25,161
edits
Changes
no edit summary
However, as confirmed by [[Mark Lefler]] <ref>[http://www.talkchess.com/forum3/viewtopic.php?f=2&t=67441&start=25 Re: Komodo 12 and MCTS] by [[Mark Lefler]], [[CCC]], May 14, 2018</ref>
implementing MCTS was already an old idea of the Komodo team - also due to [[Don Dailey|Don Dailey's]] computer [[Go]] experience building an [[UCT]] player.
<span id="MCTS"></span>==Komodo MCTS==In Komodo's [[Monte-Carlo Tree Search|MCTS ]] mode the [[Search Tree|search tree]] is expanded in [[Best-First|best-first]] manner based on [[Pawn Advantage, Win Percentage, and Elo|winning probabilities]] determined not neither by random playouts nor by a [[Neural Networks|neural network]], but a tiny [[Alpha-Beta|alpha-beta search]] plus [[Quiescence Search|quiescence]] and [[Evaluation|static evaluation]], also similar to [[UCT]], dealing with the crucial trade-off between [https://en.wikipedia.org/wiki/Monte_Carlo_tree_search#Exploration_and_exploitation exploration and exploitation]. While [[Playing Strength|playing strength]] is lower with MCTS, positional play and judgement may well be better in many positions <ref>[http://www.talkchess.com/forum3/viewtopic.php?f=2&t=67441 Komodo 12 and MCTS] by [[Larry Kaufman]], [[CCC]], May 14, 2018</ref>, not to mention a more risky and entertaining playing style.
=Tournament Play=
