Skip to content

Latest commit

 

History

History
159 lines (114 loc) · 6.09 KB

README.md

File metadata and controls

159 lines (114 loc) · 6.09 KB
Raw

License: MIT

os os os

This repo is support for a global project - to create a smart ML bot for a browser-based multiplayer battle royale game surviv.io

Autoregressive neural environment for training an RL-agent by using expert trajectories only

Game surviv.io does not have any environment in which it would be possible to train an RL agent. Environment is a function that gets the current state and the agent's action and returns the next state and reward: state_next,reward=ENV(state_curr,action). That's why we present:

  1. our neural environment-engine (deep learning models) which perform all required features of the real game environment needed for training bots locomotion:

  1. our DQN RL bot (trained in neural engine) which aims to run up to all the stones that fall into his field of vision. We will use this proof-of-concept technology to train our agent to loot vital items and other skills in a global project.

Motivation

In all our experiments environment state is the RGB frame. The development of this technology will allow some designed robots to be trained in cooking/cleaning/repairing with huge video information from YouTube and twitch.

Playable demo app

We present an interactive demo application to demonstrate the work of a neural engine with the ability to control the movement of the agent by the user. Moreover, in this application, the user can watch the game of a trained RL agent. Main goal our bot - to run up to all the stones that fall into his field of vision. Anyone can run interactive demo application on their machine (see below installation guides).

Windows installation guides

For windows users we offer to compete with our trained RL agent. Main goal is to collect a reward for running up to the stones at the frame for the same number of steps. Good luck, have fun!

Initial usage

1. In anaconda prompt: dowland or clone this GitHub repository

2. Create python virtual environment and install requirements

conda create -n survivio_venv python=3.8
conda activate survivio_venv
python -m pip install -r requirements.txt

3. Run the agent

python run_demo.py

Later usage

1. Activate python environment

conda activate survivio_venv

2. Run the agent

python run_demo.py
Ubuntu\MacOS installation guides

For ubuntu\mac OS users we offer to compare a random agent with our trained RL agent. Main goal agents is to collect a reward for running up to the stones at the frame for the same number of steps. Press any key to update agents states. Just watch and have fun!

Initial usage

1. In terminal: clone GitHub repository

git clone https://github.com/Laggg/neural_env_surviv

2. Create python virtual environment and install requirements.txt

cd neural_env_surviv
python -m venv survivio_venv
source survivio_venv/bin/activate
python -m pip install -r requirements.txt

3. Run the agent

python run_demo.py

Later usage

1. Activate python environment

source survivio_venv/bin/activate

2. Run the agent

python run_demo.py

Our DL CV experiments

Experiments demo 
Сравнение генеративных моделей S_next=model(S_curr,action) (плашки гифок слева направо):
  - init state (RGB frame)
  - Loss = 0*Lgan + MSE
  - Loss = 0*Lgan + MSE + PL/100
  - Loss = 0*Lgan + MAE + PL/100
  - Loss = 0*Lgan + 3*MAE + PL/100 (best, in DemoApp)
  - Loss = Lgan/100 + 3*MAE + PL/100 (pix2pix)


Другие эксперименты для generative models without GAN (плашки гифок слева направо):
  - 3*MAE+PL/100+1*SSIM
  - 3*MAE+PL/100+1*LPIPS
  - 3*MAE+PL/100+PieAPP(stride=10)/4
  - 3*MAE+PL/100+PieAPP(stride=7)/2
  - 3*MAE+PL/100+2*HaarPSI
  - 3*MAE+PL/100+2*SSIM (веса утеряны, метрики хуже)
  - 3*MAE+PL/100+3*DISTS (веса утеряны, метрики хуже)

  • 1. generative models without GAN
  • 2. additional losses, /mssim/style-texture loss/perceptual path length/
  • 3. generative models with GAN (pix2pix)
  • 4. VQ-VAE/GameGAN/Dreamer2
  • 5. RL для приближения агента к кустам/камням/луту (в зависимости от качества нейронного движка)
  • 6. интерактивный фронт для взаимодействия человека с нейронным движком (чтобы можно было поиграть игру, движком которой была бы нейронка)

Team/video/presentation

  1. Dmitry Fomin (RL) ods: @dfomin tg: @dfomin
  2. Pavel Danilov (FD) ods: @Pavel tg: @KalabiYauu
  3. Dmitry Gaus (CV) ods: @QwertyNice tg: @QwertyNice
  4. Klokov Alexey (CV,RL), ods: @laggg tg: @laggg5