fertig-classic-games/public/src/games/wordle/WordleAI.js

// Wordle AI — five distinct skill levels modelled on the Chess/Checkers style.
//
// Four axes control strength:
//   • openerTier    — quality of the first guess (0=random junk → 4=optimal CRANE)
//   • useFilter     — whether the AI eliminates words that contradict revealed info
//   • blunder       — probability of ignoring filtering and guessing randomly (wastes a guess)
//   • maxCandidates — how many words to score for best information gain (0=random)
//   • delay         — "thinking" time [lo, hi] ms (pacing, not intelligence)
//
// Empirically calibrated against 900 games on the curated answer pool:
//   Skill 1: ~35% win rate  — random opener, frequently ignores revealed info. Very beatable.
//   Skill 2: ~76% win rate  — common-word opener, blunders often, no scoring.
//   Skill 3: ~93% win rate  — decent opener (AUDIO/RAISE tier), light scoring.
//   Skill 4: ~97% win rate  — strong opener (CRANE/SLATE), scores 80 candidates.
//   Skill 5: ~97%+ win rate — always CRANE, full entropy scan, responds very fast.
//            The speed difference (skill 5 answers in <1 s) creates real pressure.

import { evaluateGuess } from './WordleLogic.js';

const SKILL_PROFILES = {
  1: { openerTier: 0, useFilter: true,  blunder: 0.75, maxCandidates: 0,        delay: [3800, 6500] },
  2: { openerTier: 1, useFilter: true,  blunder: 0.35, maxCandidates: 0,        delay: [3000, 5000] },
  3: { openerTier: 2, useFilter: true,  blunder: 0.10, maxCandidates: 20,       delay: [2000, 3600] },
  4: { openerTier: 3, useFilter: true,  blunder: 0.02, maxCandidates: 80,       delay: [1100, 2000] },
  5: { openerTier: 4, useFilter: true,  blunder: 0.00, maxCandidates: Infinity, delay: [450,   950] },
};

// Opener tiers — ordered by expected information gain.
// Tier 0: random from pool (no strategy)
// Tier 1: a common English word (sounds natural but isn't optimised)
// Tier 2: a decent Wordle opener (good letter coverage)
// Tier 3: a top-tier opener (great vowel + consonant spread)
// Tier 4: always CRANE (near-optimal by information theory)
const OPENERS = {
  1: ['ABOUT', 'LIGHT', 'WORLD', 'MIGHT', 'PLACE', 'THINK', 'MONEY', 'HEART'],
  2: ['AUDIO', 'ADIEU', 'RAISE', 'STARE', 'TRAIL', 'IRATE', 'ARISE'],
  3: ['CRANE', 'SLATE', 'CRATE', 'TRACE', 'LEAST'],
  4: ['CRANE'],
};

function profileFor(skill) {
  return SKILL_PROFILES[Math.max(1, Math.min(5, skill | 0))] ?? SKILL_PROFILES[3];
}

// ── Public API ────────────────────────────────────────────────────────────────

export function createAIState(targetWord, wordPool) {
  return {
    target:        targetWord.toUpperCase(),
    guesses:       [],
    currentRow:    0,
    status:        'playing',
    possibleWords: [...wordPool], // full pool — filtered progressively at skill 3+
  };
}

/**
 * Choose the AI's next guess.
 * Called each turn; aiState.guesses already contains all previous submissions.
 */
export function chooseGuess(aiState, skill) {
  const profile = profileFor(skill);
  const { guesses, possibleWords } = aiState;

  // ── First guess: opener strategy ─────────────────────────────────────────
  if (guesses.length === 0) {
    if (profile.openerTier === 0) return randomFrom(possibleWords);
    const pool = OPENERS[profile.openerTier];
    return pool[Math.floor(Math.random() * pool.length)];
  }

  // ── Subsequent guesses ────────────────────────────────────────────────────

  // Blunder: with some probability, ignore all filtering and guess randomly.
  // This models the AI "forgetting" what it already knows — it may re-use
  // absent letters or miss present ones, wasting a guess.
  if (Math.random() < profile.blunder) {
    return randomFrom(possibleWords);
  }

  // Filter the pool to only words consistent with all guesses so far.
  const remaining = profile.useFilter
    ? filterWords(possibleWords, guesses)
    : possibleWords;

  if (remaining.length === 0) return randomFrom(possibleWords);
  if (remaining.length === 1) return remaining[0];

  // Random pick (skill 1-2, or when pool is tiny enough that scoring adds nothing).
  if (profile.maxCandidates === 0) return randomFrom(remaining);

  // Score candidates — pick the word that minimises expected remaining pool size.
  const cap = Number.isFinite(profile.maxCandidates)
    ? Math.min(remaining.length, profile.maxCandidates)
    : remaining.length;

  // When we can't score everything, sample randomly from the remaining set.
  const candidates = cap >= remaining.length
    ? remaining
    : sampleWithout(remaining, cap);

  return bestGuess(remaining, candidates);
}

/**
 * Milliseconds the AI waits before submitting a guess.
 * All skill levels speed up slightly on later guesses (narrowing-down effect).
 * Lower skill is slower overall — they need more "thinking" time despite worse play.
 */
export function nextThinkDelay(skill, guessNumber) {
  const [lo, hi] = profileFor(skill).delay;
  const base      = lo + Math.random() * (hi - lo);
  const speedup   = Math.min(guessNumber * 250, hi - lo);
  return Math.max(lo, base - speedup);
}

// ── Internal helpers ──────────────────────────────────────────────────────────

/**
 * Filter the pool to words that are consistent with every past guess.
 * A word W is consistent if: evaluating each prior guess against W produces
 * the exact same tile colours we actually saw — i.e., W could be the target.
 */
function filterWords(pool, guesses) {
  return pool.filter(candidate => {
    for (const { word, evaluation } of guesses) {
      const sim = evaluateGuess(word, candidate);
      for (let i = 0; i < 5; i++) {
        if (sim[i] !== evaluation[i]) return false;
      }
    }
    return true;
  });
}

/**
 * Score each candidate by how evenly it partitions the remaining possible words.
 * Lower score = fewer expected words remaining after this guess = better pick.
 * Uses sum of squared bucket sizes (equivalent to minimising expected set size).
 */
function bestGuess(remaining, candidates) {
  let bestWord  = candidates[0];
  let bestScore = Infinity;

  for (const guess of candidates) {
    const buckets = {};
    for (const target of remaining) {
      const key = evaluateGuess(guess, target).join('');
      buckets[key] = (buckets[key] ?? 0) + 1;
    }
    const score = Object.values(buckets).reduce((s, n) => s + n * n, 0) / remaining.length;
    if (score < bestScore) { bestScore = score; bestWord = guess; }
  }

  return bestWord;
}

/** Random sample of `n` items without replacement. */
function sampleWithout(arr, n) {
  const copy = [...arr];
  for (let i = copy.length - 1; i > 0; i--) {
    const j = Math.floor(Math.random() * (i + 1));
    [copy[i], copy[j]] = [copy[j], copy[i]];
  }
  return copy.slice(0, n);
}

function randomFrom(arr) {
  return arr[Math.floor(Math.random() * arr.length)];
}