Rybka Search: a Guess

[fern]

I suppose all of you have read statements by Kasparov and many others top players saying more or less this very same thing: that they does not go soooooo far in the search, but simply know what is best after, say, 5-6 ply or, at most, in the case of combinations, 10-12 ply.
I dare to say that NO HUMAN being play chess calculating 18-25 ply as many top programs do. Not even Kasparov or only so in very rare endings, when one-branch trees are available.
Rybka, I believe, do the very same thing at least in most of the moves in middle game, where opening seeds begins to develope.
Rybka, believe it or not, is more a full width searcher than a selective one!!!
It is a witdh search engine capable of going deeper than old programs of that sort went, but not deeper than current selective machines.
And as such, having the full gamut of possible moves in the crucial first 5-6 ply of any position, it put into play a far more complex set of knowledge rules.
It is not a random fact that rybka climbed even higher once a Top master came to support the proramming. g
Rybka see no further than any other program, but see best the crucial first plys, even more, the crucial FIRST ply.
Haven 't you seen that Rybka rarely change his first move after only very few ply?
Chess is played move by move: if you can play always the very best move available, you could even disdain the need to examine more than ONE ply.
As far these days, that very best move can only be assesed after examining its consequences far ahead. That is what normal engines does.
But if by chance you have a tool to discover that without so much searching ahead, then...
I could bet my yearly earnings that Rybka incomplete PV hide not longer, but shorter lines than usual.

Waiting bets regards
Fernando

[Dr.Wael Deeb]

From an interview with Vasik:

Why is Rybka so much stronger than all the others?

My personal opinion is that Rybka's evaluation is better than that of her main competitors (and very different), while her search is of comparable quality (and somewhat different.

[GenoM]

Yes, Fernando, it looks like Rybka already knows which move is right in the position, so it has not search deeply ahead to play fine. It seems that we, non-programmers, think in similar way. May be wrong way, but I think Rybka knows the moves that leads to good positions, so it's mainly a matter of evaluation and not so much matter of search. Looking at the moves of Rybka I have the feeling that all its evaluation is about mobility -- the goal of every its move is to reach a maximum mobility for its own pieces and minimum for opponent's.
Looks easy, heh?
Best regards,
Geno

[Uri]

What is a selective machine?

[GenoM]

What is a selective machine?

Selective machine is a machine that takes a decision based on a selection. Why?

[Karlo Bala]

From an interview with Vasik:

Why is Rybka so much stronger than all the others?

My personal opinion is that Rybka's evaluation is better than that of her main competitors (and very different), while her search is of comparable quality (and somewhat different.

Let see for a moment what Vas like:
1. He likes fast optimized program
2. He likes statistics

There is no chance that Vas would sacrifice speed for some slow speculative knowledge. Maybe he add some simple stuff. I think that it is more likely that he uses some sort of material statistic for extension/reduction/pruning decisions. For example, if you are in branch where you are queen ahead and your opponent doesn't have enough material to make a threat then you don't want to make a NULL move, you don't want to reduce move, you just want to stop search. There is lot of room for improving reduction and/or pruning based on material. You can use variable margin and depth. You just need to play lot of games and count when and where your program make good and bad prediction of reducing/pruning

[fern]

With speed you refers to number of knods?
I suppose so and in that case just remember that speed can be used in different ways. You can use it to go deeper or you can use it to get a most exhaustive full-widht search with some more deep that the the traditional full-width searcher got in the past.
I do not believe in some cumbesome set of knowledge rules as a key to Rybka, but I do believe he is not just doing the same stuff of everybody, going deeper with the so called "fast searcher". I believe his is doing something else with the search AND that some specific set of clever rules -knowledege- are in use, not just an encyclopedia of theory.

My best
Fern