Evaluation vs. Search

[AdminX]

Thanks to the current Zappa vs. Rybka Match, There has been a very interesting discussion about Evaluation and Search. Mark Uniacke on this topic posted the following:

"I agree with the view that Rybka's main strength is in its fast well aimed search. Like you I have come to realise that search is king in the land of chess computers once a reasonable level of eval has been reached. (Warning this does not necessarily apply against humans)"

"Sadly, the breakthroughs in computer chess have mainly come from improvements to the search, whereas the eval improvements have been a more gradual creep."

Where as Anthony Cozzie had this to say:

you are caught in the trap of "evaluation = position", "search = tactics". Evaluation can see tactics and search can see positional moves.

So I really think search and evaluation are simply different options to try to improve play. Obviously I think eval works better while Vasik thinks search works better. So far in Mexico things are proving to be more or less a tossup

And Rolf has this to say on this site:
http://www.talkchess.com/forum/viewtopic.php?t=16717

With respect I must disagree. What the author of Zappa doesnt see is that he's caught in a lingual trap. If a program would evaluate the deep positions carefully and exactly enough it had no horizon limitation. I think one cannot argue that Rybka has a "better" search and Zappa the "better" evaluation, no, if the search is well, it's because the evaluation is also well.

I don't know about you, but I think it will be great to hear what Vas has to say on this topic once the match is over today. Even if this match does not prove anything I give a small edge to search here. Hey Dr. Hyatt if you are listening, would like to get your thoughts on this.

Thanks in Advance

For Complete Context see link below:
http://www.hiarcs.net/forums/viewtopic. ... &start=105

[swami]

There has been a very interesting discussion about Evaluation and Search. Mark Uniacke on this topic posted the following:

For Complete Context see link below:

And where is the link?

[Ryan Benitez]

There is no doubt in my mind that search is the fastest way to the top of the mountain. Look at what happened with the last set of Fruit releases. Fruit 2.3.1 was almost all eval changes from Fruit 2.2 and Fruit051103 was only small search changes from Fruit 2.2. Only with the setting from Marc did the 2.3.1 version do slightly better than the version with search changes. Per time spent search clearly wins.

[Bill Rogers]

Hello Ted
I disagree with the above premise for the following reasons. Over the years here have been some great improvement in the evaluation function and I would say that most of them have not only been made public but are employed in almost all modern chess programs. While the faster seach algorythms have helped I don't think they have helped as much as people want to give them credit for.
There is a very simple way to prove the difference for those who want to give it a try. Take the source code from an old, old engine and rewrite the search routine only and then test it against some newer engines. After that testing is done take the same engine and only rewrite the eval function and then test it once again. In this manner you can have three seperate engines that you can run various test with. I feel resonably sure that the eval improvements will win out in the long run that is of course if the primary engine does more than a 3 to 6 ply search.
I could be wrong on these assumptions but I think it is worth a try and it would answer the question maybe once and for all.
Bill

[AdminX]

Hello Ted
I disagree with the above premise for the following reasons. Over the years here have been some great improvement in the evaluation function and I would say that most of them have not only been made public but are employed in almost all modern chess programs. While the faster seach algorythms have helped I don't think they have helped as much as people want to give them credit for.
There is a very simple way to prove the difference for those who want to give it a try. Take the source code from an old, old engine and rewrite the search routine only and then test it against some newer engines. After that testing is done take the same engine and only rewrite the eval function and then test it once again. In this manner you can have three seperate engines that you can run various test with. I feel resonably sure that the eval improvements will win out in the long run that is of course if the primary engine does more than a 3 to 6 ply search.
I could be wrong on these assumptions but I think it is worth a try and it would answer the question maybe once and for all.
Bill

Yes, you could be wrong. But so can I. Interesting theory to test by the way.

Thanks

[Uri Blass]

There is no doubt in my mind that search is the fastest way to the top of the mountain. Look at what happened with the last set of Fruit releases. Fruit 2.3.1 was almost all eval changes from Fruit 2.2 and Fruit051103 was only small search changes from Fruit 2.2. Only with the setting from Marc did the 2.3.1 version do slightly better than the version with search changes. Per time spent search clearly wins.

Note that it is a problem to compare between them because you may have pruning rules that are based on evaluation factors so both search and evaluation go together and better evaluation can also improve your search.

Uri

[Ryan Benitez]

There is no doubt in my mind that search is the fastest way to the top of the mountain. Look at what happened with the last set of Fruit releases. Fruit 2.3.1 was almost all eval changes from Fruit 2.2 and Fruit051103 was only small search changes from Fruit 2.2. Only with the setting from Marc did the 2.3.1 version do slightly better than the version with search changes. Per time spent search clearly wins.

Note that it is a problem to compare between them because you may have pruning rules that are based on evaluation factors so both search and evaluation go together and better evaluation can also improve your search.

Uri

You are right, eval does affect pruning decisions in Fruit so I can not say that just because the changes are in the evaluation function that eval is the only thing affected.

[Dann Corbit]

The search verses eval debate will rage forever. It is just like the bitboard verses 0x88 debate (bitboard was the rage when Crafty did well at Jakarta. Then 0x88 took over and all the strongest programs used 0x88. Then Rybka came along (using bitboard) and the tide changed again.

I think (as far as eval verses search) Christophe said it well:
Eval and Search both produce knowledge.

If a chess program is to play really well, it had better search well. It had also better evaluate well.

As far as which program is better (Rybka verses Zappa) we don't really even have a clue from a ten game match. It will take 500 games before we even have an *inkling* which program is smarter. Right now, all we can say is that both programs play spectacular chess. That's saying a lot too. I mean, if you could cough up $50 and have Kasparov at your beck and call for ten years, that would be a pretty good bargain. Well, these programs play about that well.

IMO-YMMV

[Ryan Benitez]

The search verses eval debate will rage forever. It is just like the bitboard verses 0x88 debate (bitboard was the rage when Crafty did well at Jakarta. Then 0x88 took over and all the strongest programs used 0x88. Then Rybka came along (using bitboard) and the tide changed again.

I think (as far as eval verses search) Christophe said it well:
Eval and Search both produce knowledge.

If a chess program is to play really well, it had better search well. It had also better evaluate well.

As far as which program is better (Rybka verses Zappa) we don't really even have a clue from a ten game match. It will take 500 games before we even have an *inkling* which program is smarter. Right now, all we can say is that both programs play spectacular chess. That's saying a lot too. I mean, if you could cough up $50 and have Kasparov at your beck and call for ten years, that would be a pretty good bargain. Well, these programs play about that well.

IMO-YMMV

I don't see the Zappa/Rybka match as something intended to prove what "engine" is better anyway. It has many other factors that could not be reproduced. It is about how good the overall team is in the given conditions.

[Dann Corbit]

In a contest like this (where the engines are probably very evenly matched -- with Rybka being a bit stronger in general, but Zappa a bit more efficient with SMP and on a highly SMP system) I think that the dominating factor may be probability itself.

It is also possible that careful preparation lead to a couple of ambushes, but I suspect that it's simply a case of 'this penny came up heads.'