Glossary · Term

multi-armed bandit

← all terms

Definition

Plain language

A classic decision puzzle about when to keep using what already works versus trying something new that might be better.

As stated in the literature

A reinforcement-learning framework formalizing the explore-exploit tradeoff over actions with unknown payoffs; LoopTrap's strategy selection uses bandit-style exploration weighted by behavioral priors.

Also called: bandit

Why it matters: It captures the universal explore-versus-exploit tradeoff, guiding systems that must balance using what works against discovering something better.

For example, it's like deciding each night whether to revisit your favorite restaurant or try a new one that might be even better.

Heard on the show

“That exact term is the upper-confidence-bound rule — UCB — the classic exploration bonus from bandit problems and game-tree search.”
Episode 162 — The Empty-Lake Proof: Why More Rollouts Stop Helping Reasoning Models

Mentioned in 4 episodes

  1. 162
    The Empty-Lake Proof: Why More Rollouts Stop Helping Reasoning Models
  2. 106
    Giving Agents a Notebook Instead of New Weights: How ExpGraph Lets Frozen Models Learn
  3. 067
    An AI Just Solved a 1996 Erdős Problem—and the Simplest Agent Won
  4. 030
    Why Your AI Agent Won't Stop Working — and Each Model Falls for a Different Trap

Related concepts

Related terms