Author: Ulrich Tuerke
Date: 06:40:42 06/30/98
Go up one level in this thread
On June 30, 1998 at 09:24:00, Inmann Werner wrote: >Selective deepening and Hashtables > >Problem: >In alpha.beta , there is a variable distance (distance to the horizon) , which >is decreased each call of >Alpha-beta. >If you have, lets say for example, a check, you want to deepen the search. I do >this by increasing >The variable distance. > >If you insert the position in the hashtable, you write the variable distance in >the hash table. >(I do so). >I only take hashtable entrys, where the distance in it is more than the actual >distance. (why not?) > >If a search the move one ply deeper the next time (iterative deepening) and >look at the hash table, the Position has a very good value of the variable >distance because of the selective deepenings made in the Last search. Because >the program knows nothing about the actual selective deepenings, which will come >In the search it is satisfied and takes the hash entry. > >And thats the problem. The search nearly never looks really deeper in this move, >because it takes it always fFrom the hash table until my search is deeper then >all previous extensions. > >Is my problem-analysis ok? Have I forgot something important or stupid? I think that your analysis is okay. I have worried some time ago about the same problem. I don't think that this ever had been discussed here or elsewhere. I'm afraid that there is no clean solution to this (at least none without a performance loss). In past, I have tried various attempts to be as clean as possible, e.g. stored only the depth not accounting for extensions but when reading compared the stored depth to the current one including extensions. This means, in any case of doubt whether the stored depth was sufficient, the hash table entry wasn't used. The consequence was a rather poor performance concerning hash table use. Now, I'm just ignoring the problem. If your search is speeded up a little bit artificially by the hash table probe, you have at least the benefit to proceed faster to higher iterations. In particular, the main variation root move will be thru immediately sometimes. May be there are "colleagues" with satisfying solutions ? Uli > >If not, what can I do? > >Should I write the "distance - number of extensions" when I write to the Hash >table? What follows then out of this!? > > >Werner
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.