Author: Robert Hyatt
Date: 20:45:45 12/29/97
Go up one level in this thread
On December 29, 1997 at 22:41:57, Don Dailey wrote: >On December 29, 1997 at 17:41:52, Robert Hyatt wrote: > >>On December 29, 1997 at 16:03:23, Chris Whittington wrote: >> >>> >>>But isn't the real flaw with tests that they test for finding solutions, >>>but now how to get into those positions in the first place. And which >>>posiitons steered to is very subjective. Tal would steer different to >>>Tarrasch .... >>> >>>Chris Whittington >>> >> >>The gripe I have is with "finding" in general. IE I'd like to find >>some positions that are "positional"... where we all agree that a 4 ply >>search is enough to see "why". Many positional tests are really deep >>tactical tests. Others are positional, but they require deep searches >>to see the "stinger" at the end, so faster searches again help. I'd >>like to see a pure knowledge-based test suite where the "stinger" is >>not so deep that it takes a deep search to find it. And where "right" >>is =+1, wrong = 0, and the score at the end gives some measure of how >>strong the program is. SO far I've seen nothing like this... > >Indeed, this is a good idea. We have started working recently on >"compensation for a pawn" evaluation. We want our program to be able >to correctly sacrafice a pawn and I don't mean psuedo sac where a >really deep search only can see it. > >But one step in the right direction is a "problem set" with no >solutions, only evaluations. To score this problem set you completely >ignore the key move and look only at the score. I want to construct >this set for my work on evaluation. > >The idea would be to provide a score that the computer should try to >achieve and run the test for only a few seconds on each position. The >closer the computer gets to the correct score the better it does on >the problem set. You can easily fix it so that just jacking up all >your positional scores do not enable the computer to score better >by throwing in counter-examples and such. > >-- Don I have an oddball test mode that will take a test suite (like wac) and set each position and call Evaluate() and nothing more. I typically use this to debug things. IE when working on the blocked pawn code, I would simply add some printf's to the evaluation code to display the number of black and white pawns and the files they were on. I'd then run the WAC suite thru this odd mode and then by eye, look at every position and see what crafty thought was blocked. If I agreed, I went on. If not, I debugged and fixed it and ran it again (300 positions obviously takes a fraction of a second sinc ethere were *no* searches done... I use this for lots of things, in the mode you are describing, to make sure that when black is a pawn up with king exposed, the evaluation is what I think is correct...
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.