Asyncronous RL in Tensorflow + Keras + OpenAI's Gym

This is a Tensorflow + Keras implementation of asyncronous 1-step Q learning as described in "Asynchronous Methods for Deep Reinforcement Learning".

Since we're using multiple actor-learner threads to stabilize learning in place of experience replay (which is super memory intensive), this runs comfortably on a macbook w/ 4g of ram.

It uses Keras to define the deep q network (see model.py), OpenAI's gym library to interact with the Atari Learning Environment (see atari_environment.py), and Tensorflow for optimization/execution (see async_dqn.py).

Requirements

• tensorflow
• gym
• [gym's atari environment] (https://github.com/openai/gym#atari)
• skimage
• Keras

Usage

Training

To kick off training, run:

python async_dqn.py --experiment breakout --game "Breakout-v0" --num_concurrent 8

Here we're organizing the outputs for the current experiment under a folder called 'breakout', choosing "Breakout-v0" as our gym environment, and running 8 actor-learner threads concurrently. See this for a full list of possible game names you can hand to --game.

Visualizing training with tensorboard

We collect episode reward stats and max q values that can be vizualized with tensorboard by running the following:

tensorboard --logdir /tmp/summaries/breakout

This is what my per-episode reward and average max q value curves looked like over the training period:

Evaluation

To run a gym evaluation, turn the testing flag to True and hand in a current checkpoint file:

python async_dqn.py --experiment breakout --testing True --checkpoint_path /tmp/breakout.ckpt-2690000 --num_eval_episodes 100

After completing the eval, we can upload our eval file to OpenAI's site as follows:

import gym
gym.upload('/tmp/breakout/eval', api_key='YOUR_API_KEY')

Now we can find the eval at https://gym.openai.com/evaluations/eval_uwwAN0U3SKSkocC0PJEwQ

Next Steps

See a3c.py for a WIP async advantage actor critic implementation.

Resources

I found these super helpful as general background materials for deep RL:

• David Silver's "Deep Reinforcement Learning" lecture
• Nervana's Demystifying Deep Reinforcement Learning blog post

Important notes

• In the paper the authors mention "for asynchronous methods we average over the best 5 models from 50 experiments". I overlooked this point when I was writing this, but I think it's important. These async methods seem to vary in performance a lot from run to run (at least in my implementation of them!). I think it's a good idea to run multiple seeded versions at the same time and average over their performance to get a good picture of whether or not some architectural change is good or not. Equivalently don't get discouraged if you don't see performance on your task right away; try rerunning the same code a few more times with different seeds.
• This repo has no affiliation with Deepmind or the authors; it was just a simple project I was using to learn TensorFlow. Feedback is highly appreciated.