Why the Chess Computer Deep Blue Played Like a Human | Hacker News Reader

Why the Chess Computer Deep Blue Played Like a Human

nautil.us

61 pointsdnetesn28 comments

The interesting thing about the Monte Carlo approach is that it depends on the property that a randomly chosen move (or branch) can be representative of a similar set of moves. In a game like Go this is true because there are many sequences of moves that end up in exactly the same game state, and slight variations of board position have the same utility. But in chess, slight variations in every board position will have huge impact on the outcome of the came. Raw calculation is a necessary component of strong play and intuition is much less useful.

Chess engines that use Monte Carlo don't make moves completely at random, they use a small search depth when playing through the games to ensure that fairly good moves are mode.

A short writeup here of it in practice (images broken unfortunately) here:

http://en.chessbase.com/post/rybka-s-monte-carlo-analysis

Also worth noting that IBM stock price increased by $2Billion after defeating Kasparov. So they have had everything to gain by beating him at whatever cost.

Stock prices going up have nothing to do with the incentives of companies and everything to do with the availability of information.

It'd be like a gumball machine trying to make itself more valuable. Just sell more gumballs at a hirer price.

The incentive of public companies is everything to do with stockprice, as it is the primary metric to rate the performance of a company? What does information availability have anything to do with the value of a company?

It is fascinating to compare & contrast chess, Go, and backgammon, and how quite different approaches are needed to program in each. For chess, 'pure calculation' is most effective, while the author suggests MC is good for Go & others. In backgammon, TD-gammon used a rather naive neural net approach and trained itself, optimizing its own play over time! It is also interesting how computers' success in these areas feeds back into how humans play the game (at least in backgammon and chess opening strategy).

It's been remarked that backgammon is man's struggle against change, chess is his struggle against another man, and Go is his struggle against himself. While there's some artistic license (a good backgammon player will crush a mediocre one in the long run) this comment does provide some interesting cultural.

And there's also a Shibumi quote I like: "Go is to Western chess what philosophy is to double-entry accounting." This book was the first time I heard about Go.

I've always been curious if what Kasparov calls, "human intuition" by Deep Blue was merely Deep Blue being able to search more plies than any computer could before it.

Also, not sure if it was fair for IBM engineers to change Deep Blue code during the actual match. Because now you are actually playing against both humans and machines. It would have been a huge blow for IBM if the machine they've spent to long working on couldn't beat Kasparov.

I think IBM did several things that weren't fair to Kasparov, changing the code between games being only one. I would actually say the more important piece was that he wasn't able to practice against Deep Blue or study its games against other players.

We tend to romanticize Chess as being a hyper-intellectual game where if you are good enough, you should be able to beat anyone worse than you 100% of the time. The truth is a lot more subtle. Chess is indeed a perfect-information game, so there isn't true luck involved, but due to the fact that we can't just search the entire game tree, we introduce several elements of uncertainty. Namely, that past a certain point, you must play your opponent as much as the board. Players (and AIs) naturally have certain strengths and weaknesses, and the ability to steer the game into positions that play to your strengths and opponents weaknesses is often just as important as your overall game strength. Some of the best players in the world are known for how intensely they prepare for specific opponents.

In short, Deep Blue had 'preparation' against Kasparov from all his game histories, while Kasparov did not have the same benefit. While switching the code could be considered analogous to switching to a whole new opponent (though likely only a small change in practice), the overall unpreparedness was probably the larger factor. In my opinion Kasparov was a stronger player than 1997 Deep Blue, but this asymmetry probably cost him the series.

In short, Deep Blue had 'preparation' against Kasparov from all his game histories, while Kasparov did not have the same benefit.

Kasparov had the benefit of playing all those games as well. ;)

As for your allegations of unfairness, code tweaks were allowed between games. It was part of the rules. Just not during games. And if Kasparov didn't think the rules were fair he could have declined the match.

>Kasparov had the benefit of playing all those games as well. ;)

You missed his point, its not about practice. The top chess players often study their opponents to understand what types of moves they make and in what situations they make those moves. If Kasparov were to play any other player in highly ranked matched both players would go and study their opponents move to get an insight in the way they think.

The "unfairness" in the Deep Blue match (or what made the Deep Blue match unlike other top-level play Chess), is that Deep Blue had a deep understanding of Kasparov's moves, but Kasparov had no information about Deep Blue's moves. And to make matters worse, any information he had built up in the previous games were erased once they tweaked the code.

I thought he walked out on the match because he was upset that the computer DIDN'T behave like a human. Huh. Weird, these narratives.

He didn't walk out. He played all six games in the match and lost the match in the last game after he played a opening that was know as losing by everyone after a piece sacrifice (it was speculated that Kasparov was gambling that deep blue wouldn't make the sacrifice and then he will have the upper hand).

Deep blue played pretty much "like a machine" except for a few moves that where considered very human because they hold very long term advantages.

After the match Kasparov hinted that IBM must have cheated, of course now a days a machine running in a off the shell desktop machine will surely beat the current world champion in a match, so Kasparov was just being a sore loser at the time.

A playback of that final game:

https://www.youtube.com/watch?v=cwmIHnEHfr8

When I think of Deep Blue not playing like a computer, I think more so of it leaving itself open to a forced draw the one game when it had the upper hand.

He felt that the engineers intervened DURING the games, and obviously did not feel that intervention between games was reasonable either. The former was disallowed, and the latter was taken advantage of more than Kasparov thought it would be.

He wasn't being a sore loser and there was a reasonable amount of agreement among other chess professionals (engine authors and players) that his claims were true.

Keep in mind as well that when he asked for a rematch with more stringent scrutiny in place, IBM declined and nearly immediately dismantled Deep Blue.

The unfortunate thing is that kasparov would ALWAYS have looked like a 'sore loser' unless IBM decided to implicate themselves of wrongdoing. Kasparov being silent also does a disservice to the community, particularly since he was neither the first or only person to point out the inconsistencies in Deep Blue's play.

Keep in mind as well that the 'real circumstances' make it so both parties responses were rationally founded within the context of the available information at the time.

We will never know for sure. If I was IBM, after the cheating accusations I'd probably refuse a rematch, too, even if I didn't cheat.

Why? Because unlike before, it would generate bad publicity for me now (the reason for a rematch would keep on being mentioned), and God forbid if ubermotivated Kasparov managed to triumph this time, under tighter scrutiny. Now the public is convinced that I cheated before... The atmosphere was spoiled and they had nothing to win in terms of PR anymore, really.

It was only reasonable to swipe it under the rug and move on. Obviously it does not mean that they were clean. Dropping Deep Blue is not an indicator of either in and of itself, because it makes perfect sense in either case. We just can't know.

My point about the rematch is that it made more reasonable to the chess community to believe Kasparov's claims.

From IBM's perspective, it made sense as you pointed out. From Kasparov's perspective, it also made sense for them to decline , but for a wholly different reason.

It really was just a sad situation where nobody could come out on top except for IBM. Even if they DID cheat, they were the only ones who were capable of proving that. IBM simply outsmarted Kasparov outside the game.

That would be an amazing parallel to the story of the Mechanical Turk.

The reason Kasparov hinted at cheating was that he had the measure of Deep Blue. When it played the random move it did something that fell outside of what he had learned to except. Hence, either IBM cheated by making a move for the computer or it was showing human intuition. It was pure coincidence that the move was actually better than what it's algorithm would have chosen but Kasparov wasn't to know that.

It broke his composure and concentration and the match was lost.

Afaik, I thought the engineers running the machine were reprogramming during the game to handle changes in Kasparov's technique. Please someone with more knowledge chime in, but from my memory it was more like he was competing against computationally augmented human opponent. He wasn't beaten by a computer, he was beaten by a team of engineers using a computer.

some corroboration on http://en.wikipedia.org/wiki/Deep_Blue_(chess_computer)

Not reprogramming during any game, but during the match (between games). You can argue about how much influence this has, in my mind though he was beaten by highly similar variants of a computer program (not quite a fixed concept, but not adjusted too much during the match).

Kasparov also complained that he was never allowed to see any logs or previous games by Deep Blue ahead of the match. If he was playing another GM, there would have been hundreds of games available where he could get a feeling for the player's preferences. Preparing for such a high level match necessarily requires such preparations.

He should have made his participation conditional on being shown previous games. I'm pretty sure everybody outside of IBM would have considered that fair.

IBM reminds me of a kid at school that would play everybody that was playing chess up until the point where he'd win once against them. Then he'd run off and spout to everybody that he was now better than 'x'. The rematch offer was pretty lame after the cheating accusations, they should have addressed those first. No rematch given those conditions, no prior data -> no fair victory.

If IBM wanted to do this fairly they'd give Kasparov at least as much data on Deep Blue as they had on Kasparov.

By Nate Silver's FiveThirtyEight a short doco

http://fivethirtyeight.com/features/the-man-vs-the-machine-f... (17mins)

Fun article, but does anyone else find the damned banner at the top of the page infuriatingly distracting? And is there a way to make it go away?

Ah, in firefox, right click -> inspect element, find 'position fixed', and untick it. Page behaves....

So is there any way to do this automatically whenever websites do this sort of thing?

Stylish/Userstyles might let you script at it easily, but there doesn't seem to be a dedicated app.

Reply on news.ycombinator.com