AI research
Sweeper
A Minesweeper AI system that benchmarks symbolic reasoning, exact inference, neural networks, and hybrid agents against the same boards.
The project
Sweeper studies how an agent should make decisions in Minesweeper. It starts with symbolic proofs, switches to exact mine probabilities once deduction runs out, and evaluates neural and hybrid strategies in the same board environment. The browser interface makes the work inspectable through a playable board, assisted modes, and published benchmark results.
01
Reasoning first
Turns visible board clues into constraints and identifies provably safe moves before it ever guesses.
02
Exact probabilities
Counts valid frontier assignments to rank candidate moves when the board has no certain answer.
03
Comparable strategies
Runs symbolic, neural, and hybrid agents through one reproducible environment so the numbers mean something.
Built with
Primary
- Python
- PyTorch
- NumPy
- React
- TypeScript
- Cloudflare
Other
- Gymnasium
- pytest