Python 3.6 + PyTorch

This implementation is not designed for high performance and flexibility. These are the main DQN extensions in a simple and clean code, solving classic CartPole problem. It's basic playground for understanding algorithms, supplementary material for some great RL course.

Notebooks work on Colaboratory, no setup required. Open and train from scratch, it takes one minute.

1. Double Dueling DQN colab
2. Prioritized Experience Replay colab
3. DQN with GRU RNN and n-step updates colab
4. Implicit Quantile Networks (IQN) colab paper

See also:

• higgsfield/RL-Adventure
• google/dopamine