RainbowDQN for Super Mario Bros.

This project involved training a deep reinforcement learning agent using modern techinques. It involved reading and implementing algorithms from current research. The main paper used was "Combining Improvements in Deep Reinforcement Learning" which is a paper combining many techniques, published by DeepMind, 2017.

Videos

Full Game Performance

Multi-level Training 1

Multi-level Training 2

Multi-level Training 3

Implementation

Technologies:

• PyTorch for Deep Neural Networks (CNNs, Custom Layers)
• CUDA acceleration with PyTorch
• Reinforcement Learning algorithms developed by our group
• Environment provided by Christian Kauten (2018)