Function runOneTournament

app/scripts/rank-base-models.ts:82–129

(name: string, matchesSlice: typeof matches, epochs: number)

});

const runOneTournament = (name: string, matchesSlice: typeof matches, epochs: number) => {
  const ratingsCache: Record<string, { mu: number; sigma: number; matches: number }> = {};

  const getRating = (model: string) => ratingsCache[model] ?? { mu: 24, sigma: 1, matches: 0 };

  const updateRating = (model1: string, model2: string, score: number) => {
    const rating1 = getRating(model1);
    const rating2 = getRating(model2);

    const [team1, team2] = rate([[rating1], [rating2]], {
      score: [score, 1 - score],
      model: bradleyTerryFull,
    });

    const newRating1 = team1?.[0];
    const newRating2 = team2?.[0];

    if (!newRating1 || !newRating2) throw new Error("Expected new rating");

    ratingsCache[model1] = { ...newRating1, matches: rating1.matches + 1 };
    ratingsCache[model2] = { ...newRating2, matches: rating2.matches + 1 };
  };

  for (let i = 0; i < epochs; i++) {
    for (const match of shuffle(matchesSlice)) {
      if (match.score !== null) updateRating(match.model1, match.model2, match.score);
    }
  }
  const ratings = Object.entries(ratingsCache)
    .map(([model, r]) => ({ model, ...r, ordinal: ordinal(r) }))
    .sort((a, b) => b.mu - a.mu);

  const ratingScale = 1200 / rating().mu;

  console.log(`${name} (${matchesSlice.length} matches, ${epochs} epochs)`);
  console.log(
    table([
      ["Model", "Mean", "Std Dev", "Ordinal", "Matches"],
      ...ratings.map((r) => [
        r.model,
        (r.mu * ratingScale).toFixed(0),
        (r.sigma * ratingScale).toFixed(1),
        (ordinal(r) * ratingScale).toFixed(0),
        r.matches / epochs,
      ]),
    ]),
  );
};

runOneTournament("All", matches, 50);
runOneTournament("All (augmented)", augmentedMatches, 20);