Author: Christophe Theron
Date: 21:24:54 10/22/00
Go up one level in this thread
On October 22, 2000 at 15:23:30, Ratko V Tomic wrote:
>> I cannot think of any other purpose for all this other than to improve
>> the program's rating in the SSDF.
>
>There is no harm all by itself in being at the top of SSDF.
>
>> You can't possibly use this for
>> human players, and you aren't talking about offering playing styles,
>> as that is already rampant.
>
>Human players routinely adjust their play to exploit better the stylistic or
>capabilty weaknesses of their opponents, programs or humans. Programs aren't as
>good (yet) at this kind of 'outside the board' strategizing, but they will get
>better.
I believe that the current chess programs/hardware combinations are already at
the level or even better than GMs "on the board". Don't shoot at me right now.
But the GMs are much much better at taking advantage of the "off the board"
knowledge (they know their opponent and can adapt their strategy to it).
This "off the board" knowledge and ability to adapt gives them an serious edge
over computers. There is a big gap to close for computers because of this. It is
maybe 50 elo points at this time (maybe even 100 elo point for a few
individuals, but there are not many of them), and it might well be like the
horizon. No matter how close you get, the horizon moves away.
Increasing the computers speed will not solve the problem immediately. The gap
should be closed by introducing some opponent knowledge into the programs
(opponent modeling).
Christophe
>> Creating a number of personalities to defeat specific opponents would
>> require a lot of work and would only be of interest to the
>> programmer. Why would I want to have a Rebel especially geared to
>> defeat GM Svidler for example?
>
>I wasn't talking about specific opponent but only a specific relative advantage,
>no matter who the opponent is (i.e. the program is being tuned to a set of
>opponents with particular strength range in particular types of positions). The
>usefulness of this is surely greater than just better SSDF rating. After all if
>the program adjusts itself to maximize its strength when playing against you (or
>whoever else you pit it against), for all you care you're getting a stronger
>sparring partner, as if you bought a faster computer and used the same program
>without self-tuning.
>
>GT is quite effective against programs which don't have as good king-attack
>algorithms, more so than other good king attackers, including regular Tiger,
>probably because it does take risks. The proper degree of risk taking ("proper"
>meaning as empirically determined) is the performance maximizing strategy. By
>taking risks, GT makes positions in which it excells more frequent. While
>regular Tiger may be as good, even better, once the king attack is under way, by
>virtue of taking fewer risks it won't find itself in a king-attack as often as
>GT will. To maximize performance (rating) it is not enough to have an edge in
>some area. One has also to try maximizing the frequency of positions in which
>one can use that edge, even if some risks are involved.
>
>Now, once you agree that it is reasonable to take some risks in order to
>increase frequency of positions in which one will have an edge, and since the
>degree of advantage varies for different opponents, the immediate consequence is
>that one could do even better if the amount of risk-taking is tuned to the
>specific degree of the advantage one has against the current opponent. And that
>can be learned in a longer match or hand input if the operator knows the
>opponent. For example, GT might perform better against regular Tiger if it were
>to play with risk-taking "instinct" minimized.
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.