Twenty Questions

Think of an object. I'll find it by asking whichever question shreds the most uncertainty.

—
—
—
—

posterior entropy 1 bit/question noise-limited best surprise

Question ranking expected bits gained

    A guess is just a question about a single object, so it is scored by the same formula and competes in the same list.

    Object–feature graph

    Posterior top 12 of 104

      Asked so far actual bits gained

        Let it play itself

        The policy as a decision tree

        The reasoner never stores a tree — it recomputes the best question from the posterior every turn. But that computation is deterministic, so the whole policy is a tree: the first question is fixed, and each answer determines the next one down to a guess.