Author: Tord Romstad
Date: 10:58:52 08/17/03
Go up one level in this thread
On August 16, 2003 at 17:43:09, Uri Blass wrote: >On August 16, 2003 at 17:17:26, Tord Romstad wrote: > >>The last few weeks I have been experimenting with MTD(f). It still does not >>work quite as well >>as my old aspiration alpha beta search, but I intend to keep working on it for >>some more time. >> >>One of my biggest worries right now is that the search slows down almost to a >>full stop when >>the hash table starts getting full. For instance, here are the time (in >>seconds) needed to >>finish a 11-ply search from the starting position for different hash table sizes >>from 5 to 320 MB: >> >>5MB: 102 seconds >>10MB: 88 seconds >>20MB: 78 seconds >>40MB: 55 seconds >>80MB: 47 seconds >>160MB: 34 seconds >>320MB: 27 seconds >> >>Is this normal, or do I have some kind of bug? I really hope I don't need a >>320MB hash table >>to play decently at slow time controls. Should I take a look at my hash table >>replacement >>scheme? Currently I always replace if the depth of the new position is >= dept >>of old >>position. >> >>Another problem is that MTD(f) has very weird effects when combined with the >>various forward >>pruning techniques I use. This is not entirely unexpected, because I use the >>values of alpha >>and beta for pruning decisions. When I replace plain MTD(f) (which I haven't >>been able to >>make work very well; too many researches) with MTD(f) with a convergence >>accelerator, I >>often get entirely different search results for the same position. If I remove >>all selectivity >>(except null move pruning) from my search, changing the test driver does not >>have any >>effect on the search results, but then my program becomes much too slow. I >>suppose I >>will have to work out new forward pruning techniques. >> >>Tord > >Why? >Maybe your forward pruning is superior to MTD. > >The fact that MTD is not good for you because of your forward pruning does not >mean that your forward pruning is bad. I don't really like my forward pruning. I am beginning to become sceptical to all pruning schemes which involve the values of alpha or beta. Even in plain aspiration search there are problems when such pruning is combined with hash tables. When you encounter a previously examined position somewhere in your search and look up the position in your hash table, the information there may have been calculated with different values of alpha and beta. Unfortunately my program is so slow that I have to do some sort of forward pruning to avoid falling apart tactically. But I am sure there are many better ways to do it than my current approach. Tord
This page took 0.01 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.