muzero-super-mario-bros

Evaluating MuZero's performance using Super Mario Bros (OpenAI Gym)

This project evaluates MuZero using Super Mario Bros and compares it's performance to a custom implemented Deep-Q-Network with Double-Q-Learning (DDQN).

MuZero in Action

Some clips of the agent trained using MuZero in action are shown below.

System Architecture

How does it hold up against DDQN?

The algorithms were evaluated on a selected overworld level. The number of training epochs were limited by the available computational power. For more details, such as hyper-parameter tuning, please refer to the project report

Full Project Report

The full project report can be found here.

Citation

Please use the following citation when referring to any results from the repository or the report:

Udayashankar, S., 2022. Evaluating MuZero on Super Mario Bros. [online] GitHub.