Reinforcement learning based agents to playing Atari games.

• Deep Q-network(DQN)

  • Double DQN
  • Prioritized Replay
  • Dueiling network

• Asynchronous Advantage Actor-Critic(A3C)

  • Multiprocess support
  • Multiprocess Cuda support
  • LSTM based model
  • Generalized Advantage Estimate (GAE)
  • Frame stacking based model

• BreakoutDeterministic-v4
• PongDeterministic-v4
• KungFuMasterDeterministic-v4
• BoxingDeterministic-v4
• SapecInvadersDeterministic-v4

• DQN
• TensorBoard support
• Double DQN
• Prioritized replay
• Dueling network
• Train model for Pong
• Achive 300+ score on breakout
• A3C Agent for KungFuMasterDeterministic-v4
• A3C Agent for BoxingDeterministic-v4
• Parallel processing for A3C
• LSTM layer for A3C to replace frame stacking

The Apache-2.0 License. Please see the license file for more information.

• https://github.com/yandexdataschool/Practical_RL
• https://github.com/openai/baselines/blob/master/baselines/deepq/replay_buffer.py
• https://towardsdatascience.com/tutorial-double-deep-q-learning-with-dueling-network-architectures-4c1b3fb7f756
• https://pytorch.org/docs/stable/notes/multiprocessing.html
• https://github.com/ikostrikov/pytorch-a3c
• https://arxiv.org/pdf/1506.02438.pdf
• https://arxiv.org/abs/1602.01783