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.

Paper: 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

Learning curve 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 10 minutes.

• Learning curves (mean_episode_rewards) of other games in an hour.

How to use

Dependencies

• paddlepaddle>=2.0.0
• parl>=2.1.1
• gym==0.12.1
• atari-py==0.1.7
• opencv-python

Distributed Training:

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

xparl start --port 8010 --cpu_num 32

Note that it is not necessary to run the command each time before training. We can reuse the xparl cluster for distributed training if we have started it before. documentation

Then we can start the distributed training by running:

python train.py

Reference

• Parl Cluster Setup.
• deepmind/scalable_agent
• Ray