Author: Robert Hyatt
Date: 19:12:59 02/27/06
Go up one level in this thread
On February 27, 2006 at 19:35:54, Tony Thomas Karippa wrote: >On February 27, 2006 at 14:53:26, Robert Hyatt wrote: > >>This will be relatively short and sweet. >> >>As most know, I've been doing some major revisions to Crafty, preparing for the >>next WCCC event. These revisions are affecting the evaluation code which has >>been really ripped asunder and partially restored to sanity, and the search >>which includes some new reduction code replacing the older futility pruning done >>near the frontier, where the reduction stuff is done almost everywhere. >> >>I knew there would be a few eval issues as king safety has mainly been >>simplified with a couple of glaring holes left open for later work. I was >>really interested in the new search code because the speed/depth looked very >>good. >> >>So I'll start there, briefly. I've tested the new code in lots of nunn-type >>matches, as well as in test suites. So far, the new search solves every >>tactical suite I have tried in less total time than previous versions, which was >>surprising since the late move reduction idea often delays tactical solutions by >>hiding some strange threat moves that get reduced and therefore look >>ineffective. But happily, it has gotten better tactically in the same time >>frame. Yes it might take an extra ply or two to find the key move, but it is >>getting those 1-2 extra plies done more quickly so that the key move is _still_ >>found faster than the older versions. So that looks good (so far). >> >>Another thing I watched for was for a sudden "fail low" to pop up unexpectedly >>in a game, and I didn't see a one. Yes we had fail lows, but they were >>progressive and were the result of bad positions getting worse, not being +1 and >>suddenly seeing -3 after our opponent made an unexpected move. So for the >>moment, the current search appears to be solid. I have spent a ton of time on >>it in past weeks running test matches against older versions, running test >>suites and going over the output carefully, etc. >> >>Now I'm back to finishing up the eval. >> >>If you look at the Rascal game (round 8 I think) Crafty played a really lousy >>move 28. Nh7+. After looking at this a bit, white has two choices. The knight >>must move (attacked by pawn) and it can retreat to f3 (sane) or check on h7 and >>get trapped (insane). Crafty chose the latter for lots of reasons, mostly >>wrong. The king has to move to the e-file, as the g8 square is attacked by a >>white bishop, and the f8 square where the king now stands is attacked by the >>knight. So Crafty is pushing the black king into the center of the board. The >>only problem is, queens are gone, and there is not a plethora of material >>available to attack the king in the center, and in fact, with the knight stuck >>at h7, nothing happens at all. This was just an evaluation error caused by >>recent changes to king safety that were considered "temporary at best". I had >>to reduce the king safety scores so that we could tune the other scores for >>pieces and not keep seeing unusual moves that were a result of big kingsafety >>swings, screwing up our ability to compare changes to see which was better. >> >>Another issue was that Mike/I did zero book preparation, which showed. On Sat >>and again on Sunday we played a horrible Sicilian line (I think against Fruit, >>then again against Glaurung) where we then played Bg7 taking the bishop out of >>play, and then castling that way as well resulting in a cramped position. >>Against Glaurung we actually broke the bind, but the bishop never got into the >>game still and we were eventually squeezed to death. I should have warned Mike >>to avoid Sicilian positions because the king safety is simply not ready yet and >>it particularly gets into wild things when both sides castle opposite, which >>didn't happen here. >> >>My next plan of attack is to take the evaluation and continue to simplify and >>clean things up, and then fill in the missing holes as necessary after watching >>lots of games. Crafty's endgame skills are slowly returning as other eval terms >>are fixed and no longer swamp the important terms in endgames. I doubt any >>parallel code will be changed unless I end up on some sort of hardware with an >>unexpected "issue" that needs addressing, so until May, the Eval is going to be >>the focus for the group of folks helping me work on this stuff... >> >>I was quite happy to not play "certain opponents" and get drowned with incessant >>banter about nonsense, so that was a welcome change. :) Although we had some >>of that "banter" on channel 64 frequently, but I just generally tuned channel 64 >>out and enjoyed watching a couple of games along with the Crafty game... >> >>I believe this new version, when done, will end up being significantly stronger >>than anything released in the past from the Crafty series. How it will compare >>to the "front runners" will be seen in a few months, although it will >>occasionally play on ICC as well. I will add that I have a "random rotation" >>set up on ICC so that the most recent version does not play all the time to >>avoid any tuning issues that might come up. :) I have several older versions >>that now "claim" to be version 20.3, to make it more difficult to tune against >>prior to the WCCC. Once the WCCC arrives, the current version will again become >>public and I'll start to work on ideas for next year... >That's one of the thing I like about you, you never give up, always thinks >positve and continues to improve crafty. Many who started programming the same >time as you have already quit, or hadnt made any progress in many years. Good >luck in WCCC Professor. >Tony "many who started.."??? My first program played its first move in 1968 on an IBM /360 model 40. Know anybody that has actually been working on a chess program that long? Much one that is still doing the grind? :) two more years and it will be 40 years ago... :)
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.