Distributed A3C algorithm for playing Atari games

This repo contains the code for our work on the distributed training of RL agents for playing atari games. For further details please visit Solving Atari games with Distributed Reinforcement Learning. It's a distributed version of the A3C algorithm in Asynchronous Methods for Deep Reinforcement Learning.
We based our implementation on Tensorpack version of A3C.
It's prepared to run on large CPU clusters with Slurm manager (it was tested on up to 1500 cores)

Requirements

Python 2.7.13
Slurm 17.02.7
Tensorflow 1.2
other Python requirements as described here

Notes

In our experiments we've used Tensorflow with MKL. Our code should work with usual TensorFlow 1.2, but we neither tested nor have any benchmarks for it.

Install

git clone <this repo address>
Install Python packages : pip install -r requirements.txt
In distributed_tensorpack_mkl.sh:38 set this paths:
* EXPERIMENTS_DIR - directory where experiments will be saved
* VIRTUAL_ENV - path to virtualenv you will be using
* DISTRIBUTED_A3C_PATH - path to this repo
* TENSORPACK_PIPEDIR - path to directory for storing sockets which are used for interprocess communication

To train agent on Atari game:

Minimal command to start training:

python run_job.py -n 68 -g 60 -c 12 --use_sync --name neptune_job_name

To reproduce our best results use:

python run_job.py -n 71 -g 60 -c 12 -o adam --use_sync --name neptune_job_name -l 0.001 -b 32 --fc_neurons 128 --simulator_procs 10 --ps 4 --fc_init uniform --conv_init normal --fc_splits 4 --epsilon 1e-8 --beta1 0.8 --beta2 0.75 -e Breakout-v0 --eval_node --record_node --save_every 1000

Games

Below we showcase our solution performance on several Atari 2600 games. Left column is novice performance, middle column is after approx. 15 minutes of training and right is after approx. 30 minutes of training.

Breakout