| --- |
| license: mit |
| pretty_name: "Minesweeper Solver Benchmark (win-rate by difficulty)" |
| language: |
| - en |
| tags: |
| - minesweeper |
| - solver |
| - constraint-satisfaction |
| - probability |
| - benchmark |
| - csp |
| task_categories: |
| - other |
| size_categories: |
| - n<1K |
| configs: |
| - config_name: winrate |
| data_files: data/minesweeper-benchmark.json |
| --- |
| |
| # Minesweeper Solver Benchmark — Win-Rate by Difficulty |
|
|
| By **[LK Forge](https://lkforge.com)**: privacy-first browser tools & games, nothing to install, nothing uploaded. |
|
|
| Full-game win-rate for the [LK Forge Minesweeper Solver](https://lkforge.com/tools/puzzles/minesweeper-solver/) |
| across the three standard difficulties. The shipped engine turns every revealed number into |
| a constraint and applies **counting → the subset rule → exact enumeration of all valid mine |
| arrangements**; when logic is exhausted it opens the hidden cell with the lowest computed mine |
| probability. 1,000 seeded games per difficulty. **Rules:** classic board sizes, the first click is |
| guaranteed safe (a single non-mine cell in the centre — not a guaranteed large opening), and a game |
| counts as a win only when every non-mine cell is cleared. |
|
|
| - **How the solver works:** https://lkforge.com/tools/puzzles/minesweeper-solver/how-the-minesweeper-solver-works.html |
| - **Try it:** https://lkforge.com/tools/puzzles/minesweeper-solver/ |
| - **Producer:** [LK Forge](https://lkforge.com) — client-side AI games, solvers, and tools. |
|
|
| ## Headline numbers (1,000 games each) |
|
|
| | Difficulty | Board | Mines | Win rate | Solved with **no guess** | Avg guesses/game | |
| |---|---|---|---|---|---| |
| | Beginner | 9×9 | 10 | **95.5%** | 78.5% | 0.30 | |
| | Intermediate | 16×16 | 40 | **85.6%** | 52.4% | 0.89 | |
| | Expert | 30×16 | 99 | **36.2%** | 3.8% | 3.82 | |
|
|
| Expert Minesweeper is famously hard even for optimal play — under this **safe-first-click** rule the |
| best-known solvers land in the **~34–40%** range on classic 30×16 Expert, so 36.2% is squarely in the |
| expert-solver band. Win rates quoted **above** this band (e.g. ~54%) generally use the easier |
| **guaranteed-opening** rule, where the first click always reveals a large zero-region; that is a |
| different rule, not a stronger solver. The steep drop from Beginner to Expert is driven by how often a |
| position **cannot be resolved by logic and forces a guess**: 78.5% of Beginner games need no guess at |
| all, versus 3.8% of Expert games. |
|
|
| ## Files |
|
|
| - `data/minesweeper-benchmark.json` — per difficulty: `winPct`, `solvedNoGuessPct`, `avgGuessesPerGame`. |
|
|
| ## Method |
|
|
| - **Engine:** the shipped `minesweeper-engine.js` (`LKMines`), used unmodified as the decision-maker. |
| - **Play policy:** flag proven mines, open proven-safe cells; when none exist, open the hidden |
| cell with the minimum exact mine probability. |
| - **Reproducible:** seeded RNG for mine placement; center first click (guaranteed safe). |
|
|
| ## Reproduce |
|
|
| ```bash |
| node scripts/benchmark-minesweeper.mjs |
| ``` |
|
|
| ## Citation |
|
|
| > LK Forge (2026). *Minesweeper Solver Benchmark — Win-Rate by Difficulty.* |
| > https://lkforge.com/tools/puzzles/minesweeper-solver/ |
|
|
| Everything runs client-side, no login, no tracking. |
|
|
