Re: handicapping engines
- From: Major Cat <epikuros@xxxxxxxx>
- Date: Fri, 30 Dec 2005 15:57:49 -0800
Thank you for a most interesting post. Before I comment on it,
would you be kind enough to clarify for me whether your sug-
gestions are aimed at programmers as opposed to just power
users of chess programs? It seems that the former is the case
but I am not entirely sure. In any case, for a power user to
be able to implement your interesting ideas, two things are
necessary. First, the "levers" must be available and accessible.
Second, sufficient documentation must be made available to power
users so that they are able to intelligently calibrate things...
Johnny T wrote:
> All you are doing with the variety of things that you are listing is to
> shorten the computers horizon. This limits the kinds of mistakes the
> computer can make, and will affect some programs worse than others.
>
> I think the correct way to handle mistakes is two fold.
>
> First is incorrect pruning. This makes entire sections of the tree
> "invisible" to the computer. This is very similar to the kind of
> "blindness" that humans have. I think the solution for this can be
> found in all places "D&D". In essence there is a certain percentage
> chance that a branch will be incorrectly pruned. The lower the ranking
> the higher the percentage all the way down to hanging a piece, to
> missing certain tactical combinations.
>
> You can also have certain types of branches more easily pruned. Things
> that are common tactical errors for a given class of players, forks,
> skewers, sacrifices etc. For instance, a low ranking setting may never
> "see" tactical combinations that involve sacrifices. Then may only
> see their own, or may never see knight forks, for instance.
>
> You then have a certain percentage chance that a branch will become
> "visible" on the next turn. This allows the computer to also behave
> like a human who all of the sudden discovers his opponents plan.
>
> The second method is unbalanced chess knowledge. Normal Human Players
> (you know, kasparov excluded), make incorrect positional plays, or don't
> follow through on the plan correctly.
>
> Again using the "D&D" methodology, there is a certain percentage chance
> that things like king safety, pawn structure, openness, and other
> heuristics will become unbalanced, either by a lot, or by a little, and
> either often or not very often, depending on the "rating" that you are
> trying to set the program for.
>
> And again there should be two levels of change, one sort of overriding
> kind of "wrongness", second a dynamic changing to the "wrongness" during
> the game.
>
> Unfortunately, it will be a sheer guess whether or not you have set
> these correctly, unless you could do a lot of testing against say
> 1400-1600 players. But I think many of your customers would tell you
> if you were getting it right or not.
>
> Remember, that you have to let the computer make the best move from time
> to time, just like humans do. It is how often, and the quality of the
> wrong moves that you are trying to improve.
>
> Unfortunately just lowering the "horizon" tends to create "computer"
> errors, and tends to encourage "anti-computer" chess, which is not what
> you are really trying to do. You are trying to give the human hope
> that the computer will err, in the same quality and quantity of those
> that play just a little bit better than them. To provide struggle,
> discovery, and possibly victory.
Regards,
Major Cat
.
- Follow-Ups:
- Re: handicapping engines
- From: Johnny T
- Re: handicapping engines
- From: David Richerby
- Re: handicapping engines
- References:
- handicapping engines
- From: Magnulus
- Re: handicapping engines
- From: Johnny T
- handicapping engines
- Prev by Date: pgn alternative
- Next by Date: Re: software to practice openings
- Previous by thread: Re: handicapping engines
- Next by thread: Re: handicapping engines
- Index(es):
Relevant Pages
|
