Author: Dave Gomboc
Date: 17:17:29 05/19/99
Go up one level in this thread
On May 19, 1999 at 19:38:43, Mark Young wrote: >On May 19, 1999 at 14:25:38, Dave Gomboc wrote: > >>On May 19, 1999 at 08:05:28, Mark Young wrote: >> >>>On May 19, 1999 at 02:55:48, Dave Gomboc wrote: >>> >>>>On May 18, 1999 at 22:52:23, Mark Young wrote: >>>> >>>>>On May 18, 1999 at 22:18:47, Dave Gomboc wrote: >>>>> >>>>>>Some balance here is necessary. Some people are content to say that if a move >>>>>>is selected as best, this is good enough. Other people insist on seeing a score >>>>>>accompanying the move that accurately reflects the theoretical value of the >>>>>>position after the move is played. >>>>>> >>>>>>It is important that an accompanying score is accurate. This certifies that the >>>>>>software has seen its way through whatever complications may be present. >>>>>>Changes to the position that do not interfere with the themes that support the >>>>>>main line of play will not prevent a solution from being found. >>>>>> >>>>>>It is important that the best move is played, even if the accompanying score is >>>>>>not accurate. This certifies that the software understands enough about the >>>>>>position to make its way forward correctly -- for the moment. It is useful to >>>>>>identify that a move is likely to be the most promising, even if a definitive >>>>>>conclusion has not been reached. >>>>>> >>>>>>It is fair to say that the while the former is better, the latter is often >>>>>>adequate. If the software's assessment of a position is highly perturbed by >>>>>>"small" (read: irrelevant) changes to a position, then it could be said that >>>>>>much "luck" is involved in it choosing a particular move or another. More >>>>>>often, though, assessments do not vary widely between positions with "small" >>>>>>(again read: irrelevant) changes, and in this case, it is partially the >>>>>>consistency of the assessing that allows the proper move to frequently be >>>>>>chosen, even without the discovery of a tactical verification, or tactical >>>>>>refutation of alternatives. >>>>>> >>>>>>So, in testing, prefer a proper score, but do not ignore a proper best move >>>>>>without a proper score. It is still doing something right: give it half credit. >>>>> >>>>>What is the proper score, any + score, a huge +score....the only proper score of >>>>>a winning positions is mate in n. >>>>>> >>>>>>Dave >>>> >>>>This is a valid question. I think that a "proper score" is one that is high >>>>enough relative to the alternative scores that a certain path is absolutely >>>>guaranteed to be selected (including from perturbations of the position to be >>>>played from.) This is different from the usual notions of "seeing that it can >>>>win the piece", et cetera: a sufficient, but not necessary criteria for matching >>>>my "good enough" score. >>>> >>>>If the scores for all of the moves but the best one are abysmal, then it has >>>>found that only one move is playable, even if you don't see that it wins (or is >>>>actually not good enough anyway!, but is still the best move because it holds >>>>off a resignable position as long as possible.) If the scores for four or five >>>>moves are roughly even, this isn't really good enough to score "full points" >>>>IMO. >>>> >>>>We can see that if the program reaches a position that it presumed to be okay >>>>(in an earlier search), but sees that it is resignable when it gets there, it's >>>>toast. Finding a "proper" score at some earlier stage would have allowed it to >>>>avoid the error completely by not playing to the bad line. The major idea >>>>behind singular extensions is to see what is happening at the "end of the >>>>tunnel": you can see those fail-highs and fail-lows early enough that you can do >>>>something about them (head towards the position, or avoid it.) Programs that >>>>see "proper scores" from further back in the game are going to be stronger than >>>>those that don't, ceteris paribus. >>> >>>So what you are saying is the program has to see it tacticly, or it does not >>>count. Programs have other aspects to them then just there search, and I see >>>nothing wrong with a program finding the correct move by seeing the other line >>>get killed, instead of seeing this one move wins. >> >>This is something that I definately did not say. At least, I tried to be very >>careful not to. In some positions with some chess engines, even a steady >>one-third of a pawn difference might be enough to make such a guarantee. The >>margin of difference required depends on the position and on the engine. >> >>>You did not really answer the question, and we are back at square one. All of >>>what you are saying is subjective thus anyone can reject any BM played by any >>>program producing almost any score. You are trying to split the baby and I don't >>>think it can be done. >>> >>>You need to show me how this is workable with some examples. >> >>I am not about to enter a comprehensive scientific experiment to do so, sorry. > >Exactly, it not so cut and dry your solution. It's still better than a cut and dried, but incorrect, solution. If you like, use the marking criteria for Test Your Chess IQ on them. That is what I was trying to simplify. >> >>>>Dave >> >>Dave Dave
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.