tree: 4cf1ce5c2f52aaa01d76245425f031b2274a7803 [path history] [tgz]
  1. README.md
  2. a3c.py
  3. launcher.py
  4. rl_data.py
  5. sym.py
example/reinforcement-learning/a3c/README.md

A3C Implementation

This is an attempt to implement the A3C algorithm in paper Asynchronous Methods for Deep Reinforcement Learning.

Author: Junyuan Xie (@piiswrong)

The algorithm should be mostly correct. However I cannot reproduce the result in the paper, possibly due to hyperparameter settings. If you can find a better set of parameters please propose a pull request.

Note this is a generalization of the original algorithm since we use batch_size threads for each worker instead of the original 1 thread.

Usage

run python a3c.py --batch-size=32 --gpus=0 to run training on gpu 0 with batch-size=32.

run python launcher.py --gpus=0,1 -n 2 python a3c.py to launch training on 2 gpus (0 and 1), each gpu has two workers.