Author: Peter Fendrich
Date: 09:58:09 04/08/04
Go up one level in this thread
On April 07, 2004 at 01:22:53, rasjid chan wrote: >On April 06, 2004 at 17:01:12, Peter Fendrich wrote: > >>On April 06, 2004 at 16:34:27, Bo Persson wrote: >> >>>On April 06, 2004 at 11:12:50, Peter Fendrich wrote: >>> >>>> >>>>Agree, but most interesting is how often. It must be much more often in MTD. >>>>I just have presentiment of that the increased entry size wont pay off in PVS >>>>with that much more cut offs. >>>> >>> >>>But there is an additional value, besides the potential cutoffs. In the PV >>>search, the bounds might also help you reduce your window by moving alpha, or >>>beta, or both. Even if you don't get an immediate cutoff, you might get a pruned >>>tree. >>> >>>I once tried this by mistake, while moving from PVS to MTD(f). I started by >>>implementing the two bounds in my hash table, but forgot to add the modified >>>root search. It still reduced the search tree significantly, maybe 20%. >>> >>> >>>Bo Persson >> >>Fooling around with alpha/beta values is nasty business and would require >>thorough testing, at least in my program! >>Reduced tree size mustn't necessary be a good sign here... >>/Peter > >feedback for your info - I have now completed implementing ETC and >I think it is "well done" because it is not relatively tough. > >I don't test but use my eyes + log printout and it seems beneficial >and no hint that it is a drag on performance, but seems to search slightly >deeper... [sum of many slightlies] == better. Ok, but maybe a little test would useful anyway :-) >Now I'm introduced to hashing[LB, UB]. > >ETC make use of bounds info and MAYBE 2 bounds better than one. >There are 2 passes of looping thru the moves and we may have a cutoff in >the 1st pass or, as alpha improves in the 2nd pass,more moves may be pruned. >Added to the above, a raw guess is ETC + 2 bounds may be more beneficial >to MTDf. The average probability that a node fails high is much higher >in MTDf than in other algorithms... but PVS is also most of the time >null window. Well ETC increases the probe rate (I think...) and the usefulness of 2 bounds increases so maybe you're right. It would be interesting to test. I'll put on my already too long todo-list. /Peter >Rasjid
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.