Based on PARL, the A2C algorithm of deep reinforcement learning has been reproduced, reaching the same level of indicators as the paper in Atari benchmarks.
A2C is a synchronous, deterministic variant of Asynchronous Advantage Actor Critic (A3C). Instead of updating asynchronously in A3C or GA3C, A2C uses a synchronous approach that waits for each actor to finish its sampling before performing an update. Since loss definition of these A3C variants are identical, we use a common a3c algotrithm parl.algorithms.A3C for A2C and GA3C examples.
Please see here to know more about Atari games.
Mean episode reward in training process after 10 million sample steps.
Performance of A2C on various envrionments
- paddlepaddle>=1.6.1
- parl>=1.2.1
- gym==0.12.1
- atari-py==0.1.7
At first, We can start a local cluster with 5 CPUs:
xparl start --port 8010 --cpu_num 5Note 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