Author: Robert Hyatt
Date: 19:55:35 10/20/00
Go up one level in this thread
On October 20, 2000 at 11:24:34, Ratko V Tomic wrote:
>>>> IE if my program plays Rc6 and I can prove it is correct, I am happy.
>>>> If I can prove it is bad, even though it won the game, I am not happy.
>>>> If I can't prove it either way, I am concerned. That was the point
>>>> here. I want my fate in my hands, not resting on whether my
>>>> opponent overlooks something or not.
>>
>>>You are idealizing ability of risk-averse programs. If it were tic-tac-toe
>>>you can prove move is correct. But in chess, just because some hand-put
>>>tangle of evaluation terms gives, say, 0.3 pawns more for move A than
>>>for other moves B, C,... you haven't proven move A is correct. It is
>>>only "correct" within the model game (-tree) your program substitutes
>>>for the full chess tree (where every leaf is win, draw, loss).
>>
>>You are making the assumption that "heuristics" cannot be "accurate". I
>>can give you lots of examples where this is a false assumption. IE try to
>>play a simple k and p vs k ending against Crafty. With no tablebases.
>>It only takes a few heuristics to play this perfectly, as any good endgame
>>book whill explain.
>>
>
>I didn't say that risk-averse programs have no abilities. (Endgame rules,
>table-base positions or any other position with shallow enough min-max outcome
>are all within their abilities.)
>
>What I am saying is that you have set up "happiness" criteria which ought to
>make you "concerned" in most positions since you won't be able to prove the move
>seleceted is the best/most accurate one. Just because it may appear "best" in
>your Crafty's model game tree, it doesn't mean you have proven it the best and
>that now you have your "fate in your own hands." It so only to the degree the
>model game your program creates approximates the real game tree. If GT's model
>game, after sufficient number of games, turns out to be closer to the real game
>than Crafty's model, than you can only stay "concerned" with fingers crossed
>when playing against it, no matter how happy Crafty may appear with its accurate
>evaluations.
It is _still_ a question of magnitude. If it says +.3, and it is wrong, that
isn't a game-losing error. If it says +3 and it is wrong, that stands a much
greater chance of being a game-loser.
That was my point. If we both make errors at the same frequency, but yours
cost you 10 times as much as mine, you go broke first, assuming we start with
the same amount of money.
>
>
>> I don't know whether the speculativeness of GT is good or bad. I
>> simply brought up the possibility that if a program 'gambits' away
>> positional advantages for other positional advantages, it is perhaps
>> not a game-winning/game-losing decision. But if you gambit away a
>> piece, a mistake will change the game outcome significantly...
>
>
>Almost every move "gambits away" all the alternatives. You are enshrining
>"material value" as if it were part of the rules of the game, i.e. as if somehow
>the chess rules allow you to 'cash in' at will your knight for, say, 3 pawns,
>but they don't allow 'cashing in' at will some king attack poise for 3 pawns.
>The rules allow one as much as the other, no matter what the program's wood
>count says is the "value" of this or that.
>
>Now, it's true that it is much easier, especially for the less sophisticated
>algorithms, to hold onto a material than to keep an initiative or a king attack.
>But if GT knows how to maintain and transform favorably those seemingly more
>"fragile" ("fragile" for the less sophisticated algorithms) traits of a
>position, they're as good and solid measure of position as the "material value."
If your eval is _accurate_ then your kingside attack scores could easily be
+100, rather than +3, with no problems at all. But if they _aren't_ that
accurate, then it is a form of Russian Roulette. That is what brought this
thread's subject to CCC. I saw games with scores that were _way_ outside of
being reasonable. I simply commented on them.
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.