a3c-tensorflow

This repo contains python code for replicating the asynchronous advantage actor-critic algorithm as described in https://arxiv.org/pdf/1602.01783.pdf

Requirements

• tensorflow
• scipy
• gym (Atari)
• skimage

Training

For training a3c algorithm in BreakoutDeterministic-v3 using 8 parallel actor learner threads execute the following command:

python a3c.py --game BreakoutDeterministic-v3 --num_concurrent 8

Testing

For testing a trained a3c agent execute the folowing command

python a3c.py  --game BreakoutDeterministic-v3 --checkpoint_path path_to_checkpoint --testing True

Results

Below you can find 2 plots of training a3c in Breakout and Pong

Code and Algorithm explanation

Full explanation can be found here: https://papoudakis.github.io/announcements/policy_gradient_a3c/

Resources

https://github.com/miyosuda/async_deep_reinforce

https://github.com/coreylynch/async-rl

https://github.com/muupan/async-rl