Definition
Plain language
A simple way of deciding which option to try next, balancing what has worked so far against what you haven't tested enough to judge.
As stated in the literature
An arm-selection policy for a multi-armed bandit that trades off exploitation of high-estimated-reward arms against exploration of under-sampled ones; here used to choose among candidate edit strategies.
Also called: bandit rules
Why it matters: Without a principled way to split effort between proven and untested options, a search either locks onto the first decent idea or wastes its budget sampling everything equally.
For example, a cafeteria diner who has loved the soup three times but never tried the curry uses a bandit rule when deciding whether to order the safe favorite or finally sample the unknown dish.
Heard on the show
“A bandit rule decides which one to try next, weighing what’s paid off, against what hasn’t been tried enough to judge.”Episode 276 — An AI Agent Rewrote Its Own Scaffolding For Eight Days. Here's What Survived