tag:blogger.com,1999:blog-4459360717297142573.post8098432136741980024..comments2024-03-27T20:44:56.139+11:00Comments on chessexpress: Don't panicShaun Presshttp://www.blogger.com/profile/00897215011002594039noreply@blogger.comBlogger1125tag:blogger.com,1999:blog-4459360717297142573.post-55534426042954585202017-12-14T12:39:09.336+11:002017-12-14T12:39:09.336+11:00Actually, if you read the AlphaGoZero paper (the N...Actually, if you read the AlphaGoZero paper (the Nature one) carefully, and the current preprint, it seems there is no randomization (or even roll-outs!) in the MCTS they use (except in the learning stage, with "Dirichlet noise"). When playing a game, it's more of "book-building" according to the computer experts, as various GUIs allow. But I'm sure than the NN-based Google implementation is honed to their domain (80K new "MCTS" nodes per second).Anonymousnoreply@blogger.com