Fence/ACER_tf

Implementation for ACER in tensorflow and sonnet by deepmind

Implementation of ACER (Actor-Critic with Experience replay)

Contains the tensorflow and sonnet implementation for SAMPLE EFFICIENT ACTOR-CRITIC WITH EXPERIENCE REPLAY by Ziyu Wang, Victor Bapst et al from Deepmind (https://arxiv.org/abs/1611.01224).

The current version is tested only for MuJoCo gym environments,.

Major dependencies

• Tensorflow v1.3 (https://www.tensorflow.org/install/)
• Sonnet (https://deepmind.github.io/sonnet/)
• Python 2.7

Running

python train.py --model_dir ./tmp_model/ --env InvertedPendulum-v1 --eval_every_sec 60 --num_agents 4

See python train.py --help for a full list of options.

You can monitor training progress in Tensorboard:

tensorboard --logdir=/tmp_model/

Components

• train.py contains the main method to start training.
• agent.py contains the code for the agent threads and actual ACER algortihm
• advantage_net.py contains code for building the stochasitic dueling net
• policy_net.py contains code for building the policy network
• memory.py contains the memory class for experience replay

References

• ACER, ICLR 2017 (https://arxiv.org/abs/1611.01224)
• Denny Britz a3c implementation (https://github.com/dennybritz/reinforcement-learning/tree/master/PolicyGradient/a3c)
• TF (https://www.tensorflow.org)
• Sonnet (https://deepmind.github.io/sonnet/)