I know that most people think that computers have overtaken humans and the tide will never shift again, but I'm not convinced. Computers have been increasing in strength exponentially but they're going to hit serious diminishing returns soon if they haven't already. On the other hand, humans as a whole are getting better at a more sustainable pace. People seem to forget that, e.g., Rybka is stronger than the last generation of human GMs. Carlsen, Radjabov, Nakamura, and countless other young talents have grown up with super-strong engines at their disposal. The coming generation of GMs is going to shock the world and play better chess than the strongest engines.

If you watch Nakamura on ICC (handle Smallville), you already know that he can draw the strongest engines on there virtually at will in 3 0 (granted, with White, but still: 3 0!). If Naka were allowed to look through as many games played by an engine as its programmers were allowed to feed the engine games of his, and they played a very long match at a super-long time control (something like 40/3, 30/2, G1), Naka could just draw every game until the computer made a mistake and then play cautiously for a win.

gg

I would take the computer side for just about any amount. Computers still have a very large potential for improvement. The method they use to evaluate positions will continue to improve and obv their calculation power is unmatched.

The problem is that software is not improving at nearly the same rate that hardware is improving, and computers are not going to be calculating another ply in every variation any time soon because the combinatorial function grows faster than the exponential function. What makes you think that the engines' evaluations are just going to "keep getting better"? Will they just get better by magic? No, positional evaluation is the true AI component of chess engines, and it's much more complex than you think.

You also severely underestimate the significance of the fact that the current generation of young GMs grew up with super-strong engines. Kasparov at his peak could probably score 40% against Rybka and there were no super-GM strength engines around when his brain was still forming new neural pathways.

Another factor that you don't want to overlook is the fact that the new generation of GMs already have ten times the experience of the previous generation due to the internet. These kids have played hundreds of thousands of games already.

What about the fact that Nakamura has no problem drawing the best engines in the world in speed chess?

I think engines are pretty much irrelevant. Chess is about ideas and understanding. A computer is about calculation. Humans are flawed in calculation but have great understanding of chess. A computer has zero understanding of chess, but incredible calculation. The man-machine thing has just become passe. Engines are excellent blunder checkers and their short term calculation is enough to consistently beat most of all chess players. It doesn't really mean anything.

A somewhat interesting experiment might be to give a human access to some sort of an engine while he faces off against another engine. I think it would be a slaughter in mankind's favor, but maybe not? And again, I don't really think it's relevant either way.

I don't think drawing with white, against anybody, is an accomplishment for a 2700 player.

Maybe not just yet it isn't. If Carlsen, Nakamura etc can routinely draw with white against the top chess engines 5 years from now, I'll be surprised.

You think that the engines are going to improve more in the next five years than Carlsen is? Are you serious? You do realize that the engines aren't going to be looking any deeper in all variations in five years than they are now, right?

I think that drawing 100% of the time (even with White) against the best engines in the world in a blitz game is a huge accomplishment.

It doesn't bother you that the best humans aren't as good as the best engines, or that the general public mistakenly believes that chess is solved?

If Carlsen were to win a match against the best engine in the world there would be a huge resurgence of interest in the game. People don't like to try to get good at things at which no one in the world is as good as their PDA.

Oh, I'm sure positional evaluation is amazingly complex. In the last year I had a chance to speak with an independent programmer who believes he has developed a significant improvement in the way that chess engines calculate and evaluate positions. Maybe he was full of it, but he claimed the Rybka people had purchased his engine and that it was being tested on ICC. This is secondhand, so maybe completely unreliable. I just think that even if he was full of it, it's unreasonable to believe that the current model for chess engine calculation/evaluation is the best we will see. If it gets even marginally better, computers will be out of reach.

I'm a hobbyist programmer and have written game-playing engines, and I can tell you that it would take a fundamental shift in how chess engines evaluate positions to put them "out of reach".

Serious question: do you know anything about AI or programming, or is your claim that a "marginal" improvement in engines' evaluation functions will put them "out of reach" just snatched out of the air? No offense, but you seem to have admitted to not knowing much about it only to later come to a pretty strong conclusion.

To put it another way:

In order to evaluate even a single position, an engine has to look at tens of millions of other positions. In contrast, a GM only has to look at the current position and a few possible lines--perhaps as few as ten other positions. So if we're talking about strict evaluation, GMs are currently about a million times better at evaluating positions than engines are. The engines make up for this deficit with raw calculation. So a "marginal improvement" in how the engines evaluate positions isn't going to make a difference. Now, if the engines got, say, ten or a hundred times better at evaluation, they would become out of reach. But in computer science, finding an algorithm that is ten times better than another algorithm is a monumental task. I would be surprised if computers are even twice as good at evaluation five years from now.

I think computer are going to be looking much, much deeper five years from now than they are now. It's also likely they'll have much better pruning of the trees they look down, which will only enhance the depth of the variations they do examine.

That has to help a lot even if the computer's ability to do a static analysis of a position doesn't improve.

No, I know very little about programming. Maybe that makes that statement invalid. I don't know. I do know that the results show humans barely hanging in the matches vs machines and that any increase in strength at all will put machines out of reach. BTW, since you get to wonder about machine strength, what makes you so sure that Carlsen will become so amazing? Sure he is a phenom and has the potential to become one of the best ever, but I don't see any reason to believe he will ever be stronger than Kasparov. I guess the question is what evidence do you have that humans have not peaked wrt chess strength? (obv meaning the best players, not inflation or average strength)

Edit- strated typing before I saw your second post. I tried to get this guy to explain the program to me and he just talked circles around me about on-the-fly tablebases of likely positions and some way of pruning. It made zero sense to me. Would a marginal or moderate increase in a program's ability to prune dramatically increase its strength?

Almost everyone tries to understand, but sometimes humans are "out of
their depth": e.g., positions in five- and six-man endings in databases have
"difficult to find" optimal moves. You're correct that the very best human
players have a good understanding of the game and "bots" only play primarily
by calculating an optimized evaluation function for positions "out of book".
"Bots" misevaluate positions where one side has a material disadvantage but
a clear (at least to "strong" humans) winning plan.

Time controls are an enormous factor. If a "world class" human has much
more time and can use any resources whatsoever and games were played at
a rate of approximately one move a day, IMHO any "bot" could be
"slaughtered" ( maybe the human could be about +2=8 out of ten ). OTOH,
if a "dedicated computer" has almost the entire database of a human
opponent's games, together with a team of humans (including strong GMs)
searching for any and all types of "mistakes" made in advance (not to add
fuel for conspiracy theorists in IBM versus Kasparov), the "world class"
human could get "slaughtered". At "rapid" time controls, it's extremely
unlikely that a future human player could play at the same level as the "best
bot". It's all not really relevant.

One major point is that humans can see "beautiful ideas" in chess, not just
in actual OTB competition, but in chess composition; a "bot" or program is
brutely calculating, oblivious to the remarkable beauty that humans can
truly appreciate.

A fundamental shift in how chess engines evaluate positions is a lot more likely than a fundamental shift in how human chess players think about chess. I'm a hobbyist programmer as well and I wrote a chess engine (albeit a rather weak one) as a computer science undergrad, so I have some familiarity with the subject.

There are certainly alternatives to the traditional methods of static position evaluation. For example, and maybe this has already been done with Rybka, you could evaluate a position by playing out several simulated games to their conclusion (using a very crude, very fast algorithm to make the moves for each side). The evaluation would be the percentage of wins, draws, and losses achieved in simulation (obviously there would need to be a stochastic element involved in selecting the moves for the simulation, otherwise each simulated game would come out exactly the same). Now that might not be any better than the traditional static evaluation in practice, but its another avenue to try, and might find deep positional factors that would tend to be missed by a static evaluation. My point being, a fundamental shift in position evaluation doesn't seem all that unlikely. Where minimax with alpha-beta is pretty clearly the right way to go for search, refinements and optimizations can be added incrementally but that paradigm isn't likely to change, the same can't be said for the position evaluation side of things.

No, they're not. They're not going to be looking any deeper in all variations than they are now. If you don't believe me, google "combinatorial explosion", "exponential growth", "combinatorial growth", "moore's law", "horizon effect", etc.

Possible. I wouldn't say "likely", but possible. However, pruning is a much more difficult problem than searching. You can't just blindly expect pruning to get significantly better.

Pruning can't significantly improve unless evaluation significantly improves, unless I'm unaware of some pruning technique that isn't logically dependent on evaluation. From a common-sense standpoint, and with my "programmer's intuition", I would say that since you have to at least do a cursory evaluation of a position before you decide not to pursue lines beginning with that position, you need evaluation to prune.

That's a non sequitur. If the best humans can currently score, say, 35% against the best engines, then the engines could improve marginally and the humans' winning percentage expectation could drop to 33%. If 35% isn't out of reach, 33% isn't out of reach. Obviously this is a problem of the "how bald is bald" sort, but you get my point.

It's generally accepted that each world champion has been objectively stronger than the previous world champion. They have grown up in a more advanced world with better nutrition, better education, less hardship, etc.; they have had access to more sophisticated training methods; they have benefited from advances in opening and endgame theory; and, perhaps most importantly, they have been able to learn from the past world champion whereas the past world champion didn't have the chance to learn from them.

The current new crop of GMs has grown up in a world with super-strong engines for analysis, enormous searchable databases of GM games, strong opposition at their fingertips (via the internet), huge repositories of exercises, etc., etc. I think it would be utter folly not to expect that Carlsen (or whoever emerges as the strongest of the current generation) will not be significantly stronger than Kasparov was.

I couldn't disagree more. Every generation of human chess players has virtually re-invented the game.

That's a good point. So a lot depends on static analysis.

Well, I mean, you can prune based on preliminary analysis, which can cut away a lot of useless branches from the game tree, but this kind of simple pruning is almost certainly already implemented optimally or near-optimally.

I don't think the general public is aware of much of anything regarding man-machine chess except perhaps that Kasparov played some computer a while back, and ended up on a superbowl commercial parody afterwords. I am positive the general public has no clue who Carlsen is.

Remember, it was only a couple of years back that Kramnik 'defended' humanity against Fritz drawing the match after starting out 2.5/3 against it. Of course the score doesn't tell the half of it. The computer would have been brutalized if not for mid-match human opening assistance and of course the famous brain farts where Kramnik simply blundered into mate in 1 in a drawn position. Or him resigning another drawn position. But all those interesting anecdotes aside, nobody cared! It changed nothing besides verifying the complete lack of interest in man-machine chess!

If anything the match just again verified what I've been saying. Computers are terrible at chess. Short sighted blunders and oversights are just a part of chess. Completely getting rid of them is a literally inhuman task. That it's been accomplished by a computer resulting in their ability to perform at the highest levels means absolutely nothing. It's analogous to the fact that a baseball pitching machine can make Nolan Ryan's fastest fastball look limp. Who cares?

I was completely unaware of this. I'm so sure this is correct...

http://en.wikipedia.org/wiki/Compari...ughout_history

Well there's always a balance. The more pruning, the more depth/less accuracy. The less pruning, the less depth/more accuracy. The thing is I think there's a good chance this balance has been near perfected by now. It's something you can test pretty easily just by using position test suites, and it's a balance problem that's been around since the first chess engine.

How can computers be "terrible at chess", while at the same time being able to achieve a plus score against the top GMs? Unless you're also saying that humans are even worse at chess.

Granted, there are positions where the top GMs still tend to find better moves than the top chess engines, where conceptual understanding outweighs brute force. If you define chess as the set of such positions, I'd agree that humans are better at chess. But historically that set of positions has been shrinking, and I see no reason to believe that it won't continue to shrink.

That article talks about domination over peers, which is how competitors in any game or sport are usually compared--because it's a foregone conclusion that the modern players are objectively better than the players from back in the day.

That's not a direct correlation, but there's a definite positive correlation between time and absolute playing strength, at least according to whoever compiled that list.

Seriously, Kasparov was 10x stronger than Lasker. Lasker was 10x stronger than Morphy.

This is not even a bet, there is absolutely zero chance Carlsen will be better than the strongest engine in like 5 years.

this isn't "generally accepted" at all. Do you really think Euwe was objectively better than Alekhine? Petrosian better than Botvinnik? Karpov better than Fischer? Kramnik better than Kasparov? I don't, and I don't think most chess historians would.

Of course they were all extremely strong players and deserving world champions, and all but one beat his predecessor fair and square in a match. But I think all you can really say about relative strength is that the strength of the top players has tended to get consistently stronger.