Author: Vincent Diepeveen
Date: 05:48:54 09/13/01
Go up one level in this thread
On September 12, 2001 at 22:51:48, Robert Hyatt wrote: >On September 12, 2001 at 16:46:35, Vincent Diepeveen wrote: > >>On September 12, 2001 at 09:55:11, Robert Hyatt wrote: >> >>>On September 12, 2001 at 09:49:09, Vincent Diepeveen wrote: >>> >>>>On September 12, 2001 at 00:19:33, Robert Hyatt wrote: >>>> >>>>>On September 11, 2001 at 22:46:43, Vincent Diepeveen wrote: >>>>> >>>>>>On September 11, 2001 at 12:32:27, Uri Blass wrote: >>>>>> >>>>>>>On September 11, 2001 at 10:43:01, Robert Hyatt wrote: >>>>>>> >>>>>>>>On September 11, 2001 at 10:27:40, Uri Blass wrote: >>>>>>>> >>>>>>>>>On September 11, 2001 at 10:19:31, Robert Hyatt wrote: >>>>>>>>> >>>>>>>>>>On September 11, 2001 at 08:51:20, K. Burcham wrote: >>>>>>>>>> >>>>>>>>>>> >>>>>>>>>>> >>>>>>>>>>> >>>>>>>>>>> >>>>>>>>>>>i understand your explanation for the Rg8 and the Rf5 moves bruce. >>>>>>>>>>>that deep blue might have seen a loss in both of those lines. >>>>>>>>>>> >>>>>>>>>>> >>>>>>>>>>>i only use the word blunder when during normal game mode or in >>>>>>>>>>> analysis mode the score will jump maybe 2 or more points, >>>>>>>>>>>when the next move is made. >>>>>>>>>>> >>>>>>>>>>>for example in this deep blue blunder SOS scores this position >>>>>>>>>>> black is down -1.65. at depth 15. you can see in the analysis that >>>>>>>>>>>the score imidiately jumps and climbs to +6.41 for white with 44. ...Rd1. >>>>>>>>>>> >>>>>>>>>>>in the case of the deep junior vs shredder, in the world championship >>>>>>>>>>> i have analyized the 5+ change in score. this was not a single >>>>>>>>>>> move blunder like defined above. in the deep junior game >>>>>>>>>>> shredder didnt have a clue of the deep pawn value and its ability >>>>>>>>>>> to stop them. then when it finally saw what was really going on >>>>>>>>>>>shredder started adjusting its eval very quickly, and the score jumped >>>>>>>>>>> 5+ points. >>>>>>>>>>> >>>>>>>>>>>and i am aware that you already knew all of this----i was just explaining >>>>>>>>>>> my logic for my applicaton of the word blunder. >>>>>>>>>> >>>>>>>>>>this just means that SOS doesn't understand the position yet. When I ran >>>>>>>>>>this, I got +3.5 or so. On Rd1 my score gets significantly worse. Which >>>>>>>>>>simply means that they probably searched the alternatives deeper than I did >>>>>>>>>>and found that they were bad also. >>>>>>>>> >>>>>>>>>I rememeber that they admitted that Rd1 was result of a bug. >>>>>>>>>Their score for Rd1(-1.80) does not make sense >>>>>>>>>in every reasonable depth >>>>>>>>> >>>>>>>>>They did not play Rd1 because they found >>>>>>>>>that the alternative is worse. >>>>>>>>> >>>>>>>>>Uri >>>>>>>> >>>>>>>> >>>>>>>>I don't believe -180 is the "score" I think it is an indication of a fail >>>>>>>>high. They didn't resolve a fail high unless a second fail-high occurred, >>>>>>>>since knowing that A is better than B is enough to play A. If you know >>>>>>>>that A and B are both better than C, then you have to re-search A and B to >>>>>>>>find out which is the better move. I believe their bug was in the code that >>>>>>>>handled this when a time-out occurred. >>>>>>> >>>>>>>-260 was the score for Rf5 based on their output and it means that the score >>>>>>>for Rd1 was more optimistic for black. >>>>>>> >>>>>>>It seems clear to me that the stupid mistake was result of a bug. >>>>>>> >>>>>>>I guess that the bug happens only after failing low and not being able to solve >>>>>>>the fail low or to finish the iteration on time. >>>>>>> >>>>>>>Uri >>>>>> >>>>>>In diep i only play a move that failed high after research is finished. >>>>>>if time gets out then i do not play the move failing high at this moment >>>>>>i play th ealternative which was searched better. >>>>>> >>>>>>Still many programs however to today would play the failed high move. >>>>>> >>>>>>As bob indicates this looks easy case to me without much discussions. >>>>> >>>>> >>>>>If you fail high, but can't resolve the fail high before running out of time, >>>>>I don't see any problem whatsoever with playing the fail-high move. However, >>>>>if you fail low, and resolve that, then fail high on two moves without resolving >>>>>either, then playing one of them is _very_ risky... >>>> >>>>It is very risky to play a move that fails high. I have all kind of >>>>stupid extensions like SE, some threats, checks. All is based upon alfa and >>>>beta values, so when i research a mainline diep sees tactical way deeper >>>>than when it gets the fail high. Hence the 2 searches are not based upon >>>>the same lines, in short that means that a fail high can't be reliable. >>>> >>>>Suppose next horror scenario: >>>> >>>> program searches very long onto a move becaus eof whatever reason >>>> (fail low or whatever delay with the opponent). Then suddenly you search >>>> real deeply. >>>> >>>>The opponent makes a non expected move. Program gets a fail high for >>>>a nonsense move, doesn't have time to resolve and plays that nonsense move. >>>> >>>>Please test some games you lost with crafty and try to figure out how many >>>>moves that failed high in the end didn't become a new PV, just the *risk* >>>>of it which was no problem in the past, is just too much to take in nowadays >>>>computerchess where every move must be from high quality. There is no space >>>>for worst cases to happen. That's why i don't play a move when it fails high, >>>>only after it has been researched. NO risks! >>> >>> >>>I have done this extensively. I've not found any cases where a "phony >>>fail-high" occurred, other than on the PVS searches (null-window). I don't >>>accept fail highs there without a valid research. But on the root alpha/beta >>>window, I do accept a fail high and have never found a case where it was wrong. >>>If I did, I would consider that a bug and fix it. >> >>For diep results were different. >> >>Note that we do things different in the root. I am always searching >>the first move with [-oo,oo] and the rest with alfa,alfa+1 in the root. >> >>Best regards, >>Vincent > > >Ok.. that is why I search like I do. With an aspiration window. If I fail >high on alpha,alpha+1, I _always_ research and don't accept the fail high >unless it fails high on alpha,beta re-search as well. However, if it fails >high there, I bump beta to +infinity and research again, and I don't have to >get a score here to play the fail high move. yes i know, so the initial research we both wait. the aspiration window works great for testsets but in practical game play i have had too many worst case to keep using it.
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.