Author: Peter Fendrich
Date: 14:26:08 08/18/02
Go up one level in this thread
On August 16, 2002 at 23:20:17, Will Singleton wrote: >On August 16, 2002 at 16:37:19, Peter Fendrich wrote: > >>On August 16, 2002 at 15:12:36, Odd Gunnar Malin wrote: >> >>>On August 16, 2002 at 13:14:37, Will Singleton wrote: >>> >>>>On August 16, 2002 at 03:09:05, Odd Gunnar Malin wrote: >>>> >>>>> >>>>>I made a tool for automatic tuning of an engine but there have been zero >>>>>feedback on this so I guess most found this sort of tools useless. >>>>> >>>>>Odd Gunnar >>>> >>>>What tool was this? >>> >>>This was a tool that looped through a testset (epd-file) and adjusted evaluation >>>for each run. >>> >>>If you f.ex. had command in the engine like this: >>> >>>set bishoppair 40 >>>set rookonseventhrank 30 >>>set doublepawnpenelty -20 >>>etc. >>> >>>You tell the program with a input file like this: >>> >>>(Command|Min value|Max value|Step) >>> >>>set bishoppair|0|80|10 >>>set rookonseventhrank|0|70|10 >>>set doublepawnpenelty|-50|0|10 >>> >>> >>>It could loop throug all possible values for these settings and save the best >>>setting to a file. >>> >>>Ex. first run >>> >>>set bishoppair 0 >>>set rookonseventhrank 0 >>>set doublepawnpenelty -50 >>> >>>Next run would be with step = 10 for bishoppair >>> >>>set bishoppair 10 >>>set rookonseventhrank 0 >>>set doublepawnpenelty -50 >>> >>>etc. >>> >>>If you download the file >>> >>>http://home.online.no/~malin/sjakk/download/GradualTest.zip >>> >>>It should be in there, it is called adjusteval.exe >>> >>>The source file is here: >>> >>>http://home.online.no/~malin/sjakk/download/GradualTest_src.zip >>> >>> >>>The hard task is naturly to find enough position where the right move could be >>>found with pure evaluation, or at least with a searchdept of 2 - 3. >>> >>>I used Dann Corbit's quiet possition for this with first take away all moves >>>that needed some searchdepth to find. I used the gradualtest (also in the >>>packed) for this sou I could automatic the process. >>> >>>I also had another thing in my thought when I made it. To make personalities for >>>my engine. It could be easier to find position for this because you don't look >>>for the overall best move. >>> >>>Odd Gunnar >> >>I tried that but couldn't get it to work for some reason so I build the function >>inside Terra instead.... >> >>Peter > >Did either of you find a large enough test-suite to make this worthwhile? What >was your experience as to the benefits vs. hand-tuning? > >Will I used automatic tuning and hand tuning in an iterative manner. One problem is of course to find the best relative weights among the evaluation terms but even more tricky is to find the right terms to use and that can't be done by automatic tuning. I used the big file from Dann Corbit mentioned by Odd Gunnar. I tried to find out how sensitive each parameter (and groups of parameters) is for changes, like raising and lowering the value to see how much the result increases or decreases. It is always a risk for local optima so it can't be done fully automatic. At least from my experience. Peter
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.