Re: handicapping engines
- From: Johnny T <not@xxxxxxxx>
- Date: Mon, 02 Jan 2006 20:50:29 -0800
Major Cat wrote:
Yes, I would imagine that a drastic and, by definition, _uniform_ restriction of the ply-depth horizon would go a long way towards _simulating_ chess novice play.
No, it really wouldn't. It would exactly behave as a computer that cannot even see one move ahead. If a novice had no knowledge of chess, and just randomly moved pieces, I really wouldn't see the fun or use of creating a computer rendition of that.
But creating a program that actually played as a novice. Knows how to move the pieces, and nearly always at some point hangs a piece. That could be very useful. It could help people play against novices. Providing you know, those people are 6 and are playing for their school against players who do exactly that.
I don't know why you think time is different than ply depth. By definition you are limiting the depth the computer can see by limiting it's time. It is just a different way to do the same thing.
Talking about "chance", as you know, the probability of a proba- bility reduces to just a probability! 8>) What I am driving at here is this. Does it really matter whether the program misses some deep tactic because of a time (_not_ ply-depth)
The point of randomly pruning the tree, is that it creates the same kind of blindness that humans do. It gets wrapped up in it's own plan and is blind to other possibilities.
Now, class players do blunder tactically from time to time. The challenge would be to believably incorporate this aspect into the coding.
This doesn't happen from time to time, it happens nearly every game. Especially after "book" openings. The class player will almost always will miss a tactical opportunity or lay one open.
Unlike tinkering with search depth and permanent brain, time control calibration is subject to way less granularity.Well actually all these effect the same point as they just limit the horizon the computer can see. The same way as my P166 limits the horizon of course. Using a slow processor just gives you a chance to be able to enter certain seconds for the search depth ;)
Well yes. Limiting the ply-depth horizon indirectly and non- uniformly seems to _simulate_ class player tactical behavior quite well. Excepting the simulation of rare, "obvious" tactical blunders of class player behavior, why is this not adequate?
Because they are always "horizon" errors. "Horizon" errors are not the errors that class players make, it is the type of error that computers have historically made. This leads to "anti-computer" chess. Which is not the way to beat humans, but to beat computers. It is nowhere near the kinds of errors that class players make.
The blindness is not the same. Class players can often see somewhat deep and have plans. They often miss the obvious, one way or the other. A computer will never miss the obvious, but never see deep. Two completely different ways to error, and they types of errors are quite different. You are not likely to ever see computer type errors from humans.Moreover, the ensuing "blindness" is less uniform (from move to move). Also, why are "classic computer errors" so different from those, say, one comes across in games between class players?
Actually I do not think they allways are, but if you think about this "mate in 4" error mentioned in some earlier post, the computer will not find a mate in 2 if the horizon is to short and he has no time to come out of his valley (defined by "I have a mate in 4").
With permanent brain on, this is a virtual impossibility...
It depends on how the computer ponders. Against a class player a permanent brain may often not be useful, because the class player isn't playing the "best" move.
Yes, but are these differences _radical_ enough? Ok, the issue of positional play has come up numerous times. But, when it comes to tactical evaluation, are computer programs _consistently_ different from one another in "OTB" play mode?
Yes, they are all very different. In many "key" situations, they may all find the "correct" answer given enough time, but in many non-key situations, they have different coding for "chess knowledge" that will evaluate the positions differently, and one one-hundredth of a position value point is all the difference in the world to a computer.
Well, I have been "complaining" about this kind of "black box" ignorance imposed upon power users of software programs for half a generation... 8>)
That is because until recently there has been little real research done on how to create computers that play as poorly as humans do. But rather to have them play as well or as better as the best. But the race is over, at least until Hydra gets beat.
Now the next race is to take a step back, now that we know how to play as well as some humans, the key would be to go from there backwards to play as poorly as most humans. It is much more likely that it will be by it's mistakes a chess program passes the Turing test, not by it's machine like disposal of it's opponent. That is when a chess programs primary purpose is to help find the tactical truth of a position, but to truly be a worthy opponent regardless of the opponents strength. For it to be your slave not your master.
Actually if you leave out ChessBase here (these engines are DLL's, so pure binary code, additionally only for Windows...), all the UCI enigines export these knobs by default as it is definied in the interface. Ok, the programmer might decide not to export really all knobs available, but quite a lot are really there. You can even have a look at them (if your gui hides them e.g.) by just starting the eninge on a command line and issue "uci". Though I agree that there lacks the documentation on the details. Anyway, I don't think that this is just "keeping things secret", but more or less that it might be clear to the programmer what a setting like "Rook (Engdame)" means and that it also obvious why to set it to 110 or 90.
I would disagree about this somewhat. Yes, you can change some of these settings in some engines, including chessbase ones, but UCI ones as well. But it is the incorrect application of the rules, at different times *DURING* the game that will mark a class player. It is the dynamic change of application of chess knowledge that a computer needs to do, not to statically change. You have to both miss tactics, and make mistakes positionally, and sometimes you are allowed to be perfectly correct. But neither you nor your opponent should know at the time.
I doubt it. This has been going on for a long, long time. It is a behavioral canon. For example, Microsoft deliberately moved to "exterminate" the power user when it released Windows 95...
I agree, and disagree. I think you are going to see first a bearing of the soul for a bit, as people experiment with "intelligent mistakes". As they get comfortable with what works, then you will see those get reburied again.
- Re: handicapping engines
- From: Major Cat
- Re: handicapping engines
- Prev by Date: Standardized rating reporting.
- Next by Date: whistle
- Previous by thread: Re: handicapping engines
- Next by thread: Re: handicapping engines