@mab-kit/core

Framework-agnostic Multi-Armed Bandit core — Beta-Bernoulli Thompson Sampling with zero dependencies. Pure, deterministic (inject a seeded RNG), and easy to test.

pnpm add @mab-kit/core

API

import {
  statsFromCounts,
  probabilityOfBest,
  applyFloor,
  applyDamping,
  toIntegerPercentages,
  mulberry32,
} from "@mab-kit/core";

// 1. Observed counts -> Beta posteriors
const arms = [
  statsFromCounts({ id: "control", trials: 1000, conversions: 100 }),
  statsFromCounts({ id: "test", trials: 1000, conversions: 140 }),
];

// 2. Population-level Thompson weights (probability each arm is best)
const weights = probabilityOfBest(arms, { iterations: 10_000 });

// 3. Shape into safe, deployable integer percentages
const safe = applyDamping(prevWeights, applyFloor(weights, 0.02), 0.1);
const pct = toIntegerPercentages(safe, 100); // sums to exactly 100

Exports

• sampleGamma, sampleBeta — distribution samplers (Marsaglia–Tsang).
• statsFromCounts(counts, prior?) — {trials, conversions} → Beta(α, β).
• probabilityOfBest(arms, opts?) — Monte Carlo Thompson weights (sum to 1).
• selectArm(arms, rng?) — single-draw argmax (per-request allocation).
• applyFloor, applyDamping, normalize, toIntegerPercentages — weight shaping.
• mulberry32(seed) — seedable RNG for reproducible runs/tests.

License

MIT