Deep Fritz Rybka amp future

  

Back to forum


Thibault de Vassal    (2006-12-01)
Deep Fritz, Rybka & future

The Chess Challenge 2006 in Bonn between classical world champion Vladimir Kramnik and chess engine Deep Fritz 10 confirms (who ignored ?) the best chess programs can rivalize with the world champion in a match, but it first shows us these calculating monsters still have weaknesses.

Question is : What are the real improvements in Fritz 10 compared to Fritz 9 (engine speaking only) ?

Here is what I think about chess engines nowadays (Fritz 10, Shredder, 10, Junior 10, Hiarcs 10 and particularly Rybka 2.2) :

The way of think to play correspondence chess is (or should be) mostly human one combined with a chess engine algorithm. We follow the tree of moves like a program with our selective algorithm (much better than chess engines), applying our judgement of the position when necessary only. The point is we evaluate moves and we almost never evaluate a position twice.

Chess engines are very good analysis tools but are surprisingly not designed to be very good chess players. I think a major improvement in chess engines should be recognition of 'sufficient moves' : ie. it is no worth to always find the best move at a particular point of the tree, this reflection time could be used later... It depends on the evaluation of the position, on the clocks... Iterative model is quite basic (in a game at least !).

Another point is recognition of traps. This is the start of psychology in chess engines, and basics of the art of war. It first depends on who your opponent is, and on the clocks too. Finally, at the end of the tree, chess engines evaluate positions, but how many evaluate moves ? .. Speculative moves were a step, but it first shew chess engines were not able yet to see what move is worth to be analysed really deeper, consequently creating a 'human' weakness, particularly against some other chess engines.

I don't know how Rybka works, but as far as I read about this one that calculates much less positions (about 10 times) than Fritz, I wouldn't be surprised that Vasik Rajlich had implemented a better approach of human way of think, which is undoubtly the future of chess engines.

A good 'centaur' in ie. Playchess rapid tournaments is first a good choice between Chessbase engines according to the position and clocks. Fritz qualities probably apply best in standard games, where clocks are really designed for him. Among Chessbase engines, Hiarcs is probably the best Blitz player and could be the best correspondence chess player (even if it isn't the best CC tool for humans). Rybka is probably a kind of centaur itself (sorry, herself ;)), knowing when to use (in the tree !) brute force and more selective approachs - not to be compared to Hydra or Deep Blue which, on contrary, use most brute force.

My conclusion is chess engines have much to learn from humans yet, we'll see a Rybka 5 and Fritz 13, with much better results against other chess engines, but their results shouldn't increase a lot against the best humans in future. Finally, it will never be a good correspondence chess player :)

My two cents.


If I find time, I'll continue to implement my own chess engine..... but it's a lot of work :/


Elmer Valderrama    (2006-12-01 21:07:50)
intuition

I haven't got even Fritz 9, so can't answer the question..but regarding engine development, I believe they lack any chess intuition, it would be a major breakthrough to implement one

...although apparently brute force at 40-50 plies = intuition :(


Lionel Vidal    (2006-12-01 21:42:25)
Intuition?! what for?

Don't you think intuition in any abstract game is in fine just a nice word to hide our (that is human) limitation in analytical power?
In many very good chess books (see for instance Watson opus), intuition is indeed shown as not an adequate compensation for a good, reliable, concrete analysis. Of course, for us humans, it is still very useful because the experience of already seen patterns may suggest the very best move in a given position, without even any calculation... but if you had the power to make a complete analysis, would you still use your intuition?

My feeling (and I am not very happy with that, but I don't see any evidence to contradict it) is that in 98% of positions, the brute-force stupid way of computers is already deep enough in the tree of possibilities to find the very best move (at least in any practical sense)... and the 2% left is only interresting for correspondence players... and then, only for the very best who can claim enough expertise, or enough time :-)

Now I am sure chess can still be fun: the old and only true chess way has just been re-edited: "tempête sur l'échiquier" (sorry I don't know the name of the english version)... at least I feel competitive enough :-))


Thibault de Vassal    (2006-12-01 22:29:35)
Intuition?! what for?

The problem is a "complete analysis" is not possible, in most cases... So intuition (and psychology) - in these 2% of moves - will always have a small place, in a small gap of... a few hundred elo points ! :)


Wayne Lowrance    (2006-12-02 00:23:07)
Intuition?! What for ?

I do not understand " a few hundred elo points" ? That is huge. I think I mis understand your point. Wayne


Thibault de Vassal    (2006-12-02 02:12:06)
Intuition?! What for ?

That is huge, indeed. That's why computers alone couldn't play correspondence chess at a high level... These 2% of moves are enough to beat them, at least to score 3 out of 4.


Lionel Vidal    (2006-12-02 08:55:46)
Intuition?! What for ?

3 out of 4? Really? Which test-matches are you refering to?
Leotard made a test (won :-) and with grand manner) but that was years ago, and besides, he is one of the very best :-)
Then there is the match against a panel of different engines by Ham: even if he does not play at the same level than Leotard, he is quite a good player!... and the results were very far from 3 to 4 for human :-( (that was also years ago!) Then there are the hydra matches... :-( the results are not very good also for humans and the game comments are very instructive: against first class expertise chess knowledge and intuition, the 'dump' brute force machine managed to handle quite well complex ending positions...at least as well as all correspondence players but the very best (I would say the top 20 at most :-()
Maybe there is some recent test I am not aware of?

But the point really is: who can play like, say, Leotard? Of course, he says he can crush computers, just by playing them like 2500 rated players... well, I can believe that... but when I play a fritz-push-button opponent, I am only a 2300 player... have I to use also an engine to have a chance (and one game out of ten, be very proud to have chosen another move than one of the few the engine suggested as best and still not have lost... ok, just kidding :-)?
It can be still fun, but I think it is not the same kind of chess Leotard alludes to when speaking of himself in his after match interview :-)


Lionel Vidal    (2006-12-02 09:54:32)
A lone engine in CC :-)

Suppose I make the following test (it has certainily be proposed before, but let's do it again, for the fun of the argument):
- I buy a recent engine (say the new Fritz10)
- I play in some CC tournaments (I do not want to pay fees, so let's say, here at FICGS of course :-), and at iecg)
- I choose the first moves of all my games based on some statistics made on a CC base (just to avoid some openings statistically bad in CC)
- starting from a few moves before the engine goes out of its opening book (to be defined, maybe 4 moves) I let my average computer run 10 hours by move (around one night per move... I know, I sleep too much :-)
- I *always* play the very move the engine finds as best
- I play as many tournaments as I can, considering the time constraint that limits the number of games (just to get a meaningful rating as fast as possible)

Now, what rating do you think I can reach at most, strictly following these guidelines?
(note that if I know some basic maths to do the stats, I do not even have to know chess rules... although a basic knowledge is assumed to ease the play in practice)
Are you ready to bet on your guess ? :-)

In pratice, the test does not work, because the tester dies from boredom long before he gets any rating :-))


Thibault de Vassal    (2006-12-02 13:40:46)
Chess engines CC ratings

It is my estimation. 3 out of 4 represents about 200 elo points. I doubt Chessbase will organize a Man(+Machine) vs. Machine correspondence chess match... However there are a few examples : Arno Nickel - Hydra (2,5-0,5), Hydra which beat Adams over the board 5,5-0,5 ... And I suppose Arno Nickel did not have access to the program, but knowing better his opponent I'm sure it's possible to reach such a score against any program.

About your test, it's been discussed here already :) .. In my opinion such a player's rating would travel between 2200 and 2400 (at most) mark !

Waiting for a match against Rykba :)


Wayne Lowrance    (2006-12-05 03:26:24)
Chess Engines CC Ratings

Waiting for a match against Rybka ? :) hummmm Wayne