Author: Ratko V Tomic
Date: 18:27:28 10/14/99
Go up one level in this thread
>> >>Even the 18 plies in complex middle game (which may be 2-3 years off) still >>hinge on the same old simple-minded evaluation of the terminal nodes. >> Unless you find a clear cut advantage at that depth, the judgment of >> such position by a strong human player is superior to material, square >>count and such simple criteria. > >There is a point that you are not taking into consideration. A human will not >even look at the position 18 plies down unless there is a basically forced >variation. To say that a human can evaluate the position better 18 plies down >ignores the fact that the human does not have the time to evaluate the billions >of positions to get that deep. > I may have not expressed very clearly, since I don't disagree with your statement as it stands, but neither I disagree with my earlier statemnt. My point is that computer evaluation at 18 plies, unless it discovers something decisive (i.e. a tactical shot) isn't going produce much more strategic understanding or coherence (other than what is tactically based). E.g. if it sees that move m1, 18 plies later leads to a gain of mobility by one square, that is no more important than finding such mobility gain only 4 plies deep. Mobility shifts back and forth almost every move, so at 19 or 20 plies it will be something different again. Unless program can see something with a more tangible (stable, permanent) gain, the fluid component of the evaluation is a tossup which doesn't improve noticably (sometimes not at all) with depth. In a non-tactical/quiet position you will have 3-4 or more moves giving almost identical score, with tiny essentially random fluctuations due to such inconsequential fluid components of the scores from deep down the tree. So, as the game goes on, the program picks one or the other among these for irrelevant and incoherent reasons, like a sophisticated random number generator. A human player may have a long term plan, with its goals and intermediate subgoals. While he won't compute brute force 18 or more plies, his plan may unfold over twice as long span, by the time all the subgoals are realized, and his moves at this moment do fit some far future objective. So in a strategic battle, the difference is like the one between someone laying stones along the plan for a house and someone tossing stones optimized to make pile as high as possible. The program's style will switch essentially randomly between moves optimal for different long term plans, doing and undoing their long term effects (as related to those plans). These small effects will thus not add up coherently into a winning strategy as they do in high level human play. Of course that doesn't mean program has to lose to the strong human player. The human's plan may be unrealistic or the means or subgoals may not be well chosen, or there may be a tactical loophole in the particular means chosen, etc. But they are acting as if playing two different games, almost independently from each other. That "two separate worlds" (or two models) picture becomes quite obvious in positions where you are setting up for a king side attack, while the program is plotting how to snatch your b2 pawn in a brlliant 14 ply combination, oblivious to your pending attack 25 plies ahead, as you're to the tiny b2 pawn within your model (which even if it could see the pawn loss, wouldn't care much anyway). >Humans play a VERY stupid game, even Kasparov. It just so happens that > they play in this stupid manner against others who play even more stupid. > In our arrogance, we believe that the GMs and superGMs are playing > brilliant moves. And to us simple mortals, they do seem brilliant. But > to a program that can avoid the tactical pitfalls for 50 ply down, some > of the moves would appear to be extremely weak. I wouldn't say stupid, just different model from the program's model of the game. I would also suggest that our type of model is more sophisticated and powerful, due to great flexibility in addapting evaluations and types of thinking to the positions. >Chess is a game of tactics and only tactics. True but only from a perspective of fully evaluated (down to checkmate) tree. Otherwise it is a computer style guesswork, with its inexact node evaluations on the truncated tree, against the human imperfections. At the moment the computers, while a bit behind the top players, are catching up, but the tide may easily turn decisively against the present alpha-beta searchers, hardware advances notwithstanding, once the anti-computer strategy (which is in its infancy) evolves enough. So far it was mostly a free ride for the programs, with the strong humans playing mostly with the same fundamental positional strategy they use against other humans (with few limited value ad hoc guidelines for play against computers, similar to strategy shift a positional player might adopt against a stronger tactician human player). The right strategy may be something entirely different and seemingly paradoxical, something which would tend to lose to other strong humans. > >This is somewhat misleading. Although there are endgame positions and won >positions where the tactics are understandable, the vast majority of positions >are ones where there are hidden tactics, at least hidden to humans. > And hidden to programs with inexact node evaluation (or incomplete tree). I am sure there is huge number of tactical variations which "win" extra square of mobility or similar evaluation function point in program's book, but so waht. Unless the tactics hidden to human and visible to program gains something tangible of persistent value, convertable eventually and ultimately to the checkmate, it's no good by itself. You're speaking as if programs "discoveries" of some minor gain by their count are same as point gain in a game. Unless program sees it through to the end or to a decisive advantage, whatever it sees which is hidden to its human opponent is an idle speculation. If that were not so, programs would be winning serious games decisively against strong humans, and they're surely not doing that. Namely, you can only compare how well and how fast programs do in solving problems compared to humans. If there is a decisive gain within horizon, they'll find it quickly and accurately, much better than humans. But since they're not winning decisively against strong humans (other than blitz or other forms of play where human slowness handicaps them), that only means the bulk of chess positions encountered don't have a decisive tactics, but have kind of tactics I described at the top, the almost random picks of "wins" of irrelevant evaluation points (shifting randomly well within the error fluctuation band of the evaluation functions themselves).
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.