This project implement some reinforcement learning algorithms, including DQN, Double DQN, C51, etc. It contains subset of algorithms from dopamine. It's for self study.

• python == 3.6, other version may work as well.
• tensorflow >= 1.6.0
• retro, video game emulator for RL.
• ffmpeg (optional), for record video.

Using C51, After training 48 hour on my GTX 960M.

# train the model
python train.py

# record actions
# python eval.py --record_dir <path to store record file>
python eval.py --record_dir record

# make video. Need to have ffmpeg in the path.
# python -m retro.scripts.playback_movie <record file>...
python -m retro.scripts.playback_movie .\record\Breakout-Atari2600-Start-000000.bk2

• Using a memory pool with 100 million elements will consume about 7 GB memory.
• C51 has faster convergence speed.

• Implement store and restore for memory pool.
• Add more algorithms.

• dopamine, a research framework for fast prototyping of reinforcement learning algorithms.
• baseline, OpenAI Baselines: high-quality implementations of reinforcement learning algorithms.
• DQN-tensorflow, Tensorflow implementation of Human-Level Control through Deep Reinforcement Learning.