Author: Robert Hyatt
Date: 10:00:57 01/13/04
Go up one level in this thread
On January 12, 2004 at 19:15:48, scott farrell wrote: >On January 12, 2004 at 10:02:07, Robert Hyatt wrote: > >>On January 12, 2004 at 03:09:10, scott farrell wrote: >> >>>On January 11, 2004 at 21:18:30, macaroni wrote: >>> >>>>I have recently been fiddling arround with some passed pawn code, however I just >>>>can't seem to get it working very well. Can you give me an idea of what sort of >>>>values you give different types (in different stages) of passed pawns? >>>>cheers >>>>Tor >>> >>>I think it needs lots of knowledge. >>> >>>What I did is use several pieceSquareTables, for different types, like blocked, >>>isolated, defended, passed, connect passed (probably the most important), >>>defended passed, etc >>> >>>I then tweeked values to solve certain problems. With various psTables, you end >>>up with lots of levers to play with. >>> >>>Scott >> >>Watch out about "connected passers." >> >>give white two pawns on d2/e2. Give black two pawns on a2/h2. Both kings >>in the center. Which side do you like? :) >> > >mmm.. I thought, how often would that happen, then just watched chompster lose >an engine vs engine match to that very thing. > >I actually check for knight/bishop and score the connect passers close to >knight, as you often have to sac the piece to stop it. If only rooks left, score >it closer to a rook (piece tables still). > >I'll just add, if no pieces, just dont give bonus for connect passers. > >thanx for this > >Scott You need to be _very_ careful doing what you describe. It leads to a scoring discontinuity, or a "boundary condition" that can be very bad with minimax/alpha-beta search. If you know what a discontinuous function is, you just defined one (IE in Laplace transforms, you see "unit step functions" that have this property.) The problem here is that when you are searching positions where you _cross_ the boundary in the search, the program gets to pick when and where it crosses, and since you are talking about a positional score of +/- 3 pawns or more, what it might do is give up two pawns rather than lose the 3 pawn advantage. Not a good plan, usually. Your evaluation function _really_ should be a continuous function that has no such "discontinuities". IE you might do just fine where you have lots of pieces and use that trick, and once there are no pieces, you are also OK. But around the "edge" where you go from "have to have not" the search can do some amazing things to use that discontinuity to produce results wildly different from what you would normally expect to see. IE early in Crafty I did that for endgames, where material <= N. And at the transition point, I saw some _ugly_ ideas pop out of the search. Now it is a smooth transition from one phase to the next with no sudden large jumps.. I think Hans Berliner wrote a paper about this topic but I can't recall the title for the life of me... But, back to the idea, the reason for _many_ of my evaluation ideas came directly from watching games on ICC against Grandmaster and IM players. GMs are really amazing. I once spent a month tuning the value for the bishop pair. Roman played game after game and won way more than he should have, and he finally called one night and said "I have finally figured out what is wrong, your bishop pair score is too high." He could exploit that like crazy as Crafty would prefer to enter a blocked-up pawn structure rather than give up a bishop for a knight. And the resulting blocked-up position often made one of the bishops nearly useless. I tuned and he played and he finally said "that is playing reasonably..." and on we went to the next "topic". ICC is the best test-bed I could imagine, it is hard to believe so few actually use it... Inbred testing against a couple of computer opponents is nowhere near as valuable, IMHO. > >>This happens a lot vs GM players, once they see the hole. They will >>happily give you those two connected passers. :) >> >>Connected passers are good when you have pieces... But with no pieces, >>look out.
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.