Author: Robert Hyatt
Date: 19:00:54 10/27/98
Go up one level in this thread
On October 27, 1998 at 14:18:45, Roberto Waldteufel wrote: > >On October 27, 1998 at 08:44:14, Robert Hyatt wrote: > >>On October 26, 1998 at 16:09:42, Robert Hyatt wrote: >> >>>On October 26, 1998 at 12:43:26, Bruce Moreland wrote: >>> >>>>Here is the tail end of a game played between my program (Ferret) on a 533 mhz >>>>Alpha and Crafty on a dual 450 mhz Pentium II, at time control of 5 0, on ICC. >>>> >>>>My program got smoked in a K+P ending it traded into (apparently) voluntarily. >>>> >>>>One of the things you hear about if you talk to old-timers in this field is that >>>>the square of the pawn is a totally important evaluation term. I'm not sure who >>>>said it, but someone said that it's worth 100 Elo points. This was a long time >>>>ago, back when programs would get killed in endings routinely, but I get the >>>>idea that this wisdom hasn't been re-evaluated much since then. It is my >>>>opinion that square of the pawn is worth almost nothing in practical play, since >>>>the number of K+P endings that actually occur, with play still in them, seems to >>>>be approximately zero. >>>> >> >>I am probably the "old-timer" that provided this "quote" I suspect. :) >> >>Context: >> >>Back in the very early 1970's, we (bert gower and myself) were entering "blitz" >>in USCF chess tournaments. We had several at the University where we were >>working (University of Southern Mississippi). Our official USCF rating was in >>the 1400's and stayed pretty constant for a couple of years. >> >>one common way we were losing was passed pawns that could out-run the king and >>promote (we were typically doing 4-5 plies back then, which was *not* enough to >>see a pawn running from h2 to h8. After watching yet another loss where we had >>won a pawn (or a piece for a couple of pawns or whatever) the humans would do >>their best to simplify into endings where they could trade away the last piece >>and be left with a pawn our king couldn't catch, even though we were "winning" >>when you counted material only. >> >>Finally I decided to stop this nonsense, and added the simple "square of the >>pawn" code to blitz (added it in 1972 if I recall correctly). Our USCF rating >>immediately went up from 1410 to 1550, with (effectively) no other changes of >>any kind. The funny thing was this code was *always* a part of the blitz/cray >>blitz program, yet I only recall one game where it made a difference in an ACM >>event. You can probably find the 1984 ACM event in Los Angeles where Cray Blitz >>was black against NuChess. We were really lost (down two pawns and cramped >>beyond repair) when NuChess had to choose between winning a center pawn and a >>passed pawn on the a-file. It chose the center pawn because it really wrecked >>our remaining pawn structure, but it also allowed us to trade off all pieces >>instantly. I was sick at the time, with a fever, and was taking some sort of >>medication that just about had me knocked out. For about 15 moves I had been >>looking at -2, -2, -2 evaluation. We had already clinched first place after >>3 rounds, and if nuchess beat us they would have finished in clear 2nd. Mike >>was about to give the trophies out as it was about midnight or so and the >>spectators were tired also. Between moves, I happened to notice our terminal >>was saying +2, +2.5, +3, +4 etc.. I looked over to see what Gower had done, >>thinking he had entered the move wrong. After confirming our board matched the >>real board, my response was "holy shit, we are winning." That got Mike's >>attention and he came over to look. Sure enough, our a-pawn couldn't be stopped, >>and we won that game as well, that being the *only* computer-vs-computer game at >>an ACM event where this code made any difference. >> >>But one thing is for sure, back in 1972 the 100+ uscf rating improvement wasn't >>"guesswork" it was definitely "real"... all the games during that 1410-1560 >>improvement coming on the same xerox sigma 9 computer system... > >If your code detects that one side has an unstoppable passed pawn when the >opponent has only king and pawns left, how do you establish that the opponent >can't do something nasty like capture one of your pawns on the 2nd rank thereby >creating a more advanced passed pawn on the 6th rank, or even worse trapping and >mating your king with an onslaught of king and several pawns, aided perhaps by >some squares being blocked by your own pawns? I use an array of bitboards >indexed by side to move and king position (ie 128 bitmaps) which tells me which >squares would be "unstoppable" squares for pawns. However, I only use this to >determine trivial wins when one side has only a bare king, and when prior tests >have determined that I do not have mating material unless I promote a pawn. I >think the same quick "unstoppable passed pawn test" might be useful when the >losing side has some pawns as well, but I never figured out a satisfactory way >to tell whether the unstoppable passed pawn was enough to be sure of victory or >not in this more complicated situation. I would be interested to know how you >handled this. > Simple. I do a *deep* search, and only do the pawn race evaluations at the *tips* and not on interior nodes. I depend on the search to find the mates, the ways the opponent can create an even quicker-queening passer... and only after the normal 20+ plies of search are exhausted do I do the "race" test. >By the way, I don't think a 100 point rating improvement can come from something >that is only important very rarely. I bet that some of those lesser variations >that got cutoffs in the searches in endgames were only possible because of the >new code, allowing not only gross blunders to be avoided, but also more time >spent searching the more promising alternatives. It sounds like it may have been >one of the best single improvements you ever made. If only they were all worth >100 ELO points :-) You are mixing apples and oranges. In 1972 *the* way to beat a computer was to trade down to the ending, even a pawn or two down, then "swindle" the computer. I did this in the 80's against the chess challengers and so forth as well, as they couldn't search deeply enough either. Today, most games are decided well before a K+P ending because of the 12+ ply middlegame searches. In our case, it stopped that one plan. There were others of course... but at least we stopped trading down into a dead loss... when the loss was the result of a queening pawn. We would still trade down into an ending where our opponent had a majority that could convert into an outside passed pawn which still won... and I fixed that hole in Cray Blitz and Crafty... But there are still more to fix, of course. :) > >Best wishes, >Roberto
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.