fn should_explore(exploration: f64) -> boolExpand description
Whether this draw should explore — ignore the weights and pick uniformly.
Weights are derived from observations, and observations only exist for paths that were selected. Left alone that is a closed loop: a path that falls out of favour stops being measured, so its score can never recover and it is never chosen again. Spending a small share of draws uniformly keeps every candidate under observation, which is what lets one that has recovered climb back.