Author: blass uri
Date: 10:06:09 11/28/99
Go up one level in this thread
On November 28, 1999 at 10:05:52, Robert Hyatt wrote:
>On November 28, 1999 at 04:08:19, blass uri wrote:
>
>>On November 27, 1999 at 22:28:32, Robert Hyatt wrote:
>>
>>>On November 27, 1999 at 19:46:52, Fernando Villegas wrote:
>>>
>>>>On November 27, 1999 at 17:33:13, Robert Hyatt wrote:
>>>>
>>>>>On November 27, 1999 at 12:18:07, Fernando Villegas wrote:
>>>>>
>>>>>>I do not understand your point, Bob. This is not a match between two computers,
>>>>>>but many. How a program could do well just tuning against Tiger? Maybe that
>>>>>>could mean to un-tune against any other of the concurrence. Maybe some opening
>>>>>>preparations, but...
>>>>>>Fernando
>>>>>
>>>>>
>>>>>This is easy. A year ago, due to some unusual new eval features I added, I
>>>>>ended up with a version that had very little trouble with Fritz 5 at any time
>>>>>control. It won so many games that Lonnie accused me of using a Cray to play
>>>>>against him. If I sent _that_ version to the SSDF for testing, it would have
>>>>>done very well against fritz, because fritz would be totally unprepared. But
>>>>>once they saw what was happening, some adjusting on their end (king safety and
>>>>>passed pawns in particular) and this advantage would have eyvaporated.
>>>>>
>>>>>Almost always the _last released_ program goes to the top of the SSDF. In this
>>>>>case, it is an _unreleased_ version, which means _nobody_ had a chance to look
>>>>>at the book, and the depth, at the evals, and find out what it is doing....
>>>>>
>>>>>Sort of an "element of surprise"...
>>>>
>>>>
>>>>Please let me clear this issue a little more.
>>>>a) SSDF testing is not made by the programmers so they couldn't tune his
>>>>programs according to every new opponent.
>>>
>>>I didn't say (or imply) that they did. I am simply saying that Tiger has
>>>had ample opportunity to play against other programs in private... and that
>>>once it becomes public, other programs will have ample opportunity to play
>>>against it. And as usually happens, things will then change in surprising
>>>ways...
>>>
>>>It has _nothing_ to do with the SSDF... just that the program has not been
>>>'seen' by anyone else. You will be surprised what you can learn about a
>>>program after watching its analysis for a while... So far, that hasn't
>>>happened. But it will...
>>>
>>>
>>>
>>>
>>>
>>>>b)Being so, if, let us say, anyway F6 is delivered to the swedish people after
>>>>being tuned against Tiger 12, my question is, what would happen to F6 against
>>>>other programs? Why tunning against program X neccesarily means the likelihood
>>>>to get more points in a pool constituted by many opponents different to X?
>>>
>>>It happens all the time. At one point Ed had 8 machines running auto232 matches
>>>against his 'competition'. If one program beats you consistently, you can find
>>>out what it is seeing that you are not, and fix that without breaking yourself
>>>vs other opponents...
>>>
>>>
>>>
>>>>c) That is the core of my question and the only way I found to understand your
>>>>point is to suppose that Tiger 12 brings not only some specific features that
>>>>could be anulated so and so, but general, universal improvements so if you tune
>>>>againts them, you improve your own program "in general"
>>>
>>>That is possible as well. Several years ago I added "outside passed pawn"
>>>code to Crafty. At that time, hardly any commercial programs did anything
>>>with this and as a result, crafty won many a endgame due to this. It wasn't
>>>long before it worked less frequently. Ditto for the trapped bishop on a7.t
>>>A few 'fixed it'. A few fixed it before I started evaluating it myself. But
>>>not everybody... a couple of programs _still_ fall for it.
>>>
>>>Another good example is king safety. I don't know of any program (so far) that
>>>is very good handling king safety. Once someone does a program that is really a
>>>strong attacker, everyone will either fix it finally, or get rolled into small
>>>balls over and over. (Older versions of Genius suffered badly from this,
>>>although I don't know about the newest one).
>>>
>>>
>>>
>>>
>>>
>>>>d) In this way, tunning againts Tiger 12 would means tunning against any other
>>>>program.
>>>
>>>And quite possible tuning to do _worse_ against human players. Which is not
>>>something I am ready to do yet, myself. But for those driven by SSDF rankings,
>>>anything goes..
>>
>>I believe that the result of tuning against computers will be usually also
>>better results against humans.
>>
>>You can do better results against humans at fast time control by using not
>>objective evaluation to go for positions that humans do not like but I believe
>>that this idea is going to fail against strong humans in tournament time control
>>when they do not do the tactical mistakes that they do at blitz.
>>
>>Uri
>
>
>You might believe that to be true, but I _know_ it is not. Because when you
>play against a computer, you play against something that is _very_ incomplete
>as a chess player. If your 'computer' opponent doesn't understand blocked
>positions, it will probably try to avoid them.
If the computer who does not understand blocked positions is playing agianst a
strong human at slow time control the computer is going to do a positional
mistake by trying to avoid the blocked position and is going to lose the game
because of this positional mistake.
I believe that avoiding blocked position can help at fast time control against
humans but at tournament time control against strong humans it is not going to
be productive and the right way is to teach the program to play blocked
positions.
So you will _never_ see how
>badly you do against an opponent that does understand them. If your program
>doesn't understand pawn majorities, and your opponent doesn't, then you will
>_never_ see how badly you do against a GM that understands them well.
>
>The two pools of players are far different. yes, a computer is tactically
>stronger and if you tune to handle that weakness, you get better. But that
>is the _only_ place where computers are even close to the Grandmasters, and
>all the other weaknesses they have won't show up when you play a computer that
>doesn't understand anything but tactics...
computers understand other things except tactics.
If you do a program with only material evaluation you are going to lose against
computers even if you have a 10 times faster hardware.
Uri
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.