纯前端 Web 五子棋对弈应用：完全自研的迭代加深 α-β + VCF 算杀引擎跑在 Web Worker 里，零第三方依赖、零构建、零下载，可直接托管在 GitHub Pages。
中文说明见 README_CN.md。 在线试玩：https://justa-cai.github.io/gomoku/ · 源码：https://github.com/justa-cai/gomoku
- Self-written engine — iterative-deepening Negamax with alpha-beta pruning, Zobrist transposition table, killer/history heuristics and a VCF (victory-by-continuous-fours) threat-space search. No Stockfish, no third-party library, no WASM binary.
- Zero dependencies — no build step, no CDN, no image/font assets. Stones are drawn procedurally with Canvas radial gradients. The whole app is plain source files (< 100 KB).
- Honest stats — the side panel shows the real search depth / nodes / NPS / time / score of every move. Sub-optimal moves taken by low difficulty levels are explicitly marked sub-optimal; VCF kills are marked VCF.
- Three modes — vs AI (6 levels, play black or white), local 2-player, and AI-vs-AI for calibration and spectating.
- Freestyle rules — 15×15, first to five-or-more in a row wins, no renju forbidden moves. Full board = draw.
| Level | Name | Depth | Time cap | VCF | Sampling |
|---|---|---|---|---|---|
| 1 | 入门 | 1 | 200 ms | off | temp 25000, blunder 30% |
| 2 | 初级 | 2 | 500 ms | off | temp 10000, blunder 12% |
| 3 | 中级 | 4 | 1 s | off | temp 3000 |
| 4 | 高级 | 6 | 1.5 s | on | — |
| 5 | 大师 | 8 | 2.5 s | on | — |
| 6 | 宗师 | 12 | 4 s | on | — |
At every level two hard rules bypass all weakening: if the engine can complete five in one move it does; if the opponent threatens five it must block. A losing-cap keeps sampled moves from looking broken instead of weak.
UI (Canvas board / panels / menus, ES modules)
│
Referee layer (rules.js: legality, five-in-a-row, draw, undo)
Difficulty layer (difficulty.js: single source of truth for levels & sampling)
│
Engine bridge (bridge.js: seq-matched async facade, stale-reply dropping)
│
engine.worker.js (Web Worker: iterative deepening α-β + TT + killers + VCF)
Scores are pattern-table based (FIVE 10M / open-four 1M / four 100k / open-three 90k / …) computed with a sliding 5-cell window model that handles jump shapes (X_XXX) naturally; the "distinct completion points" rule distinguishes open fours from simple fours. Leaf evaluation is O(1) via incrementally maintained per-line pattern sums.
python3 server.py # → http://127.0.0.1:6326/(Any static file server works. Opening index.html via file:// won't —
Web Workers require HTTP.)
Push to a GitHub Pages repo; .nojekyll is included. No headers, no
cross-origin isolation, no build pipeline needed.
MIT — see LICENSE and NOTICE.md. All code in this repository is original; algorithm background references are listed in NOTICE.md.