MODEL STEWARD VI · AXELROD'S ROUND under KEEPER ERIS
The Bandit Saloon
Gambler agents pull five one-armed bandits; UCB1, ε-greedy and random gamblers share a tally board and the clever ones find the best arm.
AMBER UCB1 (Auer, Cesa-Bianchi & Fischer, 2002) picks the arm maximising mean + √(2 ln t / n) and has logarithmic regret (standard result, not re-fetched here).
MODEL Three gamblers, Ucb, Eps and Rand, play hidden Bernoulli arms with payout chances 0.2, 0.35, 0.5, 0.55 and 0.7. The board posts regret.
LIT on load: over 5000 seeded pulls, UCB1's regret is under a fifth of random's, and it picks the best arm more than 80% of the time.
Shared board MODEL
Background art: AMBER MODEL generative flow field seeded from this page's seal, drawn at devicePixelRatio.
proving…
Seal sha256(name|slug|blurb): 74008074044dfb9118b9145aac3ea80ba25800c1c3c30cb10605ae8abc5cef19