Author: Robert Hyatt
Date: 13:40:04 09/01/04
Go up one level in this thread
On August 31, 2004 at 14:30:21, Chris Welty wrote: >On August 31, 2004 at 10:28:04, Robert Hyatt wrote: > >>On August 31, 2004 at 08:22:20, Chris Welty wrote: >> >>>On August 30, 2004 at 17:01:57, Robert Hyatt wrote: >>><snip> >>>>>Auto-tuning would be great. My problem is I don't know how to do it. If you >>>>>have a method I'd love to hear more. >>>>> >>>>>Dan H. >>>> >>>> >>>>It is also a _very_ hard problem. >>>> >>>>_very_ hard... >>> >>>It was quite easy to get Altamax to autotune its evaluation parameters so that >>>it played substantially better than the hand-tuned version. It took less time >>>than hand-tuning too. >>> >>>Getting autotuning to work WELL required some work, but was still easier than >>>hand-tuning the coefficients. >>> >>>I use TD-leaf, which I got from some paper on the internet. >>> >>>Chris >> >> >>How many different terms? Crafty has a bunch. auto-tuning them to be better >>than the hand-tuned default has not happened yet although we have (mainly >>Anthony) beat on it mercilously. Crafty has about 120 different eval features >>that can be adjusted, with maybe 1/2 of them represented as large arrays of >>values. Easily over 1000 different values to adjust. A tough task as I said... > >When I did the initial comparison, about 75 terms. Since autotuning is easy I've >added lookup tables; now there are a total of 570 numbers that can be adjusted. > >The piece-square table numbers do look strange (-0.95 pawns for a bishop on >D8??) but the results are good. That has to be the result of a poor test suite. IE with a score like that, the game will revolve around one program giving up a pawn to make the other program move its bishop to D8, which makes no sense in real chess... > >I suspect it also depends on how good you are at hand tuning.
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.