Chessworld.net Instructive Kasparov Game: Add spice because two eval numbers!

[AKARAKAkeke]

Nxb5 according to houdini 1.5a 22 0.00 775.9M 8:41.74 Nxb5 Nxb5 cxb5 d5 Be2 c4 a4 Qb6+ Kh1 a6 bxa6 Rxa6 Ra3 Qd4 Qc2 Qe4 Qd1 Qd4 Qc2

so... maybe there is still something to understand what is an accurate move in chess :)

[zonamamuta]

Or, it could be even easier to implement, even now. Just count the difference in evaulation points between the first and the second computer suggestion, because the most difficult to play positions are these, where there is only one good option and you have to be extremely accurate. So if the first choice is +5 and the second +1 and the third -1 than you have a difficult position.

[Nimzomyth68]

I don't get it. How can a computer give a human evaluation. Do you mean that it would just have human made evaluations in its memory?

[whoslayedgoliath]

fritz 12 in analysis mode gives the "hotness" of a position, or its tactical trickiness

[mrKreuzfeld]

if computers could make human evaluation numbers, then they would not need brute force, and could calculate "perfect" game

[TheHuggybeary]

What about IDEA in Aquarium? I believe someone brought this up on the Rybka Forum when you posted. Oh course that is human inaction. But with a computer its a lot harder. That's why when you use IDEA people tend to make a lot of different moves to see what kind of Eval you get. I have an expert IDEA user on the Rybka Forum who will look this over about the b5 move to see what kind of score is given if you are interested?

[bryanep]

An amazing game.

[davegski]

If you know Tal's anecdote, where while analyzing a sacrifice he suddenly began thinking of how to drag a hippo out of a marsh, I think what you're looking for is a computer that can make decisions based on similar rationale. As I interpret it, Tal was basically weighing risk/reward (and the pleasure of playing the sacrifice) by thinking about something totally erroneous. Computers will never be able to produce risk/reward assessments based on psychological factors - they aren't brilliant.

[you1w]

I play chess for a hobby, but have played in 3 small local tournaments winning 1 of them, so I think I do know a little of something about chess but I'm not deep into it yet.

My thing is, I don't understand what you mean when you talk about "engines". What does being a "king's engine" player mean?

[jcandnp]

great point about having two eval numbers in the engines...i think that's exactly what we'll see from the best engines in 5 or 10 years. let's hope!

[jonsploder]

I don't agree with the human evaluation function.

The closest way we could currently write such a feature would be to write an artificial neural network, and after training it on grandmaster games (which are very similar to your first engine function now, so it would have to be taking games from positional masters rather than tactical masters), have a team of strong humans to manually correct thousands of positions to give this function an credibility. Why all this for a weaker function?

[kingscrusher]

@jonsploder Why would it have to be written like that?! Why not a count of the number of moves that retain the advantage in a certain position as a measure of "easy to play" - and when for example there is only 1 move in 50 - which perhaps isn't a forced recapture etc, then that perhaps could be identified as "difficult to play".

[jonsploder]

@kingscrusher But a computer doesn't understand that, and it's not really possible to do that anyway in a computer. Even if we did program that example in, if a delayed recapture was identified as 'hard' for a human to see, ALL similar delayed recaptures would appear to be hard for humans to the machine. That's not the case, some of them are obvious and others are not.

Why I think we'd need an ANN is they go through 'training' similarly to humans. They also make human like mistakes.

[44percentburnt]

@kingscrusher exactly what I was thinking, easy enough to implement.

[solinero]

great video, i really enjoy watching it!

[mcchang1032]

I like that that concept: easy to play for a human: doesn't rely on playing the most accurate move for 10 turns in a row. Any thoughts of treating the rise of engines in the Evolution of Chess series?

[kingscrusher]

@mcchang1032 Maybe they are influencing players preparation more and amazing tactical ideas in openings and middlegame. We are in the age now of veery convenient post mortems, and being able to see more chess in a scientific way - yeah I should talk about it in the evolution of style series.

[GavGuitar]

are you playing in the 4NCL?

[TheMomemtum]

A very good idea with the evaluation numbers. And it would be quite easy to realize it as well! The main difference between "hard" and "easy" positions is that in easy positions even alternative moves are good, whereas in hard positions only the top ones are. So "human evaluation" should be 2nd-4th ranked moves in average, or something like that.

[kingscrusher]

@TheMomemtum Yes I think it should be possible to program. The current shift in chess products is for particular applications - e.g. Chessbase 11 introduces a "prepare for opponent" rather than just remaining as a more theoretical game database. If engines are to be less theoretical, then they should aim to benefit the actual results of players at varying levels - and not just offer an evaluation based on the assumption that humans can play perfectly precisely engineered moves all the time.

[kingscrusher]

@kingscrusher This subject actually has massive implications for raising the profile of gambits and ideas which generate piece activity at the expense of material. If engines could actualy in effect start officially supporting ideas which actually are creative, dynamic, and often material losing, then perhaps also this will lead to even more exciting OTB preperation, openings, etc.

[jaileroflove]

@TheMomemtum the tough thing is that sometimes only the best move is good, but also only the best move is natural to begin with. this is trivially true for stuff like, "queen takes queen, such-and-such takes back," but can also be true for getting out of a sticky situation, or landing a killing blow or whatever. i'm not sure what principles would suffice...

[Mrfoolsmate]

The problem with path B is that when you get a job in the given field you have studied, you end up lost when what you have learnt is put to the test in various ways and you end up back on path A anyway!!

[kingscrusher]

@Mrfoolsmate Yes but if you don't get your qualifications, it might be harder to get the job in the first place. In chess, if you do get the results, you can also enter stronger tournaments, etc, and aim to play more accurately later too at higher levels, or correspondence style chess, where errors are far less in theory. 

[kingscrusher]

@kingscrusher I enjoy playing gambits myself though and like the power of the pieces - so of course I'm biased towards Kasparov's dynamic approach. It did nearly backfire when he is nearly wiped out by Karpov in 1994, when Kasparov was perhaps taking some risks with the dynamic Tarrasch defence.

[kingscrusher]

@kingscrusher That said a lot of Carlsens games also show an interest in piece activity and often giving up say a pawn "unsoundly" but getting very good pieces. Putting the opponent under intense positional pressure often is a way to crush opponents.

[Mrfoolsmate]

@kingscrusher Do you prepare lines for particular opponents in tournaments yourself? Being only 166 ecf myself, it is impossible to prepare for opponents or play against someones style because obviously they have no games in the databases.

[Tarasamanator]

i'm sure a human evaluation number type thing is possible, the engine could look at moves later other than the most accurate and if these moves make the position worse, it would go for a variation that might not give such a big advantage, but there all multiple moves to play in them, accounting for the fact that humans can't play 100% accurate

[kingscrusher]

@Tarasamanator Well yes, for example some variations in the French defence winawer, where black gives up the g pawn for dynamic compensaiton. If White can squash this compensaiton, then White could end up with a dangerous passed h pawn and later win. The beauty of error-prone OTB chess against human opponents is that errors in games can be completely assumed at all levels - even GM games!.

[peterfdrucker]

I don't understand your idea of two different evaluations. Are you suggesting that some moves are more likely than others to prompt your opponent to make a mistake? The way you discussed this game, it seemed that Kasparov's win, or at least advantage was dependent on Christiansen's mistakes. This just means that the worse your opponent, the more free you are to choose a move that is rated as suboptimal by the engine. It can't mean there's a better move for you, no matter the opponent, right?

[kingscrusher]

@peterfdrucker What do you mean by "better move" ? A move which you need to play very accurately like winning a pawn might be rated high by an engine, but an engine can face the complications with great ease because they calculate by looking at millions of positions per second.

[kingscrusher]

@kingscrusher Here is an analogy - you have three weeks to prepare for your degree level physics exam. In theory you should really read the entire textbook and understand it. In practice, you decide instead to just check past exam papers and make sure you can answer the last three years sets of questions because it's the same lecturer. Now path B leaves you passing the exam - path a although theoretically best would require too much time and effort.

[kingscrusher]

@kingscrusher Question here: Is path B (just checking past exam questions) independent of lecturer?! Is it also independent of those particular subject areas?! Probably not. Will it pass the exam? Most likely. In chess, you can often play suboptimal practical chess which give winning chances. In the Magnus Carlsen lecture at Gibraltar, a Norway GM was demonstrating a Carlsen game where he was like -4 or something but kept his pieces active and chances alive,. He later won against his opponent.

[howardbeasley]

@kingscrusher the problem with your theory here is that not only does it not take human-human play into account, as you mention, it's also got a horizon problem...i've been letting rybka run for some time on the position after the continuation i gave below (since i viewed the video, in fact) & the evaluation has dropped steadily down to +.17. other more dynamic (if weaker overall) engines like junior (which i highly recommend) actually prefer kasparov's b5 at first blush. (cont)

[howardbeasley]

@howardbeasley in essence it's a human move that computers can't see the benefit of without algorithims either pushing it towards such active play or without intense calculation. while it -may- not be best overall move, i don't believe it's nearly as inferior, objectively, as you seem to be implying.

[kingscrusher]

@howardbeasley Kasparov's move is good for OTB play against a human opponent - as with all Gambits - they have more success than perhaps theoretically they should from an engine perspective.

[kingscrusher]

@kingscrusher But yeah sometimes it does turn out that if engines are left for longer, they start seeing the dynamic compensation and start giving more credit to those dynamic-creation human moves.

[44percentburnt]

KC, I totally agree with you about the 2 evaluation numbers, when analysing with rybka I don't take his evaluation too seriously because indeed some lines are just too mind-boggling to play correct for a human.

[Nola213]

Very nice game. I like "pressing the Bangko gambit button", heh.

But the idea of a computer engine giving two sets of candidate moves. The strongest computer moves, which sometimes arent that great when a human plays it, as you said cause we cannot look so many moves and tactics ahead.

So yes, a Best "Computer" and a best "Human" move would be ideal. But as you stated we're not quite there yet; and sure perhaps the moves for Human /Comp. would be the same.

It's an interesting idea

[Miir164]

Instinctively, 8. Bxe6 appeals to me.

Why don't you nor Kasparov prefer it ?

[kingscrusher]

@Miir164 Kasparov played Bxe6 a few years earlier in the other game he faced in this variation. 

[kingscrusher]

Hi all, I put a link in this videos description to the page which has many of my other Kasparov videos. So if you like this one, please check out that page too.

[IMPaulyn]

I agree - if the computer analysis engine took a minute to look through my database, it would see I'm normally at a minus rating.

[pawnscrusher]

wow, very nice commentary. undermining and invading on the light squares e4 and d5 was the plan, when playing the b5-pawn-sac "killed" the black centre and allowed some destructive queen moves on some of the light squares. also the blocking of whites attack with his own f5 (blocks pawn and rook) pawn shows such mastery of dominating a colour of squares. But I think I would have played Nxb5 as white to challenge the strong Knight on b5 ^^ I know you think its boring tho..

[howardbeasley]

@pawnscrusher even if black forces it off himself with 15 Nxb5 Nxb5 16 cxb5 d5 black gets a very menacing pawn center as well as a potential new outpost at e4 for the pawn & white has two very weak pawns. i think its possible rybka is too materialistic in this case. at any rate it is a good example of a what kingsrusher is talking about on a much higher level -- 1 computer score, & another tailored for kasparov -- i would take him in the above position any day.

[MagikoZari]

WOW!!!

[luckystrke]

crazy game! i love kasparov.. thx for a great video!