Author: Uri Blass
Date: 06:51:33 03/02/04
Go up one level in this thread
On March 02, 2004 at 09:04:13, Peter Fendrich wrote: >On March 02, 2004 at 08:07:57, Sergei S. Markoff wrote: > >>Hello all! >> >>I post an answer to Cristophe's posting, but it was too close to message list's >>end. >> >>>An idea I have played with, and which worked somehow in the 16 bits version of >>>Chess Tiger is to: >> >>I'm using in SmarThink a scheme like you described but with some differences. >>I'm searching some first moves, captures and extended moves to full depth. Other >>moves I'm searching with reduced depth (the reducing amount depends of several >>factors, for example if the opponent king safety is worse and it's a move that >>going piece closer to opponent's king then reducing amount will be less e.t.c., >>I'm using here a lot of knowledge including trajectories analysis e.t.c.). If >>the reduced search returns value > alpha, I'm researching this move to full >>depth. There are a lot of possibilities to include a lot of knowledge here to >>build really knowledge-guided tree. > >Hi Sergei, > >I've tried similar things in Terra with no success at all and the key here is, >as you point out, knowledge. My program is currently not structured for carrying >a lot of knowledge in the interior nodes and that's of course why these things >didn't work so well. I have only kept a few features. For instance reducing the >depth for moves down in the move list that never (seldom) caused Fail High. >(Kind of "reversed" History killer table). It seems to improve the program a >tiny bit but not much... > >/Peter This is the only similiar pruning that I have in movei for more than a year(even 0.0799 is using that pruning). I believe that this is the pruning that sergei talked about in the past when he talked about history based pruning(I implemented it independetly in movei before he talked about it and was disappointed when he talked about it because I understood that it means losing one relative advantage that movei has relative to most programs and as an inferior programmer without some relative advantages thanks to ideas I have no chance to compete). I believe that the main problem is to write the relevant functions to calculate the relevant knowledge. My experience with history based pruning was that it clearly helped in test suites(I was less sure about games but I decided to keep it and later made some exceptions when not to use history based pruning). Uri
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.