This is the source code for my master's dissertation project Multi-Agent Cooperation in Hanabi with Policy Optimisation. As improvements have been made after the dissertation, please see tag:final-report for the version of used in my dissertation final report.

The implementation is based on Proximal Policy Optimisation and Hanabi Learning Environment.

• Python 3.7+ (because I can't survive without f-strings)
• PyTorch (currently CPU only) (someone buy me a NVIDIA laptop plz?)
• Hanabi Learning Environment

• configs/: JSON configuration files for agent training. Some are outdated, use for reference only.
• figures/: figures for this README.
• ppo/: main source files.
• scripts/: plotting scripts. Very volatile, use at own risk (I detest matplotlib).

Hanabi-Small is a smaller version of Hanabi with a maximum score of 10. Currently we can train an agent on Hanabi-Small in under 8 hours on an 8-core CPU machine.

Hanabi-Full is the full version with a maximum score of 25. We are still tuning the hyperparameters for Hanabi-Full but this is our current results.

We can also do ad hoc evaluation using trained agents.