Author: Robert Hyatt
Date: 20:08:42 10/20/00
Go up one level in this thread
On October 20, 2000 at 14:34:25, Christophe Theron wrote: >On October 20, 2000 at 13:03:49, Albert Silver wrote: > >>On October 20, 2000 at 02:02:56, Christophe Theron wrote: >> >>>On October 20, 2000 at 00:54:13, Albert Silver wrote: >>> >>>>On October 19, 2000 at 15:28:50, Ratko V Tomic wrote: >>>> >>>>>>I disagree. >>>>> >>>>>I think we're weighing differently what we consider typical position, each has >>>>>in mind different sample. As you point out, different causes (diminishing >>>>>returns, evaluation errors, position symmetry, quiet vs active positions, etc) >>>>>act in opposite direction in how they amplify or attenuate the 1 ply difference. >>>>> >>>>>Although I agreee that current style programs may be nearing the diminishing >>>>>returns region in deepening the brute-force search, I was talking about some >>>>>future more advanced algorithms which will push that region farther. I.e. the >>>>>point I was making is that the more advanced (far sighteed, not in the brute >>>>>force sense) the program, the greater difference it will see from 1 ply >>>>>difference. >>>> >>>>Perhaps the problem is this: you believe the nature of the diminishing returns >>>>of further depths (1 ply) is inherent to the programs' ignorance. In other >>>>words, the reason its evaluation doesn't show a greater difference from that >>>>extra ply is because the program is ignorant. I believe otherwise. I believe it >>>>is inherent to the game of chess. I believe there is no forced win to chess, and >>>>that in fact there are many times many paths to a perfect draw. That has been >>>>the direction opening theory has been taking us, and that is the destination I >>>>believe we will ultimately reach. Therefore I believe that the diminishing >>>>effects will persist. >>> >>> >>> >>> >>>The concept of entropy applied to chess. >>> >>>If you just let very good players play against each other they will produce >>>draws. >>> >>>So somebody needs to shake the chessboard! >>> >>> >>> >>> >>> >>> >>>> I think you can take one of those dead equal opening >>>>positions I mentioned and give it to one of today's programs to ponder for a >>>>year, and it will give nothing special (except some spectacularly long lines >>>>showing why neither side has anything forcible in it's favor), and then give it >>>>to a new super-duper program and it will not find any improvement. In a sense, >>>>the so-called 'bean counting' approach, much despised by Chris Wittington, makes >>>>perfect sense in a sense. If no means can be found to forcibly increase the >>>>advantage, then plant the deepest trap safely possible, and hope the opponent >>>>trips over it. The nature of that trap can be quite subtle BTW, but I do not >>>>think that in the long run one can force that extra ply to yield a win for >>>>White, and that is why we are witnessing the phenomenon of diminishing returns >>>>in the first place. >>> >>> >>> >>> >>>Chris Whittington must be absolutely delighted by your post. >>> >>>Hey Chris, can you move an object somewhere close to me, so I can feel your >>>spirit is here? :) >>> >>>(Chris thinks I have not yet understood the concept he was promoting in CSTal, >>>and which is supposed to be the ultimate response to dimishing returns, the >>>horrible fate of entropy in chess) >>> >>> >>> >>> >>> >>> >>> >>> >>>>>To use your metaphor, my point there was that a more advanced program will be a >>>>>more sensitive amplifier. That is a separate issue of whether there is much >>>>>difference to amplify to begin with in any given position and what is the >>>>>"typical" position, what is the sample. Whatever the initial difference is, the >>>>>more sensitive amplifier you use the greater the output difference, or in my >>>>>terms, the more advanced, far sighted program is the more evaluation asymmetry >>>>>it will exhibit for 1 ply initial difference (in addition to other visual >>>>>asymmetries of a position). >>>>> >>>>>To put this back in the context of the original discussion, if I am looking at a >>>>>new program with a performance within the top few, Gambit Tiger for example, one >>>>>indicator that we're dealing with a new kind of algorithms, the more advanced >>>>>species of chess programs, would be larger than typical asymmetry in its >>>>>position status display. >>>> >>>>I don't think it is assymetrical at all. >>> >>> >>> >>>Indeed it is less asymetrical than Chess Tiger 12. >>> >>>But you need to listen more carefully to Ratko and his asymetry idea. >>> >>>Remember: somebody needs to shake the chessboard. >> >>Yes, but I don't think you can do this from symmetrical positions. In other >>words, don't fight fire with fire. That is where we see the application of the >>first move appears: the symmetry can't continue forever afterall. The problem is >>that small break in the symmetry produced by that single ply isn't enough to >>break the balance sufficiently to win. It's easy to evaluate a position in which >>both opponents share the same qualities and defects: >> >>Me >> >>- Kingside castling >>- pressure on the a2-g8 diagonal >>- solid yet flexible pawn centre >>- square on f5 for a knight outpost >>- c3 pawn helps prevent an opposing piece to make use of the d4 square >> >>Opponent >> >>- Kingside castling >>- pressure on the a7-g1 diagonal >>- solid yet flexible pawn centre >>- square on f4 for a knight outpost >>- c6 pawn helps prevent an opposing piece to make use of the d4 square >> >>Not much to say unless one of those features can decide the game by a ply. Since >>no one can seem to make that happen, symmetrical openings have fallen out of >>fashion as being lifeless. Telling a program that it has the advantage for any >>reason is foolish here (IMO) as it may lead to inferior moves and hurt its play. >>The players do not believe this means that assymetrical openings are winning, >>but that is where their chance lies. Try a Sicilian Dragon for example: >> >>Me >> >>- Queenside castling >>- pressure on the a2-g8 diagonal >>- Kingside pawnstorm possible with good attacking prospects >>- semi-open d-file >> >>Opponent >> >>- Kingside castling >>- pressure on the h8-a1 diagonal >>- Queenside pawnstorm possible with good attacking prospects >>- semi-open c-file (pointed at opposing king) >> >>Here things are different for any equality is a dynamic one, an estimate. The >>different features including the all-important speed of attack make an imbalance >>easier to achieve as the equality depends not only on the features themselves, >>but their harmony. Afterall, in the first example, to say that Black's c5 bishop >>is no weaker or stronger than White's c4 bishop is easy, but what about the >>second example? How do you compare the semi-open c-file to the d-file? Or whose >>attack is stronger? A single feature comparison won't work, so you need to make >>an overall judgment call that tries to evaluate the position as a whole, and >>short of forcing lines, it is very difficult to render a verdict. >>No program is identical, so that worrying that all others will imitate your >>program's tendencies is not necessary. Even today, though The King's engine from >>CM6K and CM7k has been around for quite some time, and supposedly has special >>king attacking knowledge, no one seems to really know what that could be >>exactly. I don't think a program should be oblivious to an opponent's threats, >>nor judge them inferior merely because they are the opponent's (we only add big >>bonuses if we are the attackers). On the other hand (I'm speculating), perhaps >>optimism when rendered symmetrical, yields unnecessary pessimism leading to >>self-destructive behaviour: >> >>My opponent has a strong attack so I'm -2, but if I give up a couple of pawns >>the attack doesn't look so threatening in my horizon. 5 moves later, the attack >>never went through, but my 2 pawn deficit cost me the game. >> >>Seeking to emphasize certain aspects in the eval or including more specific >>knowledge (or something else) seem fine, but to tell the program that this only >>applies to its side because (our reasoning, not the program's of course) the >>opponent knows nothing of these things seems dangerous. >> >> Albert > > > >You are right, but there are quite a number of chess programs that evaluate king >attacks asymetrically. > >Gambit Tiger 1.0 has symetric evaluation for king attacks, but Chess Tiger 12.0 >and Crafty (current) have asymetric evals. They give more weight to the attack >of the opponent than to the attack they can create on the opponent (they are a >little bit paranoid). > >This was done in this way in Chess Tiger 12.0 because I thought I was not able >to evaluate king attacks correctly. So it was natural to assume that the >opponent had more knowledge about it, and to fear the opponent's attack. > >I assume Bob does the same reasonning for Crafty. Actually my reason is simpler... If both castle king-side, then it is quite easy to "self-attack" yourself, since every move of the kingside pawns weakens both your position _and_ your opponent's position. I found it easier to stop this nonsense by simply being more cautious... But it will definitely attack kingside, but the positional edges have to add up enough to trigger this. IE in an even position, it will refuse to play g4/g5 type moves. But if it sees it can get its pieces over there first, then that might be enough to overwhelm the asymmetry and trigger the attack. I have tried various ways to get rid of this. I have not found anything good enough, yet. As it is, I don't get attacked much by GM players, and when I do, it generally doesn't work very well (most of the time of course, there are those exceptions). Tuning the asymmetry out is trivial via command options, and several have done this. In fact, the asymmetry can even be inverted to make a wild thing. And yes, it is amusing, but no, it isn't going to beat top-10 players that know how to attack/defend. > >It is a sort of "opponent modeling" (you assume your opponent has different >abilities than yours, and you take this into account either in your search or in >your evaluation). > > >Your point about evaluation of symetry is also interesting. It is true that it >is not the same thing to have a perfectly symetrical evaluation (which the >evaluation function will evaluate as 0.00) and to have a position evaluated at >0.00 but very unbalanced by nature (each side has very different kind of >strengths/weaknesses, and maybe even the material is unbalanced). > >It is not the same, but still the programs would give both positions 0.00. I >think that with opponent modeling the positions would not be evaluated the same >anymore, but opponent modeling has not been well explored so far. > > > > Christophe
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.