BlokusDuo — AI agents for a competitive turn-based strategy game

Research code + experiments exploring search and reinforcement learning for Blokus Duo. This project compares classical planning (Minimax, MCTS) and learned components (CNN evaluation/policy) across large-scale simulation.

---

Mission

Understand what makes agents strong at spatial, combinatorial games by combining:

• principled search algorithms,
• learned evaluation functions,
• and rigorous simulation-based evaluation.

Project intention

This repository is meant to be:

• a research/engineering artifact (reproducible experiments, analysis notebooks),
• a portfolio piece showcasing applied RL + search,
• and a foundation for future work (stronger training loops, faster simulation, better evaluation).

---

Results at a glance

• Ran hundreds of thousands of simulations across multiple agent matchups
• Compared:
  • Minimax variants
  • MCTS (with/without neural guidance)
  • PPO (baseline RL approach)

The README and notebooks summarize findings; the paper provides a deeper narrative.

---

Paper / write-up

• Main paper: https://docs.google.com/document/d/1t95HiT_Vk48AHe5BSArWpV6ObkvVYfmjIF5mwMdtEJw/edit?usp=sharing
  (Consider also exporting a PDF into docs/ so the repo is self-contained.)

---

Demo / images

Recommended to add:

• docs/images/board.png — sample board state
• docs/images/pieces.png — pieces overview (you already have pieces*.png)
• docs/images/results.png — headline plot from comparisons

Example:

docs/images/
├── board.png
├── pieces.png
└── results.png