Thursday, December 26, 2013

[ivohjduo] Formalizing Bayesian evaluation function

We can theoretically formalize the idea of an evaluation function (e.g., for a chess engine) which emits an probabilistic expected outcome.  The subset of the universe of all positions which have the same evaluation score (or features) have a distribution of absolute truth of being a winning, losing, or drawn position corresponding to the expected outcome. An equivalence class within which the evaluation function cannot distinguish, whose scores have gotten merged.

No comments :