Author: Tord Romstad
Date: 02:38:52 12/17/03
Go up one level in this thread
On December 16, 2003 at 21:22:56, Anthony Cozzie wrote: >Recently I experimented with adding MTD(F) into Zappa. It has been an >interesting experiment, but I am going back to PVS(). > >I thought that since Zappa has a [UL,LL] paired transposition table and an >evaluation granularity of only 1/100 of a pawn, I use 1/64 of a pawn in Gothmog. In the new engine I am writing, the granularity will be configurable by the user, with possible values between 1/8 and 1/512 of a pawn. >MTD(f) would work quite well, >but that does not seem to be the case. The MTD(f) version of Zappa does >slightly better on test suites (113/183 @ecmgcp v 106 @ 10s/move) but in the >positional test suites it averaged about 3/4 of a ply less than the PVS() >version. My guess is that because MTD(F) tries all moves, some of the >"ridiculously losing captures" ordered near the end by PVS() are tried earlier, >which accounts for the increased test suite performance. I don't understand this, I'm afraid. What does MTD(f) vs PVS have to do with move ordering? >If anyone has any suggestions, I'm keeping the MTD(F) code in Zappa (just turned >off) and I'm willing to try anything. > Some suggestions: 1. As Dan Andersson has pointed out, enhanced transposition cutoffs are very effective when combined with MTD(f). If you don't use them already, give them a try. 2. Do you hash the qsearch? You'll probably find that it helps you reduce the tree size quite a lot, even if it doesn't give you much in PVS. 3. Do you use the plain MTD(f) algorithm (as described in Plaat's paper), or do you use some kind of convergence accelerator? Plain MTD(f) never worked well for me. I now increase the step size when the search fails in the same direction more than once. Tord
This page took 0.09 seconds to execute
Last modified: Thu, 07 Jul 11 08:48:38 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.