Author: Duncan Roberts
Date: 15:50:34 11/22/05
Go up one level in this thread
On November 22, 2005 at 15:05:47, A. Steen wrote: >On November 21, 2005 at 07:32:27, Tord Romstad wrote: > >>On November 21, 2005 at 03:17:15, A. Steen wrote: >> >>>On November 21, 2005 at 03:05:16, Tord Romstad wrote: >>> >>>:: everything else agreed with :: >>> >>>>In my >>>>opinion, a better way to use EGTBs would be to use them as an oracle >>>>when implementing endgame evaluation rules. We could analyse all >>>>the positions in an EGTB and try to find general rules and patterns >>>>which enable us to classify the majority of positions as won, drawn >>>>or lost without consulting the EGTB. >>> >>>I have spent years doing this. :( >>> >>>Except for the simple cases, where algorithms already produce good results and >>>make nearly optimal moves (without much shuffling), it is just too hard. >> >>It is not easy, but far from impossible. > > >Here and elsewhere, we are usually saying almost the same thing. > >The exception I highlight with "$$$$$$$$" signs. > > > >>It has been done successfully for centuries, even without computers and >>EGTBs. Open any advanced endgame textbook. You will find big classes >>of endgames classified as won or drawn, sometimes by extremely complex >>maneuvres which a chess program without EGTBs or special knowledge >>wouldn't find even after hours of search. For endgames which are too >>hard to analyse exactly, you will often find very effective heuristic >>rules for guessing whether a given position is won, and for finding the >>winning plan. > > >ot: Same older "advanced endgame textbooks" contain an enormously high >proportion of incorrect moves, too. :) And incorrect heuristics too. The >process often failed. > > > >>Consider KRPKR as an example. The complete EGTBs for this endgame >>are about 150 MB of data, not counting the numerous transitions to >>other endgames. This is obviously far too much to memorise for >>human players. Nevertheless, even moderately strong human players >>are able to play such endgames almost perfectly, because we have >>been able to identify the most fundamental classes of won and drawn >>positions, and to formulate some general high-level rules for how >>to play. That all this work was done before EGTBs were available, >>and with very few inaccuracies, is truly amazing. Doing similar >>work today would be much easier. > > >For KRPKR I agree. A P is a unidirectional linear man and which file it is on >is the first key. The rest is all about virtual rectangles formed by rooks and >about K-access. This underlies my krpkr heuristic anyway, and yes it is nearly >100% effective. I don't guarantee to find the EGTB best - usually my route is >longer- but I won't change the game outcome except rarely. > >But KRPKR is a very easy one for you to choose. :) >The heuristics happen to be simple enough to be able to be formulated in such a >way that a reasonably intelligent human can understand them... > > > >>When we have managed to replace the detailed, but unorganised >>EGTB data with general rules, it is often easy to evaluate much >>more complicated endgames correctly. As a trivial example, >>consider the following position: >> >>[D]8/6k1/1r5p/8/8/5K2/R5PP/8 w - >> >>For a human player who has studied basic endgame theory, it is >>rather obvious that this is a draw. It is clear that white >>cannot create a passed pawn without allowing a Philidor-type >>draw. Arriving at the same conclusion armed with just the >>raw EGTB data would probably require a very deep search. > > >Until the ~150Tb 7-man base set is computed, or at least enough of them to allow >krppkrp to be computed. > > >>>Maybe some neural network type of work might yield something. >> >>It would be interesting, but I think this is better done by >>humans assisted with a chess program and EGTBs than by an >>unassisted computer. >> >>John Nunn has done some fascinating work, but I still haven't >>found the time to read it. > > >$$$$$$$$$$$$$$ > >:) > >And here is the point. >The rules he found can't be explained in some simple way. >They are black magic and vague and incomplete (and he knows that). >To call them rules is in fact wrong of me. >The best description is "observations". > >If a (say) 1Gb tablebase set (original set + daughters after promotion(s)) is >replaced by an "observations" set which itself takes 1Mb (written in some >structured formulaic but human friendly way) to set out, have we moved forward? > >The human mind has limited capacity to understand, absorb let alone retain. > >Perfect chess quite likely will look only like disorganised chaos to even the >most supreme humans (my point in mentioning the monstrous ununderstandable >sequences we see). > > > >> >>>As an extreme case, what pattern can we hope to find in the useful parts of >>>KRRNKRR or KRBNKRB, in what appear to be just shufflings? >> >>I doubt that we could find any interesting patterns here, but these >>endgames have almost no practical importance. How often do you see >>KRBNKRB in real games? > > >I quoted them only because they are computed. But even in more commonly-seen >endings, with pawns, perfection play is chaotic. What rules there are, are >buried too far (I think) to ever be found. > > > >>Tord > > > >I summarise our difference AISI - you may be an optimist who believes that under >many/most chessic structures is a set of orderly rules (however complex, but >capable of comprehension by humans if they somehow became known). > >I am a medium-optimist but who thinks that under many/most chessic structures is >a set of orderly rules whose complexity renders them totally incapable of >comprehension by humans, even if they were "known". maybe in 10 years it will be possible for software to do some data mining and discover the rules. if that can be done then humans should be able to understand it. duncan > >Best, > >A.S.
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.