Original Research
Self-play with PPO, search, and board-state encoders
The original project focused on training and evaluating Blokus Duo agents with policy and value modeling, then comparing how search changed the quality of move selection.
Blokus Duo ML Agents
Play a reduced-scope agent demo built for the portfolio site.
The original Blokus Duo work was a research-heavy self-play project. This route turns that work into something visitors can actually use: a browser-friendly board, several agent personalities, and a ruleset small enough to finish in a few minutes without losing the diagonal-placement strategy that made the project interesting.
What this route is for
- Show the strategic feel of the project without requiring Python or model hosting.
- Expose multiple agent styles so visitors can compare different move-selection tables.
- Separate the original ML work from the website adaptation in a way that is honest and interactive.
Website Adaptation
Static move tables that keep the portfolio fully playable
The web demo distills those ideas into a reduced 9x9 ruleset with static policy tables. That makes the page fast, deployable on GitHub Pages, and still useful as a tactical showcase.
Browser Demo
Play against static agent profiles
Each profile uses a different precomputed move-ranking table. They are not live neural nets in the browser, but they let the site demonstrate how different policy preferences change openings, space control, and endgame shape.
Playable Web Adaptation
Reduced-board Blokus Duo, tuned for fast ML-style matchups
This browser version keeps the diagonal-only expansion rule from Blokus while shrinking the board to 9x9 and limiting each side to six pieces. The agent profiles are static move-ranking tables designed to mimic different self-play personalities without shipping full model inference into the site.
- First move must cover your home corner.
- Your later pieces must touch your color only at corners, never edge-to-edge.
- When both sides run out of legal diagonal anchors, highest coverage wins.
Match Setup
Choose your side and the policy table you want to face
Prefers long diagonals that race toward the opponent corner and force awkward replies.
Turn 1
Your board
Blue opens from the top-left corner. Pick a piece, rotate it, and place on a highlighted anchor.
Previewing Seed from A1. Click a dotted anchor to place it.