Be sure the chess knowledges added today into a chess engine are minimalistic.I analysed a lot of games of Naum 3 ( and not only )- not even a move to avoid the weak pawns center, a bad play knight vs. bishop, no profilaxis.......& so on. I strongly suspect that a fast search function is more convenient today for most of programmers.To add new chess knowledges into a chess engine is not a simple matter.........and ........with new chess knowledges the search becomes slowly, the horizon of search not so profoundly. Take a look please to KN/s & deepth of two engines with an amount of chess knowledges above medium : Hiarcs & Zappa Mexico.
We are now far away from the perfect chess .To build up-today- a chess engine including the whole chess theory and a very fast search function simple isn't possible.The tree is
maimed by various pruning functions by the shrine of the same search goddess and a lot of interesting continuations are premature killed. To build up the 32 men TBS also isn't possible today for a lot of reasons ( computing,hoarding......).And if the 32 men TBS will exist a day we need not a chess engine but a super search soft.
So, lovers of chess, be quiet for the next 100 years.The chess remains like a mystery & our toys engines will exist !
A similar game is taking place here v a strong CC player who is not using an engine. The engine he is playing is restricted to a maximum 1 hour search. http://hiarcs.net/forums/viewtopic.php?t=1009
I am thinking that maybe one could invent "non-human" evaluation to exploit some areas of human evaluation which could be inaccurate.
They're somewhere in between OTB and correspondence players. (a little closer to the OTB atm)
In the off chance that 7-man tablebases contribute substantially more than 6-man did and some pretty optimistic iterations of moore's law, we might have something that can do the job in 5-10 years. (i assume at this point the centaur relationship would shift up a medium and the engine would beat the centaur OTB, but lose to the centaur in correspondence (and beat the correspondence WC (unaided by an engine) ) )
The fact that top programs are so strong without some knowledge that humans have prove that knowledge is relatively not important in chess.
I do not expect chess to be solved mathematically in the next 20 years but a situation when almost every game between programs is drawn is clearly a possibility that I expect in the next 20 years.
Then I thought that "explaining chess mathematically" is not worth the effort :)
I certainly agree. Maybe, instead of accusing chess programs with having no understanding of positional play, humans should consider the fact that there could be some serious inaccuracies in the literature of strategical evaluation.
I'm certainly no expert on chess, but I'm beginning to think this is really the case. Probably humans have really underestimated the tactical complexity and tactical resourcefullness of chess a little bit too much. This could be why Tal (and probably even Topalov) (from the attacking side) and Petrosian (from the defending side) has reached so much success and has become so popular. I was always considering Petrosian as a positional and defensive player, and was surprised when Kramnik said his style was, on the contrary, very tactical; it was just that he saw more tactical defensive resources than any other player could. He was probably right.
If we take another step on this argument, probably a very risky one, this could be actually the reason why chess programs are not "positionally good," and why it is so difficult to implement knowledge in them: maybe they just reject it, like a body would reject new organs.
Larry Kaufmann had stated that when he tried to teach Rybka about Pawn Islands, he said the results got worse and he couldn't find a way to teach it without this being so. He then said that either Rybka "already had" this information in the compound of other knowledge she has, or that some of the assumptions about pawn islands may be simply wrong. I think these comments are of great importance. humans really got to step back and watch these "alien-like style" of super engines more closely.
> Maybe, instead of accusing chess programs with having no understanding of positional play, humans should consider the fact that there could be some serious inaccuracies in the literature of strategical evaluation.
I agree. A human can have a strategy that covers the next 20 halfmoves and looks very human, while the engine can have a strategy that covers the next 21 halfmoves, and is better, but looks strange and counterintuitive. So the engine programmer may want to teach the human strategy to the engine because he doesn't understand the computer strategy, and this would only have worse results.
But this problem will persist because humans can't think like computers. A new question arises when you consider that we want strong engines to analyze our games, but if we want to follow the computer strategy that we don't understand, making the engine's move is only going to hurt our game. A weaker move is better than a strong one, if we can understand the former.
Not strange,not counterintuitive.In chess a single halfmove make the difference ( you must attack - for example - a position with a forced mate in 5, but you see correct only 4 moves . The continuation ? ). The top engines can see deeply in the chess tree branches -who are in fog for the top chess players ( and not only 1 halfmove ) -despite the fact that sometimes they select not the best branch because the lack of knowledges.
To have a perfect chess evaluation function ( hard but not impossible )make possible a perfect chess at the horizon accessile by the search function of each generation of chess engines. The catastrophic problem is that a massive evaluation function retard the search . If possible a perfect evaluation function + a very fast search ( the challenge for future ) it will be near the perfect chess .That's all !
All the best,
You simply choose the best move based on evaluation.
Perfect evaluation simply tell you if a move is mate for one side and how many moves and also give score that is not mate for drawing moves.
Different evaluation is simply not perfect so it is practically impossible to have perfect evaluation.
You simply choose the best move based on evaluation."
Nope ! A perfect evaluation give us a perfect evaluation of a position according the chess theory. But if in tree they are a lot of near equal promising continuations ? Even the 32 men TBS will exist we need a SEARCH engine ( not a chess one ) to find the best continuation among an awe-inspiring numbers. The search make the horizon of a chess engine.
This is not correct.
There is no book that tell us how to give a score correctly to chess positions and humans do not know what is the optimal evaluation.
There are a lot of positional factors and people do not know what is the optimal score to give them.
The problem is not only about speed and you are not going to get perfect evaluation that you cannot improve by changing the formula even if you assume that speed is not a limitation and finding if a change is improvement or the opposite is an hard task espacially with complex evaluation.
If you have 100,000 numbers that you can change and if you need one year to get the optimal value of one of them then you may need 100,000 years to get the optimal value of all of them.
The practical problem is clearly harder because possible changes is not only changing weights and you may have a local maximum(situation when changing single number make the program worse but changing more than one single number make the program better).
This is not correct.
The practical problem is clearly harder because possible changes is not only changing weights and you may have a local maximum(situation when changing single number make the program worse but changing more than one single number make the program better). "
Here you are right ! But it isn't an insoluble problem.
e4 win in x moves
d4 draw in y moves
a3 loss in z moves
a4 loss ....
b3 loss ....
So we won't need a search. When we have it, you'll see that I'm right :)
(Uri, please check your pm box.)
Yes. The same is valid with openings, by the way.
I resoundingly agree with this, as well as the rest of your comments. The simple proof is how counterintuitive certain long endgame combinations can be. If this is true for a board with six pieces on it, imagine how bewildering a 20 or 25-man EGTB would be. There can be little doubt that there are countless won positions that require playing one move exclusively that appears to our feeble sensibilities to be a ridiculous and unaesthetic move.
Humans in general love jumping to conclusions and then holding them against (or with) all evidence. Apparently psychologists even have a test where they show patients random blots. Healthy patients will find (non-existent) patterns, while mentally ill patients will see the random blots for what they really are.
On a related topic, I've had a lot of fun watching people guess various things about Rybka. Let's just say that there are a lot of mentally healthy people in computer chess :)
It only needs more work.
Please don't forget the benefits of more chess knowledge put in Rybka.
You can be right that evolution of chess knowledge used in engines will be in an unexpected way.
After we finish implementing all classical chess knowledge in engines, unusual parameters benefiting will start to be found.
If we separate "what we know in chess" and "what we guess in chess", after a century long chess culture, we will see that the only real thing that we know about chess is tablebases. This is horrible :)
No, this fact prove only the inequality:
a strong search+ weak knowledges > a weak search + strong knowledges
The great number of drawns between the top chess engines prove only near the same strength in search , near the same chess knowledges and -please understand me in a correct way -near the same architecture.Nothing more !
> "I do not expect chess to be solved mathematically in the next 20 years but a situation when almost every game between programs is drawn is clearly a possibility that I expect in the next 20 years." ( Uri Blass )
Or i interpret it this way: In the next 20 years all engines will be too strong that they will all end as draw regardles of the difference in hardware used.
My point is that even the slowest engine will find the very strong move that is enough to draw against perfect play.
This process is already started in the 400/40 games of CEGT
Naum3 got draws in 80% of the games against Rybka.
Of course Rybka3 is probably going to be better but I expect Naum3 still to get at least 50% draws against Rybka3 at 400/40 time control and probably we are going to see better naum that is going to get more than 80% again against rybka3.
> Of course Rybka3 is probably going to be better but I expect Naum3 still to get at least 50% draws against Rybka3 at 400/40 time control
Maybe Rybka will need to change her playing style to enter positions that other engines have trouble understanding so she uses her strength more intelligently and avoids draws (I.E. entering positions that are very hard to draw for either side, knowing that it's an advantage to her.)
This trend shows nothing because in 7,8,9...........32 men TBS we have a lot of nodal points !
Maybe it means that knowledge is very important to human players and not as important to computer engines.
Search depth will never be enough.
So the knowledge will always be important for engines.
As the search depths increase, the knowledge will help less against weaker opponents.
But against stronger opponents, which can calculate very deep, knowledge will keep its importance.
Knowledge has an effect like "the engine calculates n moves deeper" than without knowledge. ( n depends of the quality of knowledge)
If the extra time caused by adding knowledge is less then the time needed to calculate n moves deeper, it is better to use that knowledge.
This is what happened in the tests used for R3s "more knowledge or more depth" desicion.
Result : importance is a relative thing :)
I’ll like to try another approach, suppose that we had infinite computer processing power, and with a very simplistic search method (very low HUMAN chess knowledge) we could find all paths from the start to the end of the game. So we could find a colossal tree of moves with all the possible moves that reach the end of the game. As all of you usually say, to solve chess. If we had access to that tree of moves it would be very indifferent to have HUMAN chess knowledge. This is simply because there would be only 3 groups of paths 1) To Win, 2) To Draw 3) To Lose. And of course everybody would try to choose any path from group 1 while it is possible. I say while it is possible, because if blacks select group 1 too, in such moment of the game the only paths possible to be selected would belong to group 2. And the game would be a draw.
But coming back to the real world… There is no such tree of all chess moves, so humans need to find a method or technique that let them avoiding unnecessary analysis of moves. And that is the moment where HUMAN chess knowledge or (HUMAN) chess theory comes up, as method to help their creators (humans) to find the best path in the unavailable tree of all chess moves.
The point here (and nothing new) is that chess theory (strategic and tactical) is nothing more than an invention (a model) from humans that is not simple to be verified. It took a lot of years to prove that some classical chess ideas were not correct, indeed, some of the Nimzowitch’s hypermodern ideas had some evolution during the last years (take a look to some comments and chapter from “Secrets of Modern Chess Strategy” by Watson).
Chess theory is nothing more (nothing LESS) than a tool created by humans to help them play better chess BETWEEN HUMANS, a tool which its real value added is very difficult to be proved (which of course does not mean that it can not be proved). So how could we be sure that all the existent HUMAN chess theory is correct? Is it possible? The only concrete way to prove it would be to know the tree of all chess moves and verify that using that theory (the one that we want to verify) lets find a path belonging to group 1 (or 2) with the addition that it have to be the shortest of all possible paths.
So… besides very basic and chess mathematical principles who can say that the “others” (block, prophylaxis, overprotection, etc) HUMAN chess principles are correct?
So...How could we be sure that is necessary to add HUMAN chess Knowledge to COMPUTER chess engines?
"So...How could we be sure that is necessary to add HUMAN chess Knowledge to COMPUTER chess engines? "
1.The men mastering a great amount of (human ) chess knowledges (masters,IMs,GMs ) perform better against chess engines;
2.the chess is a logic game and the human chess knowledges were estabished in a logic way serving the chess rules;
3. the search function-itself-is based on chess knowledges.When searching we (& chess engines-of course ) must to have always some principles.Two engines having exact the same search horizont are differentiated by the (human ) chess knowledges added.
4.Von Neuman architecture of our computers is a masterpiece of human knowledges....
5.We can see in 3,4,5,6 men TBS the correctness of human chess knowledges.
Let me correct one thing,
So...How could we be sure that is necessary to add all theoretically working existent HUMAN chess Knowledge to COMPUTER chess engines?
"1.The men mastering a great amount of (human ) chess knowledges (masters,IMs,GMs ) perform better against chess engines;"
Is it still true to say that? If now a days the best human chess player of the world would play a match without odds at (we cay say) 4 hours or 6 hours time controls with one of the best chess engines of the world (Rybka, Zappa, Naum, Shredder...) which be using a modest opening book and 5 men tb... Will the human perform better? I'm really not sure about that.
INMHO now a days any 2700 elo chess engine perform better than a human, but we must see it from an open perspective, I mean considering overall results, I'm not sure we can judge a chess engine which wins 90% (or more) of the games against a human to be worse than the human because the engine does not apply some HUMAN chess principles or even more because the chess engine does not understand yet some specific positions.
"2.the chess is a logic game and the human chess knowledges were estabished in a logic way serving the chess rules;"
It could be ok, but...
"3. the search function-itself-is based on chess knowledges.When searching we (& chess engines-of course ) must to have always some principles.Two engines having exact the same search horizont are differentiated by the (human ) chess knowledges added."
I agree, but the point here (and the base of this thread) is how much human chess knowledge is currently included in chess engines. We are saying that now a days the human chess knowledge implemented in chess engines is not so big, and the fact is that without "all" that knowledge chess engines are (at least starting) to perform better against humans.
"4.Von Neuman architecture of our computers is a masterpiece of human knowledges....
5.We can see in 3,4,5,6 men TBS the correctness of human chess knowledges."
Could be ok, but I'm not sure they let to prove something practically.
"........and the fact is that without "all" that knowledge chess engines are (at least starting) to perform better against humans. "
Of course YES !
In chess a single halfmove make the difference . The chess engines can see deeply in the chess tree branches -who are in fog for a (top ) human chess player -despite the fact that sometimes the branch selected by engine isn't the best- because the lack of knowledges.
But in equal search horizons the chess knowledge becomes essential.Here is the problem: how to add new chess knowledges without any loss in search horizon.
All the best,
And nowadays, it works :)
Powered by mwForum 2.27.4 © 1999-2012 Markus Wichitill