TE
TechEcho
Home24h TopNewestBestAskShowJobs
GitHubTwitter
Home

TechEcho

A tech news platform built with Next.js, providing global tech news and discussions.

GitHubTwitter

Home

HomeNewestBestAskShowJobs

Resources

HackerNews APIOriginal HackerNewsNext.js

© 2025 TechEcho. All rights reserved.

Monte Carlo instead of Alpha-Beta for chess programs?

89 pointsby S4Mabout 6 years ago

3 comments

dragontamerabout 6 years ago
Monte-Carlo seems more akin to how humans play games.<p>Alpha-Beta pruning is the classical computer science algorithm, easier to understand, describe, and analyze... but very inhuman. Its an exhaustive search, like a breadth-first search.<p>Monte-Carlo has a lot of variations. Classically, the original MCTS algorithms for Go would play all the way to the end of a game before searching other parts of the tree. In effect, its a depth-first search, you play until you make a conclusion (potentially a &quot;bad&quot; conclusion). The reasoning is simple: on the average, a lot of &quot;samples&quot; of playthroughs will lead to a better estimate for win&#x2F;loss probability. Think of it as a random-sample of the win&#x2F;loss potential of a move.<p>EDIT: Cleaned up the above paragraph.<p>MCTS implemented by AlphaZero is different however. It uses a neural-net to find &quot;interesting&quot; positions to guide the search, and the same neural net also evaluates the position (who is winning or losing). It seems like a very good way to have a single neural network mostly perform double-duty (aside from the two output nodes). Most of the input layers &#x2F; early hidden-layers can be &quot;recycled&quot; between the &quot;explore&quot; function and the &quot;evaluation&quot; function.<p>EDIT: I don&#x27;t know how AlphaZero gets its &quot;exploration&quot; network trained up however. Its possible that the original randomized MCTS algorithm (search randomly... which btw, performs way better than you&#x27;d expect...) might be used to &quot;bootstrap&quot; the exploration weights.<p>So MCTS just naturally works with neural nets very well.
评论 #19561415 未加载
评论 #19564017 未加载
EventH-about 6 years ago
Why does this article emphasize Komodo MCTS and completely ignore LeelaZero which has had much more exciting and interesting results as far as &#x27;alternative method&#x27; engines go? This is especially strange given that Leela operates similarly to AlphaZero which the article strongly praises.
评论 #19561669 未加载
评论 #19561249 未加载
评论 #19562278 未加载
LukeWalshabout 6 years ago
I would love to use Alpha Zero. The article makes it sound like Komodo (which uses similar techniques to Alpha Zero) could also beat Stockfish (since Alpha Zero beat Stockfish). Stockfish still beat Komodo and all other engines according to the computer chess championship which uses equal-and-limited compute requirements.<p>The 5&#x2F;2 blitz computer chess championship is currently live (with both Stockfish and Komodo competing): <a href="https:&#x2F;&#x2F;www.chess.com&#x2F;computer-chess-championship" rel="nofollow">https:&#x2F;&#x2F;www.chess.com&#x2F;computer-chess-championship</a>
评论 #19561722 未加载
评论 #19561719 未加载