Author: Robert Hyatt
Date: 20:25:01 06/16/02
Go up one level in this thread
On June 16, 2002 at 17:04:47, Alessandro Damiani wrote: >On June 16, 2002 at 12:53:19, Gian-Carlo Pascutto wrote: > >>Hello all, >> >>my program has been seriously suffering from search instabilities >>since the day I started re-using my hashtables between searches >>(i.e. not clearing them). I dissected the problem into a pathological >>case this afternoon: >> >>1) We are pondering, do a deep 20 ply search and see that we have >>a combination to promote a pawn or three. We start failing high >>bigtime. >> >>2) The opponent plays a different move than expected, but because >>this is an endgame, it transposes into the variation we were pondering >>anyway, so we get a lot of hash hits. >> >>3) We start our normal search on ply 3 with a window [-50,50]. >>We get a hash hit one of the bottom nodes. It indicates a lowerbound >>of +2000. This is enough to get a cutoff, and we fail high. >> >>4) We restart our normal search on ply 3 with a window [50, inf]. >>We get a hash hit one of the bottom nodes, but this time the bound >>is _not_ enough to get a cutoff. We search this node, and (because >>we're only 3 ply deep), only see an advantage of 0.01. We fail low. >> > >I always research with (alpha, inf) when a fail-high happens. Therefore here I >would research with (-50, inf). This way I hopefully can avoid fail-lows. > >Regards, > >Alessandro It doesn't solve the problem. You get a "deep" hash entry match that says score is >= X, but when you re-search, you can't prove it with a shallow search. Either you get a fail low, or a score that is worse than the best move you have found so far, which will cause further mischief... > > >>5) We see a fail high followed by a fail low in the search driver, >>and need a way to resolve this. I do a full-window research in this >>case [-inf,inf]. Which gets a score of +0.01 back. >> >>6) We start our search at ply 4 (same thing all over again) >> >>7) This repeats until the search starts to see the 20 pawn >>advantage. We did 3 searches on each ply, two of them with an open >>window. >> >>Thank god I don't get this worst-case-scenario often, but I >>do see a lot of fail high/fail low combinations caused by this >>effect. >> >>I'm wondering if anyone knows a solution, or what other people >>that reuse hashtables do. >> >>-- >>GCP
This page took 0 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.