IMPALA

Reproduce IMPALA with PARL

Based on PARL, the IMPALA algorithm of deep reinforcement learning is reproduced, and the same level of indicators of the paper is reproduced in the classic Atari game.

• IMPALA in Impala: Scalable distributed deep-rl with importance weighted actor-learner architectures

Atari games introduction

Please see here to know more about Atari games.

Benchmark result

Result with one learner (in a P40 GPU) and 32 actors (in 32 CPUs).

• PongNoFrameskip-v4: mean_episode_rewards can reach 18-19 score in about 7~10 minutes.

• Results of other games in an hour.

How to use

Dependencies

• paddlepaddle>=1.6.1
• parl
• gym==0.12.1
• atari-py==0.1.7

Distributed Training:

At first, We can start a local cluster with 32 CPUs:

xparl start --port 8010 --cpu_num 32

Note that if you have started a master before, you don't have to run the above command. For more information about the cluster, please refer to our documentation

Then we can start the distributed training by running:

python train.py

Reference

• Parl Cluster Setup.
• deepmind/scalable_agent
• Ray