Author: Vincent Diepeveen
Date: 05:00:40 05/13/98
Go up one level in this thread
On May 12, 1998 at 16:54:32, Robert Hyatt wrote: >On May 12, 1998 at 14:43:50, Vincent Diepeveen wrote: > >> >>The new version in fact has 3.5 out of 4 against crafty >> >>2.5 out of 3 against Moron: >> 1: Crafty v15.5 (2 cpus) >> 2: Dual 333/PII Linux(2.1.101) >> >>1 out of 1 against >>1: Crafty v15.6 (4 cpus) >> 2: computer operators: limit consecutive games to 4 or less or risk >>getting +noplay'ed. I get too many complaints from IM/GM players about >>getting locked out by another computer playing Crafty... Thanks... >> 3: ALR Quad-6 P6/200 X 4, 512mb RAM, Linux 2.1.99 >> >>Here the history of DoctorWho: >>19: = 2416 B 2461 Moron [ sr 15 15] C55 Rep May 12 98 02:29 >>18: + 2414 W 2501 crafty [ sr 30 30] C45 Res May 11 98 05:23 >>16: + 2392 B 2463 Moron [ sr 15 15] B12 Mat May 10 98 01:02 >>15: + 2371 W 2484 Moron [ sr 15 15] C30 Mat May 09 98 23:37 >> >>So to answer your question Bob: 3.5 out of 4 is almost 90% against >>at least 3 times faster hardware. >> >>Vincent > > >doesn't work like that. I can pick a series of games from the week >before where I won *7* and lost *0*. so you can't cherry-pick one good >result. We're talking about results of the latest version of Diep. Not about a set of games somewhere in between. I picked them because they were in the history of Doctorwho. DiepX is bad example, because i'm running this account, and last month or so i hardly ran it up. I just ran it up at a Sunday to test some code at blitz. Lost nearly all games bye the way. To get a fair comparision i think we should take some things into account a) some things Diep still lacks where crafty takes adantage of b) gameresult ain't always the reality of how a program performs. The advantage of internet is that you get quickly masses of games within few days. Sometimes even within 24 hours. Unlike Schroeder, Chessbase and others i don't have bunches of auto232 players here, which produce a thousand of games every 2 weeks. In the past Jan Louwman sometimes played few games for me, but that usually takes a week for say 50 games of 60 moves in 60 minutes, and recently Jan got very sick. Not looking good. So from auto232 player i know how my program performance is, but Jan never played 3 minutes a move, and seldom blitz. Big surprise was that Diep played so well in blitz in Paris, beating first round Kallisto, which directly withdrew from the blitzevent (so diep didn't get this point and needed to play another game against a different program to get that point) then crafty was easy to win from. 2 endings Diep drew which were totally won, but yeah, Diep is horrible in endings with few pawns. Against Fritz diep drew an ending with a pawn up. I was happy in fact that Matthias offered me a draw. Objectively he was totally correct, fritz never could have won that endgame. Against Ferret Diep lost after it looked dangerous for ferret, but in the post mortem it appeared that Ferret played very well. Against Nimzo diep lost on book. Out of book nimzo was +3.x; would have lost anyway. Nimzo is great in blitz. Against Junior diep had a totally won position, but then diep searched 5 seconds for a move, played it and lost. at second 6 it would have gotten a fail low. 6 ply ain't enough to see tactically everything. But in that position 6 ply would have been enough to win it. Against Darkthough Diep won on clock, after darkthought played on after 2 drawoffers(english+german) of me in a rookending with a pawn up for Diep. That was the only lucky endgamepoint of Diep. Enfin, Diep lost bunches of points there in the endgame and only got 0.5 point for free. finally diep finished 7th. Big surprise of a program that's about 20 times slower than Fritz. This was an important blitz event. Diep was there. Now after this diep started playing a little more (before that handoperated) on the internet. blitzrating: 2300. Date: november 1997 All these games since then i started to think about the fact why a program like crafty at home does so bad, and at the internet does so well. Fact is that all games played with crafty over here, crafty performance is horrible. Same report i get from others who played with crafty against Diep. Usually 1 draw and that's it. What's the big difference between playing at home/tournaments and internet? Well a lot: Here crafty has hardly tablebases, just 28 mainly 4 piece and 3 piece. No single 5 piece. crafty cannot take advantage of lag. Sounds weird to you? Yes it might, but fact is that Diep is so stupid to clean ALL its hash every move. Next version this feature will be left out. In my draughtsprogram i don't do this, and after some more measurements i conclude now that this is for internet play terrible important. now at home i play levels like x moves in x minutes. At the internet this level ain't possible. Just blitz. It seems my time division still sucks. All games at the internet are blitz. at home i play more like 2 hours 40 moves, but then both programs sharing at 1 computer (winboard). Recently it is possible to match ferret and crafty at slower time controls, for which i thank them both. Before this slowest level against crafty was 15 15. Book. At the internet for ages diep played with the same small book. I tested some different openingsmoves. It appeared that 1.c4 doesn't work that well for Diep. too bad that i was too lazy for several months. Last 2 weeks it's possible now to read in big pgn books into diep. Before that there was only 1.c4, and then diep was out of book within 3 moves against crafty at the internet. Good book job Hyatt, no matter whether it was caused by learning, or by hand editting. With the introduction of Wall book for crafty recently things changed, lucky last weeks diep ain't out of book within few moves anymore. Diep has no booklearning. That's stupid of course, but a fact. I've investigated learning a lot, and know exactly how to do it now. I just need that Saturday that i have time and that i want to make it. Endgame. after Paris diep still plays the same sucking endgame. I don't mind this, but if we'd adjucate positions, then i think these 82 losses are not entirely correct. A big problem in Diep is that it doesn't have KRP KR, where Diep does like to exchange to rook endings very much. Bunches of games diep has lost because of this. Can't count them so much. This is nice for Hyatt of course, if he wants to present statistics. Time. Diep has time trouble. Sometimes it loses on clock. I'm not sure why. Probably bugs. Draw offers. Diep offers sometimes at move 1 already draw. I'm not sure why. Problem is that on ICC you need to CLAIM a draw when it's perpetual. diep didn't claim. Result is that it loses on flag. When i tried to implement this went wrong. Still is wrong. Perpetual is an excellent way to win from diep. It stops searching usually, but it won't move anymore. DiepX is hard to measure lately, because i'm the only one running it, and i must pay a lot for internet access. .75 dollar cents an hour during evening, and 3.5 dollar an hour during daytime. The only accounts we can measure are: JudgeTurpin and DoctorWho. I think 2 accounts are enough. No questions to run up another account, i always say no to this, unless you are in for a new experiment. >Here's how things *really* look, and I am excluding results from the >last >4 weeks where I had the SMP computer... This is strictly crafty vs DiepX >and crafty vs DoctorWho... on equal hardware: > >Crafty has won 82 against diepX, lost 29, and drawn 24. *total* >against handle "DiepX" > >Against DoctorWho, Crafty has won 16, lost 2 and drawn 3 since he >started >using diepx... >Now, to resolve this, I suggest we take this to ICC and play a match, >any time control you want... > >Then we can decide who's ideas work, when we have some numbers to look >at. But my numbers are *nowhere* your cherry-picked numbers... > >And none of the above include the ALR hardware... it would be *much* >worse >in that case, as I only see two losses total in the ALR data... and >they >have played plenty of games... > >And don't go into the "bug" discussion... I have had *serious* bugs too, >just fixed a bad one that screwed totally up the repetition code... note that last couple of years crafty's evaluation hardly has changed, where Diep's evaluation i experiment with every week. Lately for example i changed bunches of knight code. especially the mobility part of it. Caused it to prefer a knight always over a bishop.
This page took 0.01 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.