Bi5O8', 'Ag8Ge2Ir4Si8O7', 'Cs8Ir6Ir7O6', 'Bi6O8', 'Sn3Pb2Cs4Si9O7', 'Pb4Ir8As2Ir1O2', 'Sn3Pt2Ir8P9O8', 'Ag1P1Pb9Pb5O8', 'In2Ni7Pb7O2', 'Te8Sn3Re2As5O2', 'N2Sn7K1O2', 'Pd2Ir1Pb2O2O2', 'Dy2Ge2S4Cs8O7', 'As2As8Au8Ge2O8', 'Se2Fe2Si9Tl2O8', 'Cs4Ge2Ir9Os2O7', 'Sn8Sn1S7Sn6O2', 'Ag8Ni2P1Ir2O7', 'Cs7P7Li4N7O2', 'Ir8N9Se9O8', 'F7Sn1Cs6As9O8', 'Si8Pt4Rb2Si7O6', 'Pb1Ni2Ir4O8', 'Na2Ni2Pd2As1O2', 'P1As8S6O4O2', 'Pb9S2Au8Si2O2', 'Pd9Se3Bi4Tl8O2', 'Cs9Re2P7Pb3O2', 'Hg7As7S2Sn1O2', 'Pd9Ir1Br8O7', 'Ni8Ir1Mn8Ir8O8', 'Se1As8Ir2As6O8', 'As5Pb9Sn2O8', 'Sn8Ir8Si7O7', 'Hg2Li2Ir7Te2O2', 'B8Ir5Pb2Cs5O2', 'As7Pb6Ir2Ir5O2', 'Si7Na2O8', 'Hg7Pt2O2', 'Si2As8Pt2As8O7', 'Na8Ir1Pb2Ir4O6', 'Au2S9Sn2Pb5O8', 'Ir4P8Sn3O8', 'Si2As6Tl2As4O2', 'As9Pb6Ni2Pt4O2', 'Ir1Na2Pb8As9O8', 'Bi2Ir9S8O2', 'Re2Ir8Tl8S2O2', 'P1Sn2Pb8O2', 'Pb6Cl4O2', 'Hg1Te7Pt2N9O2', 'K8Ni2Ir2O7', 'N8O5As5O7', 'Bi9Ni2As2P1O8', 'F8O7As2Sb4O2', 'Ag9P1Ir8O7', 'Pb7Ir3N7Pd2O8', 'Bi1Se2Re2Pt7O8', 'K7Ir8Mn2O7', 'Tl9Ir1S7O8', 'Sn3S7Ir6Cs1O8', 'K8S2C8Si1O8', 'Se2Pb8Hg6As4O8', 'Ag8Ni2Si1Si7O1', 'Pd2As1Cs6Se6O7', 'Cl1As1Ni2O7', 'Si7Pb4Ir6O2', 'Rb8Se2Se8As1O2', 'Na2Ir4O7', 'Te8As8P8K5O7', 'Au8S8S4Ir6O8', 'Tl8Ir5Re2Si8O2', 'P8Ir7Pd2Ir7O8', 'Pb9Ge2O8', 'Na8P1Se2N5O2', 'F1As1O7', 'As7Pb8Bi1O8', 'Ir2P5F2P4O8', 'Pb3Ir7S2O8', 'Ag9P1P1Pt9O7', 'Au2S1Pb9As8O8', 'Bi1Ir7N9Si8O2', 'P2Sn3Pb6Pb4O2', 'F2Pb6O2', 'O8As9Ni2Ir7O8', 'P8Tl2O7Ir8O8', 'Mg8K7O1', 'Ni8Mn8Ir1O1', 'K7Au1Pb5Sn5O7', 'F8W2O9', 'As8Au2P2O7', 'As8As9As1S8O2', 'Pd9S2Ir1Cs7O6', 'P2Pt7As1Pt2O7', 'Pb5S2As8Pu2O2', 'Se8As8Ir7W2O8', 'Pb1Sb2O2', 'Ag9N2N2As5O6', 'Cs7Mn2Ni2Pt2O6', 'As3Sn5Mn2Ir4O8', 'Hg9Sn3K8O7', 'Cl9S2O8', 'Pb3Se2C2Pb5O8', 'K4Ir7S9N4O2', 'K8Ir2Tb2O6', 'As6Re2N8Si4O7', 'Sn6Ir6Ir5N7O2', 'F6Ir6Te7O8', 'N4Re8Rb8Tl8O2', 'Pd2Se3Sb9As7O2', 'B8Tl8O7Hg4O8', 'F9As9As5Sn8O8', 'In8Ni2N2O2', 'Sn7S7Ni8Na8O2', 'Pb3S2Ir7Au9O2', 'Co2V1B1Se8O8', 'Sn2Re2As3Se7O2', 'Ir2Pb2S7F6O8', 'F9Ge8As6Pb4O8', 'In2C2K9O2', 'Sn2Ir7S7Cl7O8', 'Ir2Ir7Pb2N7O8', 'As9Ge2S8Ir2O2', 'Ir7Se9Sn1P2O2', 'Br8Ge9Si2O8', 'Bi1Ni2Sn3Tl4O2', 'Rb8Bi1Pb8O6', 'Dy2Ge2As7O6', 'As5S6O7O2', 'Pb5S2Pb8Sn1O8', 'Pb8Ir2As2Si7O2', 'Pd9Ir4Cs6As8O6', 'Bi6Ge2K3N4O2', 'Cs7Mn2N4C2O7', 'As3Sn5Pt7Pt9O7', 'In8Mn7P7P9O9', 'Pd1P2S6Cl2O7', 'Li2N1O2', 'Hg7Sb2F1Si2O8', 'Se7Pt9Pt2O8', 'Cs1Re2Ir7Na7O6', 'Ir7S7W2Cl7O8', 'Ag2N2S9S2O7', 'Cl9As5F2Sn6O2', 'Sn5Sn1Au2N2O2', 'As1Tl7K7Hg7O7', 'Ir4Ni2N7O2', 'Rb2Ir7As4Tl2O2', 'Ir7As5P7Hg5O2', 'K8Ir1Na7Pt9O7', 'Pb9Ir2Se8Ge8O2', 'Ag1N7Pd6Ge2O6', 'Cl2S2As2O7', 'Sn6Ir7Mn2Ir1O8', 'Ag8Ge2Na2Sn7O7', 'Bi8Ir9N2Pb3O8', 'Bi6Ir9Pb7Ir5O2', 'Pb1Ir3Pb1As2O2', 'Pd8Ir2Bi4Na7O2', 'In2S2Pt2Cl2O8', 'P2As5O2', 'Se1Pt2Ge8O2', 'In8Mn2C2O8', 'As9Pb8Se2Ni2O2', 'Bi1Se2Pb9Pb8O2', 'Sn6Pt2N9O7', 'Ir1Pb2Na2P8O8', 'Pb6Ir2S2O2', 'Pb8Ir9Pb9Se9O2', 'Ir4Na8As4Sn8O1', 'Sn5Pt4N9Hg8O8', 'Na8S2Se1O7', 'Sn4Ni2Tl5P2O2', 'Tl8Ir5As5Pb1O8', 'Tl8Ir2N8K6O8', 'Mg8As6O6', 'Sn1Pt2P4O7', 'Te2Pt7Si2Pb3O8', 'Pd8Pt2Ir7K2O6', 'Ag9P1Pt7Sn1O6', 'Pt7As9Pt2O7', 'F1O7', 'F8Sn8As9O8', 'As8Ir8Ir5O8', 'F6Au2F8As6O2', 'Ag2N2N7O7', 'Br8Ge7O2', 'Te8Fe8S4S2O8', 'Ag1Ir5Si4Hg5O8', 'Pt8N8Ge8O6', 'Tl7C2Ir8O8', 'Sn4Pt7As2Pb5O2', 'Re2As3As8Na2O8', 'Tl1Ir1As2Sn9O8', 'Te2Pb2Ir8S2O2', 'Au2Pt8Pt2Pt9O7', 'As2Au7Ir2As4O8', 'Cs1P2Ir1Ge7O7', 'Au2Ir9Sn7Na7O2', 'Sn8Ni8Se6O7', 'Mn2Ir1S8Pd7O8', 'Bi5Ir8Ir9As5O8', 'B8Ir2Ir5O6', 'Ag9Ge2Ir9Ge2O7', 'Li8Si8Ir2Tl7O2', 'Se8Pb8Ir8O8', 'Bi2Ir9Ir1Se7O8', 'Pb6Ir8N1Ir1O2', 'Ag8Ge8Ir7Pt2O7', 'Mn8N1N2O7', 'Pb5Ge2F8Pb1O2', 'Se8Pb8Ni2Pb8O2', 'Bi4Ir8S8As9O8', 'Co2Tb2Pb1Se2O8', 'Pb6F7O8', 'Na8Ni2Si2Au1O1', 'Br8As5Pb7O8', 'As4P7Au2O2', 'K7Ir2Sn2O7', 'Ag9Ge8N8Se4O2', 'Pd8Na2N2Cs2O8', 'F9Pb2Cl9Tl8O8', 'Ir7N2As9O8', 'P2Pb7S4O2', 'In2Pt2N9Si4O7', 'S2Sn7Ir7O8', 'Ag2Ir1Tl2Sn1O2', 'Sn6Pb2K2P4O2', 'As4Pt7P4Na7O7', 'Na8Ir2F1N2O7', 'Mg2Ho8Pu8N1O7', 'Sn9S8Se2Cl3O2', 'Se1Ge9Ir8Pt8O8', 'Cs8Ir4Sn3Ir9O7', 'Pd1Ge2Pt1Ir2O8', 'Sn8S4W2O2', 'F6S9Mn2Pb9O2', 'Mn2Ir8S8Hg2O8', 'Se2Ir7Se2O2', 'Tl9Ir5Si7As6O8', 'N9Te2F4Cs8O8', 'B8Ir8Sn5Pt2O7', 'Sn3Ir4P2Sn7O8', 'Bi9Ir2Ir4N9O8', 'Hg8As8K3Ir1O6', 'Au8Pb7Pb5O8', 'Pb4S2Pu2Ir2O2', 'F9Mn8Tl1As2O2', 'Bi8Ir7As4Pb9O2', 'Cl9As5Tl8O8', 'As7As7Ni8O6', 'Mn2Ir1Sn1Pb7O2', 'Tl9Ir2Na2Se8O8', 'Ag9Ge7N2Ir7O2', 'Pb6S2Se2P2O2', 'Pb9S2Au8Pb3O2', 'K4Ir2Ge6O7', 'Pd2Na2Li6P1O1', 'Ir9Ir1Ir6Pt4O2', 'As9Pb2Pt2N8O2', 'Hg8Bi8Sn1O2', 'Sn7Ir4Ir9Se2O2', 'Ir7Pb9P2O8', 'Pd2Si1Ir7Se1O2', 'Ni2As6As8Ni8O6', 'Sn5As1Au2Ir5O2', 'Se1Ir7Ir2O8', 'Se8Ge5O2', 'Sn8Ir7Si4Ir6O8', 'As1As6S8W2O2', 'Bi1Ge2Ir9Se9O2', 'Cs2O2', 'Pb6Ir7As1Ir7O2', 'Br8Pt7P9Te2O2', 'As9Au8Ge2Mn2O2', 'Se2Pb8P7O2', 'Bi1Se2Ir6N1O8', 'N8Au8Ir9O2', 'As4Sn1Ir8Ir1O8', 'As9S9Pt6K2O7', 'Bi2K5Ir5O7', 'O2As7Sn1P8O7', 'Sn8P2Pb6O8', 'Br2Ir7Ir9O8', 'Se2Pb8K8Si8O7', 'F8As5Br7O7', 'Cl8Pt7Cs9As8O8', 'Sn5Pt6Ir5Se8O2', 'Se1As4S2Cl5O8', 'Bi3Ge2Ir9O8', 'Ir2Se2Ge8As2O8', 'Pb5Ge2As8O2', 'F2Ir8Ta8O9', 'Ir1Si8Pb4P8O2', 'As3Pb6Si2Ni8O2', 'As9Se2P7Sn1O8', 'Pb2Ir3Re2Sn2O2', 'Sn3Se8Sn1Se2O7', 'F6P8Pt7Hg1O8', 'Ag9Ir1N2O7', 'Pd9Pt2Ir5Pb3O2', 'Se1Pb2As6O2', 'Ir4N7N4Pb1O2', 'Se1Pb8Sb2O2', 'P8Pb2O2O2', 'As1As5P1O7', 'As8Re2Sn3Ir8O2', 'P8Tl8Sb9O2', 'Pb5Se2Cl9O8', 'P2Sn3As2S2O2', 'Tl2As4O8', 'O2Sn1Pu2O2', 'Ni2P8Bi1S9O2', 'Cs7Ni2As3O8', 'Dy8S2Ir7As6O8', 'As3Ni2Si2Bi5O8', 'K7S7F4Bi7O2', 'Co2N8Li9Pt9O8', 'Sn1Pt7As9P7O2', 'Cs7Re2Na8Si7O1', 'K7Ni2Pt7Pb1O7', 'Ir4Se4S8O2', 'S2As7Sn7P8O2', 'Ag8Ni2Se7N2O7', 'Er2N2Sn6F3O8', 'Bi2Ir8As7Ir4O2', 'Pd2As1K4P5O7', 'Li2Pb1Ir7O8', 'Bi6Ir9Se2Au2O8', 'N1Re8O9', 'Sn7Se2S6Pt2O2', 'Cs7Se2S7Pb3O6', 'Pb6Ir2N2Pb9O8', 'P8As3K1Ir5O8', 'F2As5O4O2', 'Pb4Ir9Pb1N8O2', 'Tl7C2N8O8', 'Bi6Ir9As1O8', 'K2Ir7Pb4Pb8O8', 'Sn4Pt6As8Se2O8', 'Ir2Pb3As2O2', 'Cs2Ni2Pt6As1O7', 'Sn8P7Cl1Cl2O2', 'Ir2Pb7N8Tl8O8', 'F8Tl2Pb7Pb5O2', 'Pb9Ni2Ge8S2O2', 'Hg4S4O2', 'Dy2Ir5P7F8O8', 'Ni2P8Cl9Se7O2', 'Er2Ir7Si2Mn2O9', 'Pb3Ir2Se2P2O8', 'F2Au2Na2Ir2O2', 'Pd9Ir7P7O8', 'Sn5Mn2P8O7', 'Sn8Pt7Se2O8', 'Sn1As5As5S7O7', 'Ni8Tl8Si9Bi9O2', 'Bi2Ir2Au2Ir1O2', 'F8Pb9Ge2P9O2', 'F9Ni9Tl4P9O8', 'Mg2Cs5Ir9O7', 'Co2As5Sn8As3O2', 'Cl8Sn6K2O6', 'Ir7Ir1Se7O2', 'P1Ir2P9Pb6O2', 'Si8P9As7Bi4O8', 'O2Pt6Pt7Pb7O2', 'Sn5Ir4S7Ge7O8', 'P1Sn7O6', 'F4S7Sn8Tl8O2', 'Cl2Te5O7Cl9O8', 'Bi4Ge2Se2P2O8', 'Cl2Ca8Mn2Na2O2', 'Ir1As5O2', 'Cs7Ir7N9Si8O8', 'Pb5Se2Br2S1O2', 'As3Ir7Ni9O2', 'Si8Ir8Cl6Mn2O7', 'Ag9Ir1P2Ir8O6', 'Sn7P7Sn1P2O2', 'Dy8Ge7As7Ir6O8', 'Cs7Bi4O6', 'Na2Pb1Tl7O2', 'Cs9P7As5Ir5O8', 'Pd2Ir4As3Te7O2', 'Ag9Ir7Na7P9O6', 'Ag9Ge2Ir4As4O2', 'Cs7Mn2Pb1Pb5O7', 'Pd9P3Ir1Li8O7', 'As1Tl2Cs6Si7O7', 'Ir2S8K5S7O7', 'Ir9Si4O8', 'K9As1As6Na2O6', 'N4O6Ni2Pt9O2', 'Re2As8N8Bi7O2', 'Tl2Ir5Pb9As5O2', 'Ni8Pt8P2P2O6', 'As4Pb6Tl2Cs1O8', 'Br2O7', 'Si2Ir7Pt8O7', 'Sn4S4Sb4Ir6O8', 'Cs7Pt7As1Ir7O6', 'As9S4Ir6Bi5O2', 'Bi1Ge2Ir8Se2O8', 'F4Tl8As9Sn3O2', 'Hg1Tl7Se2O2', 'F4Au7Na8Si6O7', 'F9Pt2Hg8Te2O2', 'Se1Au2Ir7C8O2', 'F4As2Bi1O2', 'Ni2As8Na7Sn2O7']\n"
     ]
    }
   ],
   "source": [
    "# 1. Get final rewards from random policy\n",
    "with open('./data/oxides_calcine/Q_data_random.pkl', 'rb') as f:\n",
    "    Q_data_random = pickle.load(f)\n",
    "s_material_inputs, s_step_inputs, a_elem_inputs, a_comp_inputs, Q_targets, en_targets = generate_dqn_dcn_inputs_and_targets_from_Q_c_data(Q_c_data = Q_data_random)\n",
    "final_rewards_random = Q_targets[0::5].reshape(-1) # find final rewards\n",
    "en_random = en_targets[0::5].reshape(-1) # find final en\n",
    "\n",
    "# 2. Get final rewards from trained policy\n",
    "Q_c_data_trained, en_targets_trained, final_rewards_trained, final_compounds = \\\n",
    "generate_Q_c_datapoints_and_evaluate_policy(dqn = dqn, dcn = dcn, num_eps = 1000, epsilon = 0, \n",
    "                                            en_threshold = 0.5, stochastic_top_frac = 0.20, oxide = True)\n",
    "\n",
    "# Save compounds generated                                            \n",
    "with open('./training_data/oxides_calcine/final_compounds_roost_constrained.pkl', 'wb') as f:\n",
    "        pickle.dump(final_compounds, f, pickle.HIGHEST_PROTOCOL)\n",
    "\n",
    "print(final_compounds)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "with open('./training_data/oxides_calcine/final_compounds_roost_constrained.pkl', 'rb') as handle:\n",
    "    final_compounds = pickle.load(handle)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 35,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "0.819"
      ]
     },
     "execution_count": 35,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# check charge neutrality\n",
    "len(check_neutrality_multiple(final_compounds)[0])/len(final_compounds)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 36,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "/home/synthesisproject/anaconda3/envs/dqn_roost_EMD/lib/python3.7/site-packages/seaborn/distributions.py:2619: FutureWarning: `distplot` is a deprecated function and will be removed in a future version. Please adapt your code to use either `displot` (a figure-level function with similar flexibility) or `histplot` (an axes-level function for histograms).\n",
      "  warnings.warn(msg, FutureWarning)\n",
      "/home/synthesisproject/anaconda3/envs/dqn_roost_EMD/lib/python3.7/site-packages/seaborn/distributions.py:2619: FutureWarning: `distplot` is a deprecated function and will be removed in a future version. Please adapt your code to use either `displot` (a figure-level function with similar flexibility) or `histplot` (an axes-level function for histograms).\n",
      "  warnings.warn(msg, FutureWarning)\n"
     ]
    },
    {
     "data": {
      "image/png": "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",
      "text/plain": [
       "<Figure size 720x504 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "# Plot distribution of final rewards by random vs trained policy\n",
    "plt.figure(figsize=(10,7))\n",
    "sns.distplot(final_rewards_random, hist = True, norm_hist = True, label = 'Random', color = 'grey')\n",
    "sns.distplot(final_rewards_trained, hist = True, norm_hist = True, label = 'Trained', color = 'red')\n",
    "plt.ylabel('Probability density')\n",
    "plt.xlabel('Reward')\n",
    "# plt.ylim(0,0.0025)\n",
    "# plt.xlim(-2000,0)   \n",
    "plt.legend()\n",
    "plt.savefig(fname = 'plots/oxides_calcine/reward_dist_500iter_roost_constrained.png', dpi = 300)\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Validity of compounds generated from constrained model (DQN + DCN)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 27,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Random policy: Fraction of en OK compounds =  0.402\n",
      "Trained policy: Fraction of en OK compounds =  0.694\n"
     ]
    }
   ],
   "source": [
    "# Validity of random\n",
    "count_random = 0\n",
    "for target in en_random:\n",
    "    if target.item() == 1.:\n",
    "        count_random += 1\n",
    "print('Random policy: Fraction of en OK compounds = ' , count_random/len(en_random))\n",
    "# Validity of trained\n",
    "count_trained = 0\n",
    "for target in en_targets_trained:\n",
    "    if target.item() == 1.:\n",
    "        count_trained += 1\n",
    "print('Trained policy: Fraction of en OK compounds = ' , count_trained/len(en_targets_trained))\n",
    "validity = {'random': count_random/len(en_random),\n",
    "            'trained': count_trained/len(en_targets_trained)\n",
    "            }\n",
    "with open('training_data/oxides_calcine/validity.json', 'w') as f:\n",
    "    json.dump(validity, f)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Experiment: Validity vs en_threshold"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# en_threshold_list  = [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9]\n",
    "# en_good_ratio_vs_en_threshold = []\n",
    "\n",
    "# def evaluate_en_threshold(en_threshold):\n",
    "#     # 2. Get final rewards from trained policy\n",
    "#     Q_c_data_trained, en_targets_trained, final_rewards_trained, final_compounds = \\\n",
    "#     generate_Q_c_datapoints_and_evaluate_policy(dqn = dqn, dcn = dcn, num_eps = 100, epsilon = 0, \n",
    "#                                             en_threshold = en_threshold, stochastic_top_frac = 0.20, oxide = True)\n",
    "#     # Validity of trained\n",
    "#     count = 0\n",
    "#     for target in en_targets_trained:\n",
    "#         if target.item() == 1.:\n",
    "#             count += 1\n",
    "#     ratio = count/len(en_targets_trained)\n",
    "#     return ratio\n",
    "\n",
    "# for en_threshold in en_threshold_list:\n",
    "#     en_good_ratio = evaluate_en_threshold(en_threshold = en_threshold)\n",
    "#     en_good_ratio_vs_en_threshold.append(en_good_ratio)\n",
    "\n",
    "# with open('./training_data/oxides_calcine/en_threshold_list_RF_constrained.pkl', 'wb') as f:\n",
    "#         pickle.dump(en_threshold_list, f, pickle.HIGHEST_PROTOCOL)\n",
    "\n",
    "# with open('./training_data/oxides_calcine/en_good_ratio_vs_en_threshold_RF_constrained.pkl', 'wb') as f:\n",
    "#         pickle.dump(en_good_ratio_vs_en_threshold, f, pickle.HIGHEST_PROTOCOL)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# with open('./training_data/oxides_calcine/en_threshold_list_RF_constrained.pkl', 'rb') as f:\n",
    "#     en_threshold_list = pickle.load(f)\n",
    "\n",
    "# with open('./training_data/oxides_calcine/en_good_ratio_vs_en_threshold_RF_constrained.pkl', 'rb') as f:\n",
    "#     en_good_ratio_vs_en_threshold = pickle.load(f)\n",
    "\n",
    "# plt.figure(figsize=(10,7))\n",
    "# plt.plot(en_threshold_list, en_good_ratio_vs_en_threshold)\n",
    "# plt.ylabel('Fraction en OK')\n",
    "# plt.xlabel('En threshold')\n",
    "# # plt.savefig(fname = 'plots/oxides_calcine/en_good_ratio_vs_en_threshold_500iter_RF_constrained.png', dpi = 300)\n",
    "# plt.show()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Look at compounds"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# rf_regr = RandomForestRegressor()\n",
    "# rf_regr = joblib.load(\"rf_models/optimal_sinter_RF.joblib\")\n",
    "\n",
    "# # Featurization for RF model\n",
    "# feature_calculators = MultipleFeaturizer([\n",
    "#     cf.element.Stoichiometry(),\n",
    "#     cf.composite.ElementProperty.from_preset(\"magpie\"),\n",
    "#     cf.orbital.ValenceOrbital(props=[\"avg\"]),\n",
    "#     cf.ion.IonProperty(fast=True)\n",
    "# ])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# # Load compounds generated\n",
    "# with open('./training_data/oxides_calcine/final_compounds_RF_constrained.pkl', 'rb') as f: # For final compounds generated using constrained RL (DQN + DCN)\n",
    "#     final_compounds = pickle.load(f)\n",
    "\n",
    "# # For final compounds generated using constrained RL (DQN + DCN)\n",
    "# good, bad = check_both_multiple(final_compounds)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# print('Fraction of en OK compounds:', len(good)/(len(good) + len(bad)))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# # Convert good ones to reduced formula\n",
    "# for idx in range(len(good)):\n",
    "#     good[idx] = Composition(good[idx]).reduced_formula\n",
    "\n",
    "# # Predict sintering T of valid compounds\n",
    "# max_len = 0\n",
    "# good_sinter_T = []\n",
    "# for compound in good:\n",
    "#     chemical = Composition(compound)\n",
    "#     features = feature_calculators.featurize(chemical) # No need to scale for RF model\n",
    "#     features = np.array(features).reshape(1, -1)\n",
    "#     # print(features)\n",
    "#     sinter_T_pred = rf_regr.predict(features)[0]\n",
    "#     good_sinter_T.append(sinter_T_pred)\n",
    "#     compound = Composition(compound).reduced_formula\n",
    "#     print('Compound:', compound, (18-len(compound)+3)*' ','Sinter T:', round(sinter_T_pred))\n",
    "\n",
    "# plt.figure()\n",
    "# sns.distplot(good_sinter_T)\n",
    "# plt.xlabel('Sintering T')\n",
    "# plt.show()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Formation energy of generated compounds"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# # RANDOM POLICY\n",
    "# # Load random oxides\n",
    "# with open('./training_data/oxides_calcine/random_oxides.pkl', 'rb') as f: \n",
    "#     random_oxides = pickle.load(f)\n",
    "# form_e_random = []\n",
    "# for compound in random_oxides:\n",
    "#     form_e_random.append(predict_formation_energy(compound))\n",
    "\n",
    "# # Save predicted formation energies of random oxides                               \n",
    "# with open('./training_data/oxides_calcine/form_e_random.pkl', 'wb') as f:\n",
    "#         pickle.dump(form_e_random, f, pickle.HIGHEST_PROTOCOL)\n",
    "\n",
    "# Load predicted formation energies of random oxides \n",
    "with open('./training_data/oxides_calcine/form_e_random.pkl', 'rb') as f: \n",
    "    form_e_random = pickle.load(f)\n",
    "    \n",
    "stable_random = len([x for x in form_e_random if x <= 0])/len(form_e_random)\n",
    "print(stable_random)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# # UNCONSTRAINED DQN (sintering T only)\n",
    "# with open('./training_data/oxides_calcine/compounds_vs_iter_RF_constrained.pkl', 'rb') as f: # For final compounds generated using constrained RL (DQN + DCN)\n",
    "#     final_compounds = pickle.load(f)\n",
    "# final_compounds = final_compounds[-1]\n",
    "\n",
    "form_e_unconstrained = []\n",
    "for compound in final_compounds:\n",
    "    form_e_unconstrained.append(predict_formation_energy(compound))\n",
    "\n",
    "# # Save predicted formation energies of UNconstrained DQN                                  \n",
    "# with open('./training_data/oxides_calcine/form_e_500iter_RF.pkl', 'wb') as f:\n",
    "#         pickle.dump(form_e_unconstrained, f, pickle.HIGHEST_PROTOCOL)\n",
    "\n",
    "# # Load predicted formation energies of UNconstrained DQN\n",
    "# with open('./training_data/oxides_calcine/form_e_500iter_RF.pkl', 'rb') as f: \n",
    "#     form_e_unconstrained = pickle.load(f)\n",
    "\n",
    "# len([x for x in form_e_unconstrained if x <= 0])/len(form_e_unconstrained)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# # CONSTRAINED DQN (sintering T + en constraint)\n",
    "# # Load compounds generated\n",
    "# with open('./training_data/oxides_calcine/final_compounds_RF_constrained.pkl', 'rb') as f: # For final compounds generated using constrained RL (DQN + DCN)\n",
    "#     final_compounds = pickle.load(f)\n",
    "\n",
    "# form_e_constrained = []\n",
    "# for compound in final_compounds:\n",
    "#     form_e_constrained.append(predict_formation_energy(compound))\n",
    "\n",
    "# # Save predicted formation energies of constrained DQN                                  \n",
    "# with open('./training_data/oxides_calcine/form_e_500iter_RF_constrained.pkl', 'wb') as f:\n",
    "#         pickle.dump(form_e_constrained, f, pickle.HIGHEST_PROTOCOL)\n",
    "\n",
    "# Load predicted formation energies of constrained DQN\n",
    "with open('./training_data/oxides_calcine/form_e_500iter_RF_constrained.pkl', 'rb') as f: \n",
    "    form_e_constrained = pickle.load(f)\n",
    "\n",
    "stable_trained = len([x for x in form_e_constrained if x <= 0])/len(form_e_constrained)\n",
    "print(stable_trained)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "form_e_below_zero = {'random': stable_random,\n",
    "                    'trained': stable_trained\n",
    "            }\n",
    "with open('training_data/oxides_calcine/form_e_below_zero.json', 'w') as f:\n",
    "    json.dump(form_e_below_zero, f)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "plt.figure()\n",
    "sns.distplot(form_e_random, bins = 30, hist = True, norm_hist = True, label = 'Random', color = 'grey')\n",
    "# sns.distplot(form_e_unconstrained, hist = True, norm_hist = True, label = 'DQN', color = 'blue')\n",
    "sns.distplot(form_e_constrained, hist = True, norm_hist = True, label = 'Constrained DQN', color = 'red')\n",
    "plt.xlabel('Formation energy /eV')\n",
    "plt.legend()\n",
    "# plt.savefig('plots/oxides_calcine/form_e_comparison.png', dpi = 300)\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "np.mean(form_e_constrained), np.std(form_e_constrained)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "similarity_to_nearest_neighbor(final_compounds)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Visualizing training data of unconstrained DQN"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Evolution of final rewards with iteration (unconstrained DQN only)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# with open('./training_data/final_rewards_vs_iter_RF_constrained.pkl', 'rb') as f:\n",
    "#     final_rewards_vs_iter = pickle.load(f)\n",
    "# num_iter = 500\n",
    "# final_rewards_mean = [x.mean() for x in final_rewards_vs_iter]\n",
    "# final_rewards_std  = [x.std() for x in final_rewards_vs_iter]\n",
    "# plt.figure(figsize=(10,7))\n",
    "# markers, caps, bars = plt.errorbar(range(num_iter), final_rewards_mean, yerr = final_rewards_std, mfc= 'blue', linewidth = 2, elinewidth= 1)\n",
    "# [bar.set_alpha(0.3) for bar in bars]\n",
    "# [bar.set_color('orange') for bar in bars]\n",
    "# plt.ylabel('Final reward')\n",
    "# plt.xlabel('Training iteration')\n",
    "# plt.legend()\n",
    "# # plt.savefig(fname = 'plots/reward_vs_iter_500iter_RF_constrained.png', dpi = 300)\n",
    "# plt.show()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Evolution of compounds with iteration (unconstrained DQN only)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# with open('./training_data/compounds_vs_iter_RF.pkl', 'rb') as f:\n",
    "#     compounds_vs_iter = pickle.load(f)\n",
    "\n",
    "# # FEATURIZE GENERATED COMPOUNDS\n",
    "# for iteration in tqdm(range(len(compounds_vs_iter))): # For each iteration of compounds\n",
    "#     for idx in range(len(compounds_vs_iter[iteration])): # For each compound\n",
    "#         compounds_vs_iter[iteration][idx] = featurize_target(compounds_vs_iter[iteration][idx]) # Featurize using Magpie embeddings\n",
    "\n",
    "# # Save featurized compounds\n",
    "# with open('./training_data/compound_features_vs_iter_RF.pkl', 'wb') as f:\n",
    "#         pickle.dump(compounds_vs_iter, f, pickle.HIGHEST_PROTOCOL)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# # Load compound features\n",
    "# with open('./training_data/compound_features_vs_iter_RF.pkl', 'rb') as f:\n",
    "#     compound_features_vs_iter = pickle.load(f)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# len(compound_features_vs_iter)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Standarizing data before PCA is important."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# # Load scalar\n",
    "# scaler = StandardScaler() # Fit scaler\n",
    "# scaler.fit(s_material_inputs)\n",
    "\n",
    "# # Convert each compound features from list to np.array\n",
    "# for iteration in tqdm(range(len(compound_features_vs_iter))): # For each iteration of compounds\n",
    "#     for idx in range(len(compound_features_vs_iter[iteration])): # For each compound\n",
    "#         compound_features_vs_iter[iteration][idx] = scaler.transform(np.array([compound_features_vs_iter[iteration][idx]]))[0] # Featurize using Magpie embeddings"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# features = [] # 500 x np.array() of shape 100 x 145 (100 is no of compounds per iteration, 145 is the feature size)\n",
    "# for iteration in tqdm(range(len(compound_features_vs_iter))): # For each iteration of compounds\n",
    "#     iteration_stack = []\n",
    "#     for idx in range(len(compound_features_vs_iter[iteration])): # For each compound\n",
    "#         iteration_stack.append(compound_features_vs_iter[iteration][idx])\n",
    "#     iteration_stack = np.stack(iteration_stack)\n",
    "#     features.append(iteration_stack)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# # Stack features into one big matrix\n",
    "# full_stack = [] \n",
    "# for iteration in features:\n",
    "#     full_stack.append(iteration)\n",
    "# full_stack = np.vstack(full_stack)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Dimensionality reduction using t-SNE"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# # Perform t-SNE\n",
    "# tsne_embeddings = TSNE(n_components = 2, \n",
    "#                        perplexity = 30,\n",
    "#                        learning_rate='auto',\n",
    "#                        init='random' # random/pca initialization\n",
    "#                        ).fit_transform(full_stack)\n",
    "\n",
    "# # with open('training_data/tsne_embeddings_500iter_RF.npy', 'wb') as f:\n",
    "# #     np.save(f, tsne_embeddings)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# # Load t-SNE embeddings\n",
    "# with open('training_data/tsne_embeddings_500iter_RF.npy', 'rb') as f:\n",
    "#     tsne_embeddings = np.load(f)\n",
    "\n",
    "# with open('./training_data/final_rewards_vs_iter_RF.pkl', 'rb') as f:\n",
    "#     final_rewards_vs_iter = pickle.load(f)\n",
    "\n",
    "# final_rewards_vs_iter_stacked = torch.stack(final_rewards_vs_iter)\n",
    "# final_rewards = final_rewards_vs_iter_stacked.reshape(-1)\n",
    "# print('tsne_embeddings: ', tsne_embeddings.shape)\n",
    "# print('final_rewards:', final_rewards.shape)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# # Generate list of iterations for 50k compounds\n",
    "# iteration = []\n",
    "# for i in range(500):\n",
    "#     for j in range(100):\n",
    "#         iteration.append(i)\n",
    "# # Plot TSNE\n",
    "# def plot_TSNE(train_iter = None):\n",
    "#     fig = plt.figure(figsize=(6.5,5))\n",
    "#     if train_iter == None:\n",
    "#         plt.scatter(tsne_embeddings[:,0], tsne_embeddings[:,1], s = 0.05, alpha = 1, c = final_rewards, cmap = 'seismic') # For rewards\n",
    "#         # plt.scatter(tsne_embeddings[:,0], tsne_embeddings[:,1], s = 0.05, alpha = 1, c = iteration, cmap = 'seismic') # For iteration\n",
    "#     else:\n",
    "#         row_idx = range(100*train_iter,(100*train_iter)+100)\n",
    "#         rewards = final_rewards_vs_iter[train_iter]\n",
    "#         plt.scatter(tsne_embeddings[row_idx,0], tsne_embeddings[row_idx,1], c = rewards, s = 5, cmap = 'seismic')\n",
    "#     plt.ylabel('PC 2')\n",
    "#     plt.xlabel('PC 1')\n",
    "#     plt.ylim(-125, 125)\n",
    "#     plt.xlim(-125, 125)\n",
    "#     plt.legend()\n",
    "#     # plt.text(20,100,'Iteration %.0f' % train_iter, fontsize = 15)\n",
    "#     cbar = plt.colorbar()\n",
    "#     cbar.set_label('Reward')\n",
    "#     # cbar.set_label('Iteration')\n",
    "#     plt.clim(-1000,-400) \n",
    "#     # plt.savefig(fname = 'plots/TSNE_vs_iter_500iter_RF.png', dpi = 300) \n",
    "#     plt.show()\n",
    "#     return fig\n",
    "# plot_TSNE()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# images = []  # For saving into gif\n",
    "\n",
    "# for i in range(0, 600, 1): # For stalling the last frame\n",
    "#     j = i\n",
    "#     if i > 499: # For stalling the last frame\n",
    "#         j = 499\n",
    "#     fig = plot_TSNE(train_iter = j)\n",
    "#     # Used to return the plot as an image array \n",
    "#     # (https://ndres.me/post/matplotlib-animated-gifs-easily/)\n",
    "#     fig.canvas.draw()       # draw the canvas, cache the renderer\n",
    "#     image = np.frombuffer(fig.canvas.tostring_rgb(), dtype='uint8')\n",
    "#     image = image.reshape(fig.canvas.get_width_height()[::-1] + (3,))\n",
    "#     images.append(image)\n",
    "# imageio.mimsave('plots/TSNE_reward_vs_iter_500iter_RF.gif', images, fps = 30)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Dimensionality reduction using singular value decomposition (SVD)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# # Perform SVD\n",
    "# u, s, vh = np.linalg.svd(full_stack, full_matrices=True)\n",
    "\n",
    "# # with open('training_data/SVD_embeddings_500iter_RF.npy', 'wb') as f:\n",
    "# #     np.save(f, u)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# # Load SVD embeddings\n",
    "# with open('training_data/SVD_embeddings_500iter_RF.npy', 'rb') as f:\n",
    "#     u = np.load(f)\n",
    "\n",
    "# with open('training_data/final_rewards_vs_iter_RF.pkl', 'rb') as f:\n",
    "#     final_rewards_vs_iter = pickle.load(f)\n",
    "\n",
    "# final_rewards_vs_iter_stacked = torch.stack(final_rewards_vs_iter)\n",
    "# final_rewards = final_rewards_vs_iter_stacked.reshape(-1)\n",
    "# print('SVD_embeddings: ', u.shape)\n",
    "# print('final_rewards:', final_rewards.shape)\n",
    "\n",
    "# u_truncated = u[:, :2]\n",
    "# print('SVD_embeddings after truncation:', u_truncated.shape)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# iteration = []\n",
    "# for i in range(500):\n",
    "#     for j in range(100):\n",
    "#         iteration.append(i)\n",
    "\n",
    "# # Plot SVD\n",
    "# def plot_SVD(train_iter = None):\n",
    "#     fig = plt.figure(figsize=(6.5,5))\n",
    "#     if train_iter == None:\n",
    "#         plt.scatter(u_truncated[:,0], u_truncated[:,1], c = final_rewards, s = 0.1, cmap = 'seismic') # For rewards\n",
    "#         # plt.scatter(u_truncated[:,0], u_truncated[:,1], c = iteration, s = 0.1, cmap = 'seismic') # For iterations\n",
    "#     else:\n",
    "#         row_idx = range(100*train_iter,(100*train_iter)+100)\n",
    "#         rewards = final_rewards_vs_iter[train_iter]\n",
    "#         plt.scatter(u_truncated[row_idx,0], u_truncated[row_idx,1], c = rewards, s = 5, cmap = 'seismic')\n",
    "#     plt.ylabel('PC 2')\n",
    "#     plt.xlabel('PC 1')\n",
    "#     plt.ylim(-0.013, 0.013)\n",
    "#     plt.xlim(-0.009, .002)\n",
    "#     plt.legend()\n",
    "#     # plt.text(-0.002,0.010,'Iteration %.0f' % train_iter, fontsize = 15)\n",
    "#     cbar = plt.colorbar()\n",
    "#     plt.clim(-1000,-400)  \n",
    "#     # cbar.set_label('Iteration')\n",
    "#     cbar.set_label('Reward')\n",
    "#     # plt.savefig(fname = 'plots/SVD_reward_vs_iter_500iter_RF.png', dpi = 300)\n",
    "#     plt.show()\n",
    "#     return fig\n",
    "# plot_SVD()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# images = []  # For saving into gif\n",
    "\n",
    "# for i in range(0, 600, 1): # For stalling the last frame\n",
    "#     j = i\n",
    "#     if i > 499: # For stalling the last frame\n",
    "#         j = 499\n",
    "#     fig = plot_SVD(train_iter = j)\n",
    "#     # Used to return the plot as an image array \n",
    "#     # (https://ndres.me/post/matplotlib-animated-gifs-easily/)\n",
    "#     fig.canvas.draw()       # draw the canvas, cache the renderer\n",
    "#     image = np.frombuffer(fig.canvas.tostring_rgb(), dtype='uint8')\n",
    "#     image = image.reshape(fig.canvas.get_width_height()[::-1] + (3,))\n",
    "#     images.append(image)\n",
    "# imageio.mimsave('plots/SVD_reward_vs_iter_500iter_RF.gif', images, fps = 30)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# For NeurIPS workshop paper"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# # Load random oxides\n",
    "with open('./training_data/random_oxides.pkl', 'rb') as f: \n",
    "    random_oxides = pickle.load(f)\n",
    "\n",
    "# Load compounds\n",
    "with open('./training_data/oxides_calcine/final_compounds_roost_constrained.pkl', 'rb') as handle:\n",
    "    final_compounds = pickle.load(handle)\n",
    "\n",
    "# Formation Energy\n",
    "form_e_random = []\n",
    "form_e_trained = []\n",
    "for compound in random_oxides:\n",
    "    form_e_random.append(predict_formation_energy(compound))\n",
    "for compound in final_compounds:\n",
    "    form_e_trained.append(predict_formation_energy(compound))\n",
    "\n",
    "# Charge neutrality and electronegativity\n",
    "cn_random_good, cn_random_bad = check_neutrality_multiple(random_oxides)\n",
    "en_random_good, en_random_bad = check_electronegativity_multiple(random_oxides)\n",
    "cn_trained_good, cn_trained_bad = check_neutrality_multiple(final_compounds)\n",
    "en_trained_good, en_trained_bad = check_electronegativity_multiple(final_compounds)\n",
    "\n",
    "# Uniqueness\n",
    "random_oxides_unique = list(set(random_oxides))\n",
    "final_compounds_unique = list(set(final_compounds))\n",
    "\n",
    "results = {}\n",
    "results['form_e_random_mean'] = np.mean(form_e_random)\n",
    "results['form_e_random_std'] = np.std(form_e_random)\n",
    "results['form_e_trained_mean'] = np.mean(form_e_trained)\n",
    "results['form_e_trained_std'] = np.std(form_e_trained)\n",
    "results['charge_neutral_random'] = len(cn_random_good)/(len(cn_random_good)+len(cn_random_bad))\n",
    "results['electronegativity_balanced_random'] = len(en_random_good)/(len(en_random_good)+len(en_random_bad))\n",
    "results['charge_neutral_trained'] = len(cn_trained_good)/(len(cn_trained_good)+len(cn_trained_bad))\n",
    "results['electronegativity_balanced_trained'] = len(en_trained_good)/(len(en_trained_good)+len(en_trained_bad))\n",
    "results['unique_random'] = len(random_oxides_unique)/len(random_oxides)\n",
    "results['unique_trained'] = len(final_compounds_unique)/len(final_compounds)\n",
    "results"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "results"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "print('Formation energy (random): {} ({})'.format(results['form_e_random_mean'], results['form_e_random_std']))\n",
    "print('Formation energy (trained): {} ({})'.format(results['form_e_trained_mean'], results['form_e_trained_std']))"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "dqn_roost_EMD",
   "language": "python",
   "name": "dqn_roost_emd"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.7.10"
  },
  "vscode": {
   "interpreter": {
    "hash": "13f06d7516e179eb6b00f7bc4aef61899d405efad5361ac9f38a509de5d78c01"
   }
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}