Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Hello from Edmonton (and on Temporal Differences)

Author: Vincent Diepeveen

Date: 10:26:05 08/04/02

Go up one level in this thread


On August 04, 2002 at 11:47:01, Sune Fischer wrote:

>On August 04, 2002 at 09:13:31, Vincent Diepeveen wrote:
>
>>On August 01, 2002 at 05:16:55, Sune Fischer wrote:
>>
>>We must not think too simple about autotuning. It is a complicated
>>matter. Yet the obvious thing is that the autotuner has no domain specific
>>knowledge.
>>
>>So suppose that someone *manages* to find a good way of tuning.
>>
>>Even the simple evaluation of deep blue, which had about a few tens
>>of patterns and each pattern indexed by an array or so from 64.
>>
>>We talk about 5000 adjustable patterns (i like round numbers good) or
>>so for an average program.
>>
>>to tune that in the incredible good
>>  O (n log n) that's like (using 2 log)
>>
>>==> 5000 x 12 = 60000 operations.
>>
>>Each operation consists of playing a game or 250 at auto player.
>>No commercial program ever managed to improve by playing blitz...
>>
>>250 games x 60000 = 15 000 000 games.
>>
>>Get the problem of learning slowly?
>
>No, TDLeaf is a steepest descent algorithm, if it works it will go much faster
>because it's going directly against the gradient.
>I'm not saying it will be easy, or that it won't require a large number of
>games, but I believe its potential is greater than what is humanly possible.
>
>
>>Now we talk about a simple thing called chess, just 64 squares.
>>If i autotune something to drive my car there are a zillion parameters
>>to tune ;)
>
>Yes, but you tune them _all at once_ so it's really not that bad :)

Exactly the answer i wanted to hear. This is simply impossible to tune
them all very well at once without domain knowledge. If you get a perfect
tuning at

O (n log n) you already get a nobel prize.

The tuning we need, and definitely traffic, is not 'all a little tuned',
errors are not acceptible simply. One parameter which is -1.0 instead of
+0.22, that's an unacceptible error simply. Hand tuning is even
more accurate. It even sees the difference between 0.22 and 0.40 very
clearly. In case of the pro's, even the difference between 0.22 and 0.20
is getting felt in some cases (bishop vs knight).

Obviously you won't achieve such accuracy under O (n log n) with TD learning,
it's a rude and primitif algorithm which can be considered lucky if the plus
and minus sign are already guessed right.

The pro's are not far off perfect tuning nowadays. Of course the parameters
can get improved, but definitely not their tuning. I doubt you can achieve
much better tuning with crafty too.

It's programs like DIEP where tuning can be improved bigtime, but of course
we talk about FINE tuning. Whether it's 0.023 instead of 0.020.

In DIEP i work not at 1/100 of a pawn but 1/1000 of a pawn. Another 1000
horrors more :)

Best regards,
Vincent

>-S.
>
>
>>>On July 31, 2002 at 18:10:08, Vincent Diepeveen wrote:
>>>
>>>>And he'll never do a competative chessprogram again either, he forgot to
>>>>add that too.
>>>
>>>It's the classic problem. Person A has spent 15 years on his chessprogram, he
>>>thinks he knows it all. In from the right comes a guy who knows nothing about
>>>chess programming but is very good a tuning weights in general. He applies his
>>>knowledge to a simple chess program and is very successful in increasing it's
>>>strength.
>>>Person A now concludes; "his method isn't working, his program is still weaker
>>>than mine":)
>>>
>>>>I remember Knightcap very well. TD learning had the habit to slowly
>>>>make it more aggressive until it was giving away a piece for 1 pawn and
>>>>a check.
>>>>
>>>>Then of course the 'brain was cleared' and experiment restarted.
>>>>So in short the longer the program used the TD learning the worse it
>>>>would play, from my viewpoint.
>>>>
>>>>Definitely from a chessplayers viewpoint it did. Of course we must not
>>>>forget that in the time it played online, that nearly no program was
>>>>very aggressive. So doing a few patzer moves was a good way to get from
>>>>perhaps scoring 11% to 12% or so.
>>>
>>>So in other words, if you teach it the wrong things it doesn't work?
>>>Why am I not surprised.
>>>
>>>-S.



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.