Author: Christophe Theron
Date: 00:09:08 11/12/00
Go up one level in this thread
On November 11, 2000 at 14:07:51, Enrique Irazoqui wrote:
>On November 11, 2000 at 13:51:24, Christophe Theron wrote:
>
>>On November 10, 2000 at 14:57:06, Joe Besogn wrote:
>>
>>>On November 10, 2000 at 14:12:31, Christophe Theron wrote:
>>>
>>>>On November 10, 2000 at 09:16:06, Mogens Larsen wrote:
>>>>
>>>>>On November 10, 2000 at 08:41:24, Uri Blass wrote:
>>>>>
>>>>>>The question is how do you combine the 2 programs.
>>>>>
>>>>>Exactly. That was the point of my last paragraph. I suppose it would be
>>>>>essential to analyze where the strengths and weaknesses of the programs differ
>>>>>and interpret them as patterns recognizeable for some sort of evaluation
>>>>>function.
>>>>>
>>>>>I was talking with Bob about applying SMP knowledge to run different engines at
>>>>>the same time and base the selectivity on an advanced evaluation engine. I still
>>>>>think it's possible to implement and get good results.
>>>>>
>>>>>>I thought about the idea of a root processor that decides based on the root
>>>>>>position which engine to use.
>>>>>
>>>>>What should a root processor "look" for?
>>>>>
>>>>>>It can be based on previous evaluations of the program.
>>>>>>
>>>>>>I think that starting to use tiger13 when Gambittiger shows big scores may be a
>>>>>>good idea because when gambit shows a big score the job of attacking was done
>>>>>>and the problem of tiger that I found in the game against crafty that was posted
>>>>>>by thorsten was that it did not like to go to an endgame and prefered
>>>>>>speculation of 3 pawns and not winning a rook for a bishop but I do not believe
>>>>>>that it can lead to an improvement of more than 20-30 elo.
>>>>>
>>>>>It might be useful to make certain that the high evaluation of GT is maintained
>>>>>for x number of moves before concluding that the attack has accomplished a
>>>>>winning position by transposing to the endgame. Basically that would mean
>>>>>reducing or removing "speculation" altogether at a cetain point in the endgame.
>>>>>
>>>>>Maybe it's possible to encourage or favour the speculative approach in certain
>>>>>positions where the structure is immoveable in order to open up for the tactics.
>>>>>That might not necessarily be a good idea.
>>>>>
>>>>>Mogens.
>>>>
>>>>
>>>>
>>>>This discussion between you and Uri sounds interesting, and the ideas might be
>>>>true and useable.
>>>>
>>>>However I see a much simpler way to get an already significant improvement by
>>>>using Chess Tiger 13.0 and Gambit Tiger 1.0 together, and any owner of the Rebel
>>>>11 CD can do it:
>>>>
>>>>The trick is very simple: just decide to use either Chess Tiger or Gambit Tiger
>>>>depending on which opponent you have to play.
>>>>
>>>>During the beta test period, we have gathered informations about the performance
>>>>of the two engines, and we have noticed that against some opponents Gambit Tiger
>>>>was scoring very well, while against other opponents Chess Tiger did much
>>>>better.
>>>>
>>>>I should dig into my archives to find the right information, but I remember that
>>>>for example Gambit Tiger 1.0 crushed Shredder 4, and did 50% against Junior6.
>>>>Chess Tiger 13.0 did much better against Junior6. This has been reported by
>>>>several testers, who have played matches independantly.
>>>>
>>>>These two opponents are just an example. The important point is that Chess Tiger
>>>>and Gambit Tiger have different winning profile against different opponents and
>>>>it must be possible to take advantage of this.
>>>>
>>>>Given that against a given opponent the difference in winning percentage between
>>>>Chess Tiger and Gambit Tiger can easily be around 10%, I guess that the
>>>>difference between "using always the same Tiger engine" and "using the right
>>>>Tiger engine for an opponent" could be 5% in average, which makes an elo
>>>>difference of 35 elo points.
>>>>
>>>>Given the current level of competition between chess progams, 35 elo points is a
>>>>significant advantage.
>>>>
>>>>If the idea of selecting the engine depending on the opponent sounds shocking to
>>>>you, remember that this strategy is a main source of success for human players.
>>>>Human players KNOW who they are playing, and adapt their style according to
>>>>their opponent (and also according to the specific conditions of the game).
>>>
>>>Scenario:
>>>
>>>Big deal championship match between top program (better than Deep Blue) and
>>>world human champion.
>>>
>>>1. Top program can use any style it wants, play in a way previously unseen and
>>>kept hidden by the developers.
>>>
>>>2. Or, top program has a relatively consistent style, changes openings, but
>>>style variations are like a humans, not too much, so the style remains
>>>identifiable.
>>>
>>>
>>>(2) is what happens when human plays human, each can look at the old games, and
>>>get a general idea of what they expect. At GM level this research is very
>>>important, and knowing-the-opponent is a GM strength.
>>>
>>>Your proposal, while doable, takes human-computer games into the asme field as
>>>DB-Kasparov, ie the human has no pre-knowledge. This is unfair and should be
>>>outlawed.
>>>
>>>Thus I propose Rule 1 of computer program behaviour: a programmer may not make
>>>unknown or varied styles that enable the program to gain a surprise advantage
>>>over a human.
>>>
>>>Rule 2 should outlaw similar behaviour against other programs.
>>
>>
>>
>>
>>You must be kidding?
>>
>>There is no way to force a human player to play in his usual style.
>>
>>There will be no way to force any program to play with a "consistent"
>>predictable style.
>>
>>Your "forbidding rule 2" is completely stupid and impossible to enforce. It
>>would prevent programs to reach higher levels of intelligence anyway.
>>
>>If I want to write a polymorphic program and use it in tournaments, who's going
>>to forbid me to do so???
>>
>>
>>
>>
>>
>>>Users need to know and have a relatively consistent style when they purchase or
>>>download, not some randomising monster.
>>
>>
>>
>>
>>Users call programs with predicatble style that you can beat several times in
>>the same way "stupid".
>>
>>And they are right.
>>
>>Looks like you want the programs to stay stupid forever...
>>
>>It can be a program feature to let the user set the playing style (passive,
>>agressive, whatever...), but by default the program on its strongest setting
>>should be able to vary its playing style.
>>
>>
>>
>>
>>>The things these programmers will do to gain rating points.
>>>
>>>
>>>>
>>>>The players able to adapt gain a lot of points because of this, and this
>>>>strategy is considered with respect amongst human players.
>>>>
>>>>I think that using Sarah's tournaments results (which are public) it must be
>>>>possible to see against which opponents Gambit Tiger performs better. It should
>>>>also be possible from these public results to see if by selectively using Gambit
>>>>Tiger or Chess Tiger it is possible to get a much better performance for the
>>>>"Tiger family".
>>>>
>>>>Then an experiment could be done in order to verify that the improvement is
>>>>consistent and reproductible.
>>>>
>>>>I believe an elo improvement of 35 points is possible.
>>>>
>>>
>>>Old paradigm thinking. Rating points are not important any more, Doh! It's style
>>>of play. Tiger's style of play (if it is what is reported) is what is important,
>>>not some silly chasing of rating points.
>>
>>
>>
>>
>>There are people interested by rating points. It's my work to satisfy them as
>>well as people interested by playing style.
>
>Most people are probably interested in both, as I am and I know you are. Why
>should you accept a dichotomy that doesn't make any sense?
I used to believe in strength only, but I have recently learned that it is also
possible to dramatically change the style without significantly weakening a
program.
This was a novelty for me. Before that I thought that any attempt to
specifically change the playing style would result in a badly "untuned" engine,
a much weaker one.
So let's go for both strength and style. Why not?
Christophe
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.