|
||
Title: Time spent on search vs. evaluation Post by n00bftw on May 20th, 2014, 12:45pm I am curious as to the current tradeoffs between time spent searching vs. evaluation. For example, if a single-threaded bot spends two minutes doing a typical alpha-beta search, how much of that time is spent evaluating moves and/or board positions, and how much of that time is spent generating steps/moves to search? I realize that the answer will vary by bot and by board position. I am wondering, though, if there are general ranges which are considered typical. Thanks! |
||
Title: Re: Time spent on search vs. evaluation Post by leon_messerschmidt on May 25th, 2014, 1:16am For my (unfinished) bot it is about 2/3 searching and 1/3 evaluating at the moment. |
||
Title: Re: Time spent on search vs. evaluation Post by ddyer on Oct 3rd, 2014, 1:53pm For traditional alpha-beta searches, improvement in the quality of the evaluator trumps everything else. Ideally, you would work on a playing algorithm without consideration of the search time, until you achieved the desired quality of play, and then optimize what works. In the real world, alpha-beta has very lumpy performance; you pay a huge time penalty to increment the search depth, so you'll be forced to select a depth and program to keep the performance at that depth acceptable. This leads to a lot of time wasted optimizing things that don't really matter. Another point to consider is that evaluator quality is ultimately an emergent property; If you could see all the way to the end, you'd need only a trivial evaluator that reports win or loss. By "quality of the evaluator" you could think of approximating the emergent properties that would be apparent in a deeper search, if you could afford do one. |
||
Arimaa Forum » Powered by YaBB 1 Gold - SP 1.3.1! YaBB © 2000-2003. All Rights Reserved. |