Deep Blue

A BruteForce chess machine, designed by a team lead by FengHsiungHsu, built by IBM. DeepBlue is the first machine to beat a human world chess champion in a match (1997, Gary Kasparov, score 3.5-2.5).

The work built on that of HansBerliner?'s chess machine Hitech, a successful hardware design with one custom chess chip per square for 64-way parallelism, and Feng's graduate work at CMU to implement Hitech's design in a single chip, DeepThought?. DeepBlue used many thousands of improved versions of DeepThought?'s chip to process the tips of the GameTree in parallel. DeepBlue used IBM's SP/2 for the framework for the parallel design.

It is estimated that IBM received many billions of dollars equivalent of good press from their successful challenge of the world champion. Unfortunately for computer chess fans, IBM would never dream of jeopardizing that PR with a rematch. DeepBlue was immediately decommissioned, probably never to play chess again. IBM's research into massively parallel computation continues in the form of BlueGene.

The following is from an ad for FengHsiungHsu's keynote at Oopsla '96:

 While ComputerChess? is traditionally
 considered an AI (ArtificialIntelligence) topic, Deep Blue,
 the strongest chess automaton ever built,
 is not a result of an AI project,
 but the end product of a concerted effort
 that draws from many fields beyond 
 what is normally considered AI: VLSI design,
 special purpose processing,
 algorithm design, parallel processing and so on.

Interestingly, some top grandmasters consider the very best PC chess software to now be playing better chess than DeepBlue did. Obviously that's a bit subjective, but objectively when Kasparov plays a chess match against a PC these days he's not doing much better than against DeepBlue.

Moved from ArtificialIntelligence:

The following is an ignorant falsehood; DeepBlue does not make such mistakes.

It's not false. I'm only an 1800 and won a pawn vs a 2600 rated computer (Rebel) set on maximum level (>1 day per move) that an amateur could easily determine was a weak move. Rebel moved the wrong rook to the center, very slightly weakening its queenside. Many moves later (>10), by gradually increasing pressure on the pawn, I captured the pawn with a winning position. Heuristically every human knows which rook to move but it didn't tangibly result in a weakened position until trillions of positions, something computers cannot currently do. I then fed the same position into it out of curiosity, left it on all week, and it realized it was a horrible move after 5-7 days (don't remember).

Part of why we're not impressed by DeepBlue as chess grandmaster are the stupid mistakes it can still make, that even a very average player wouldn't, in certain situations where real understanding of chess seems to be required. It just doesn't impress us as truly intelligent. Some will say that that's human arrogance. But I'd argue that it's human common sense. Something that AI finds notoriously difficult.

Grandmasters sometimes make stupid mistakes, so why hold DeepBlue to a different standard? Plus, I bet most players below the level of grandmaster would be hard-pressed to point out any stupid mistakes made by DeepBlue. One of the consultants to the DeepBlue team was grandmaster Joel Benjamin, who made the comment that he was "dominated" by DeepBlue.

[Years ago, when computer chess was just getting interesting, many chess players, Kasparov included, said that a computer could never beat a human beacuse a computer couldn't understand chess. This is human egocentricsm - it assumes that you have to solve chess the same way a human does to be a master. That's obviously false - DeepBlue, and all computers, win chess through computation. A human can't possibly compute at anywhere near the same level, which is why we rely on intuition and "understanding" to short-circuit our way through the heavy computation the computer does. It's worth noting that nobody really knows how that intuition process works, and that it may not really be that fundamentally different than DeepBlues? heavily parallel graph lookups. Regardless, DeepBlue is a better chess player than all but a tiny fraction of people - so small that it's statistcally non-existent, whether you can claim it "understands" chess or not, it is demonstrably more capable.]

The following is more than an ignorant falsehood, it's a complete fabrication:

Don't forget, deep blue played to beat one strategy only. If it had played anyone but Kasparov, it would have lost. In fact, we mustn't forget that IBM DECLINED when asked to have it play someone else. IF they can RELIABLY beat a human, I'll be impressed... If I get to pick the human after it's already programmed.

Deep Blue was not tuned to any particular strategy; it used a lot of chess knowledge, and highly optimized hardware. There's no doubt that the Deep Blue team tuned their machine for the Kasparov match, and that Kasparaov was at a disadvantage since he could not see any games the new DeepBlue was playing. But in a high-stakes match like this, it's customary to allow both sides to do as much preparation as they can afford.

[Why would you even think anyone would believe that? Kasparov, one of the greatest chess minds ever, only has *one* particular strategy, such that you can optimize a computer to beat him? Pshaw.]

It was an overhyped match. DeepBlue never was the best program vs humans, Rebel was and is (http://www.rebel.nl/) even on a PENTIUM (compared to DeepBlue's insane processing power). It plays more human style, and does human moves that no other chess program sees. Also Anand is far better against computers than Kasparov because of his insane speed (he is the fastest chess champion in history). The best human vs computer should be Anand vs Rebel. He routinely demolishes the best computer programs all the time, until he encountered Rebel. They fed the same positions and times that DeepBlue had in that tournament into Rebel and Rebel played better and would've obtained a better result. But all the marketing hype by IBM deceived people into thinking that DeepBlue was the best.

I don't know what it is about the DeepBlue match that fosters such strong feelings in the computer chess crowd. The argument that modern programs on modern PCs are better than DeepBlue is perennial on the various computer chess forums. At the time, DeepBlue could wipe the floor (as in not a single draw or loss) with the cream of the chess computer crop. There is no indication that PC's have gotten that much faster in the last decade, nor software that much stronger. However, since DeepBlue will never be resurrected, there is no way to measure current chess computer improvement, so all this arguing is just shouting into the wind, like arguing that Morphy could have beaten Fischer.

Care to back that up? All evidence I've seen indicates that Rebel was stronger than DeepBlue at the time.

In 1997, Rebel and Chess Genius were within 100 points of each other on computer rating lists. DeepBlue wiped out Genius 10-0 in a match using tournament time controls. Statistically, this works out to a rating difference of at least 400 points. QED. Furthermore, on the Man-Machine rating list available from the Rebel site, DeepBlue 97 has a 2862 performance rating whereas the 2002 Rebel at 2697 does not even top the list (Chess Tiger 2788). (It is unfortunate that DeepBlue was not allowed to play more of these kind of matches. It would have saved a lot of arguing.)

Also, the author of Rebel has specifically stated (on the ComputerChessClub? board) that there were no PC programs, including his own, that came anywhere near DeepBlue at the time of the match. I don't know Ed Schröder's opinion of current PCs and software (he's retired now).

CategoryHardware