Author: Dann Corbit
Date: 10:33:01 09/11/02
Go up one level in this thread
On September 11, 2002 at 05:19:27, Andrew Williams wrote: >On September 10, 2002 at 20:17:21, Gian-Carlo Pascutto wrote: > >>On September 10, 2002 at 19:52:32, Dann Corbit wrote: >> >>>If MTD(f) were linear [or linear behavior happened more than once in a blue >>>moon], then PostModernist would not score nearly so well, I think. >> >>I'm fairly sure Andrew uses convergence accelerators, so it's not >>strictly MTD(n,f) in PostModernist. >> > >Yes. I do 1, 1, 2, 3, 4, 5, 6, 7, 7, 7... I've tried all sorts in my time: > > 1, 2, 4, 9, 16, 25, 36, 36, 36... > 1, 1, 2, 3, 5, 8, 13, 21, 34, 55, 55, 55... > 1, 2, 3, 4, 4, 4... > >If I get FH, FH, FH, FH, FL, I start heading back down subtracting from the >first numbner. > >All of these work better for me than just using 1 for every step. Have you tried a "tiny" NegaC* centered on f? I would be interested to see how a binary search with a very small window fares. The reason I mention it is that I suspect your implementation of MTD(f) framework is as good or better than anyone else's. The rest of us just play with toys. For instance, a 256 unit wide interval will take 8 searches. If you know the average value for the number of searches applied, you could try a window one or two bits narrower and see how it comes out. Just a thought.
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.