Skip to content

A simplistic implementation of DQN that works under CartPole-v0 with rendered pixels as input

Notifications You must be signed in to change notification settings

tqjxlm/Simple-DQN-Pytorch

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

2 Commits
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Simple-DQN-Pytorch

This is a simplistic implementation of DQN that works under CartPole-v0 with rendered pixels as input. It extends the implementation of pytorch's official DQN tutorial (which doesn't actually work) https://pytorch.org/tutorials/intermediate/reinforcement_q_learning.html.

It's tuned specifically for CartPole-v0 and will possibly fail for other tasks.

Requirements

  • pytorch 1.0+
  • torchvision
  • OpenAI gym
  • matplotlib
  • Pillow

Performance

Due to the randomness of the DQN algorithm (random training samples & random network initialization), the training is not deterministic. You may have to restart a few times to get a satisfying result.

The following 3 experiment results came from the same set of hyper parameters:

Trial Max reward Max 100-mean Total episodes Solved after (episodes)
0 1600 220 5000 3000
1 900 160 5000 -
2 2500 500 10000 700

The last column indicates when the 100-mean reached 200.

image not available

Last 300 episode history of training:

image not available

  • Blue curve: episode reward
  • Red curve: episode loss
  • Orange curve: 100-mean
  • Higher green line: max reward
  • Lower green line: max 100-mean

Implementation

Methods

  • Preprocess:
    • stack last 3 frames
    • grayscale, crop, resize
  • Net work:
    • conv: 4 layers with a kernel size of 5 and stride of 2
    • flatten: simply reshape all pixels from conv output
    • output: fully connect flattened features to produce 1 output for Value and 2 outputs for Advantage
  • Double DQN
  • Duelling DQN
  • Prioritized replay memory
  • Adam optimizer
  • MSE loss
  • Separate render thread

Hyper-parameters

Parameter Value
Learning rate 3e-5
Target net update (steps) 200
Batch size 256
Gamma 1
Memory size 10000
Memory alpha 0.6
Memory beta start 0.4
Memory beta frames 10000
eps start 1.0
eps end 0.01
eps decay 10

Other Observations

  • The training loss never converges, while performance keeps improving.
  • Training is very sensitive to hyperparameter changes.
  • Although the 100-mean is quite high and stable after some training, the network may still perform some terrible moves.
  • It's quite sensitive to initialization, which can only be controlled with a random seed.
  • High rate of random sample rate (epsilon) seems to ruin the training. Only a small one works.

About

A simplistic implementation of DQN that works under CartPole-v0 with rendered pixels as input

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published