/dqn-mario

[UPDATED] PyTorch Implementation of DQN and training Super Mario Bros

Primary LanguagePython

Reinforcement Learning for Super Mario Bros using DQN

Modified to work with gym-super-mario-bros

NOTE: This is an unofficial fork of the original code published as dqn-mario.

DQN

使用PyTorch實作DQN演算法，並訓練super-mario-bros，整體架構參考openai/baselines。

Warning：訓練DQN請開足夠的記憶體，Replay Buffer以預設值1000000為例至少會使用約8G的記憶體.

Dependencies

  • Python 3.6
  • PyTorch
  • gym
  • gym-super-mario-bros

Result

  • Super-Mario-Bros

使用8顆cpu在GCP上跑16個小時，RAM開24G非常足夠，但很難收斂，無法穩定過關。 訓練的影像預設位置在/video/mario/。

References

Playing Atari with Deep Reinforcement Learning
openai/baselines
transedward/pytorch-dqn
openai/gym
Kautenja/gym-super-mario-bros