Regression: legal-move generation still dominates search, so hard never gets past depth 3

オープン
#168 コメント 1 件 リアクション 0 件 担当者 0 名 GitHub で見る

メンテナーはふだん 1 日以内に返信

まだ誰も着手していません。

評価

難易度
5/5
見積もり時間
1週間以上
初心者へのやさしさ
38/100
issue の種類
バグ
明瞭さ
明確に書かれている
活発さ
活発
技術スタック
javascript

調査の方向性

Start with js/rules.js at legalMovesFrom, allLegalMoves and allLegalCaptures, then read the search and quiescence call sites in js/ai.js. Run npm test and use the provided Node benchmark steps to compare move generation with evaluation. Done means all acceptance criteria pass, including the generator cross-checks and the hard depth 4 timing target.

索引モデルが issue の本文から書いたものです。

説明

help wanted performance severity: medium

Summary

#86 was closed, but legal-move generation is still the main cost of the engine search. It costs about 40 times as much as evaluation. As a result, the "hard" level never finishes a search deeper than its base depth of 3, and its maxDepth: 8 cannot be reached.

Where

  • js/rules.js:243-270 (legalMovesFrom)
  • js/rules.js:302-360 (allLegalMoves, allLegalCaptures)
  • js/ai.js:16 (hard: { depth: 3, timeMs: 1000, maxDepth: 8 })
  • js/ai.js:118, js/ai.js:201 (search and quiescence call sites)

Details

For every piece, legalMovesFrom tests all 64 target squares with isPseudoLegalMove and isPathObstructed. For each candidate move it copies the whole board (applyMove) and runs isInCheck. The search calls this at every interior node, and quiescence calls allLegalCaptures, which also runs one applyMove per capture.

Measured in Node on a middlegame position:

Operation Cost
allLegalMoves 685 µs per call
allLegalCaptures 173 µs per call
evaluatePosition 17.5 µs per call
Depth-3 search 1.1 s, 10.7k nodes
Depth-4 search 16.9 s, 115k nodes

Impact

hard is allowed about 1 s, or at most 3 s on the clock (searchBudget), so iterative deepening never completes depth 4. The worker then spends the rest of its allowance on a depth-4 search that is thrown away. Hard plays at its fixed depth, and every computer move and every hint keeps a core busy for the full allowance. On phones, depth 3 alone takes longer than the 1 s allowance. The same constant also makes review scoring (see the issue on review steps) and the no-worker fallback slow.

Steps to reproduce

  1. Time searchPosition at depth 3 and depth 4 on a middlegame position, using rules.js and ai.js in Node or in a jest test.
  2. Compare the cost of allLegalMoves with the cost of evaluatePosition.

Suggested fix

  1. Generate pseudo-legal moves from piece offsets and sliding rays instead of scanning all 64 squares.
  2. Check legality only where it can fail: king moves, pinned pieces (one ray scan from the king), en passant, and positions in check.
  3. Or defer the legality test until the search actually plays a move, and count a node with no legal child as mate or stalemate.
  4. Make and unmake moves on one board (or a flat 64-entry array) instead of copying it for each move.
  5. Then add move ordering (MVV-LVA, killer moves).

Keep the public functions in rules.js as wrappers so callers and tests don't change. The existing generator cross-check tests ("quick generators agree") act as a perft-style safety net.

Acceptance criteria

  • allLegalMoves is at least 5 times faster on the benchmark position.
  • Hard completes depth 4 within its 1 s allowance on desktop.
  • All rules tests and generator cross-checks pass unchanged.

Getting started

  • Relevant files: js/rules.js, js/ai.js, tests/unit/rules.test.js, tests/unit/ai-strength.test.js
  • How to run tests: npm test

Regression of #86.


Filed by repo-audit/full-audit at 01c6a35. Severity: medium. Effort: M. audit-key: performance|js/rules.js|legalMovesFrom

主要言語
JavaScript
スター
4
フォーク
2
平均マージ
22時間 16分
マージ済み PR(30日)
9

環境構築

はじめの一歩

  1. issue を最後まで読み、次にプロジェクトのコントリビューションガイドを読みます。
  2. 着手することを issue にコメントします — 二人が同じ作業をするのを防げます。
  3. リポジトリをフォークし、ブランチを切って変更します。
  4. issue 番号を参照したプルリクエストを送ります。

TMHSDigital/Static_Chess のほかの issue

TMHSDigital/Static_Chess の issue をすべて見る

似ている issue

JavaScript の issue をもっと見る

新しい issue をメールで受け取る

初心者向けの GitHub issue を短くまとめたダイジェスト。