PyTorch implementation of the M-DQN algorithm based on the paper Munchause Reinforcement Learning.
For a short introduction check out the Medium Article!
Discrete Action Space:
Continuous Action Space:
Compared to the original algorithm I did some changes:
- Instead of doing a hard update every 8000 frames I implemented a soft-update. By personal experience this worked better.
Comparison runs between M-DQN and DQN for the CartPole-v0 environment and LunarLander-v2.
Comparison of IQN and M-IQN for LunarLander-v2
Comparison IQN and M-IQN for Breakout
