Minesweeper RL Arcade
Retro arcade cabinet where you race reinforcement-learning policies on 8x8 Minesweeper
Episodes & demos
- S01E01 Play the arcadehttps://mines.livive.net/ ► Watch (opens in a new tab)
About
Minesweeper RL Arcade is an in-browser retro arcade cabinet built around a deterministic 8x8 Minesweeper engine and a library of ONNX-exported reinforcement-learning policies. It has five modes: PLAY, where you clear a seeded board with a guaranteed-safe first click; AI, where a selected checkpoint plays the same seed with live confidence, per-cell distributions, and a legal-action mask; RACE, a side-by-side human-vs-policy comparison on an identical seed; CONCURRENT, where every checkpoint in the library steps through one board in lockstep with a CLEARED/DETONATED/STUCK scoreboard and points-sorted leaderboard; and MODELS, a details panel for each manifest checkpoint (family, method, architecture, training recipe, file size, and fixed-seed evaluation). The v2 library ships nine documented checkpoints across four families — a uniform baseline, solver-imitation behaviour clones, REINFORCE policy-gradient models, and kernel/single-conv ablations — evaluated from 0% to 61.4% win rate. The app is a static ES-module bundle that fetches models/manifest.json and runs inference through the pinned onnxruntime-web CDN; the engine, featurization, and action-mask contract stay frozen so scores remain comparable across checkpoints. It deploys to GitHub Pages from main at mines.livive.net, guarded by a smoke workflow that asserts the app, the manifest, and every model file are served live.
Summary
An in-browser retro arcade cabinet running ONNX-exported reinforcement-learning policies on a deterministic 8x8 Minesweeper engine. Play a seeded board, watch the AI reason with a live action mask and confidence heatmap, race a checkpoint on the same seed, or run the whole nine-model library in CONCURRENT mode.
Tech
- JavaScript
- HTML
- CSS
- ONNX Runtime Web
- Python
- PyTorch
- Stable-Baselines3
Tags
- arcade-cabinet
- browser-game
- minesweeper
- onnx
- onnxruntime
- policy-imitation
- reinforcement-learning