“In the 1990s, world chess champion Gary Kasparov played two historic matches against IBM’s Deep Blue supercomputer. He won the first match but lost the second by just a single point. As a graduate student at Stanford writing a thesis on artificial intelligence at that time, I was fascinated by the match. I’d been a computer hobbyist since the 1980s as well as a chess buff. Over the years I’ve tried practically every commercially available chess program on every platform, including Sargon, Socrates, Chessmaster, and others. I used TRS 80 and Apple II computers, and then IBM PCs running DOS, followed by Macintosh and Windows systems. Since that time I’ve wanted a chess computer as powerful as Deep Blue – my own world-champion-level sparring partner.”
Big Blue….. It seems like only yesterday, and how far away it seemed from being even remotely affordable. How times have changed. You could far surpass Deep Blue’s performance on what is in most companies’ server rooms and clusters these days (and in a few years, on one machine), and the advances in chess software and in evaluating moves are mind blowing. If it was done again I doubt whether any human would win.
Pretty interesting read.
Deep Blue vs Kasparov, the greatest publicity stunt ever. A match that is remembered almost a decade later, and will be for some longer.
Article gets a bit boring in the middle when discussing what other hardware he put in his ‘chesscomputer’ but all in all a good read if you’re interested in chess AI. He could’ve done with slightly more details about Deep Blue itself.
Deep Blue partially won because it was tweaked to defeat Kasparov; focus of Kasparov’s anger afterwards. But it was an impressive feat to create Deep Blue itself. Some interesting optimisations were made from an AI perspective. It’s required reading for an AI course
The article says that four 2.2 GHz processors are equivalent to a 8.8 GHz processor; this guy either is simplifying things too much so that nontechnical people can understand him, or has a very deep lack of understanding on basic concepts like processor clock, architectures, multiprocessing, and how these things are related.
This simplification is pretty reasonable in the context of a chess program, because the problem is embarrassingly parallelizable. It is easy to break the problem of searching the game tree into many independent subproblems, so the four processors can typically run with almost no synchronization.
Yes, in practice, an 8.8 GHz processor would not behave the same, were we able to build one, since it would suffer more from cache misses, peak memory bandwidth limitations, etc. This is putting a pretty fine point on things though…
I bet some of the readers could overclock that puppy so it would wipe the floor with Deep Blue 😛
On a serious note, it’s dissapointing that there was no actual game of chess between Deep Blue and Deep Blitz
And there never will be, since IBM retired Deep Blue from chess playing.
One cabinet from Deep Blue is on display in the Smithsonian Museum of American History, right next to WOPR.
“How about a nice game of chess?” — WOPR
While it was impressive for its time. Computer Scientists and artificial intelligence people are still trying to create algorithms and computers that can beat Go players. IIRC the most advanced ones can barely beat high-level beginners to low-end intermediate players. To beat an advanced player, a ridiculous handicap is needed, this wikipedia link says 25 stones.
http://en.wikipedia.org/wiki/Go_(board_game)
Edited 2006-01-27 05:26
For those who love chess, be sure to check this out:
http://www.chessgames.com/perl/chessgame?gid=1269891
Kasparov vs X3D Fritz. One of the collest computer X human games ever played.
It’s interesting that although there are languages specifically targetting AI programming (such as Lisp and Scheme) they choose to write Big Blue in C
Does anyone know if these languages have been discarded (in general) by the AI community?
“Does anyone know if these languages have been discarded (in general) by the AI community?”
What AI community?? 😎
Joking aside, this was a great article!
I used to play quiet a lot of chess with the author Ron Herardian in the 80s and 90s. His style was quite headstrong and he was capable of some memorable and imaginitive sacrificial combinations which I was usually unable to refute over the board. He was a big fan of Morphy, Lasker, and Alekhine.
At the time we used Socrates on our PCs as an analysis tool. Socrates found an entertaining 16 move continuation which would have saved me after I unwisely took the bait (a knight for a pawn if I recall) on one of his better brilliances. We’d team up and take on the machine, developing our own anti-computer strategies.
I hope you can enter your machine in the tourney for the computer championship. You might have to travel to Spain to do it, though.
Cheers!
After a few pages this article gets pretty boring in talk about high end Opteron PC HW. If I were involved in chess engines I would go straight to FPGA parallel HW development as did the team that was in the news last year at the UAE? show which also had Opteron servers. The HW used for that was Xilinx FPGA based built by Nallatech in Scotland. I don’t think it really takes so much general purpose compute power at all except in the sense of post processing the vast nos of candidates the FPGA can grind out.
I would go further, use a simple FPGA RISC cpu core that needs about 1% of FPGA and surround it with move eval engines and do as much reduction locally. These smarter cpu-move engines can then be replicatedd quite a few times per FPGA and what comes out is at a much higher level for the host cpu to analyze. These cpu-move engines would only chugg along at 150MHz or so but reduction has to be done close to where its generated. Such cpus are much closer to the old PPCs used by DeepBlue and the FPGA logic is probably more capable today than those old ASICs too. As the paper says, its the HW-SW synergy that gives the 200M positions/sec, you really don’t want to use dumb brute force SW only on massive Opteron farms except that at the end of the day when you shut the thing down, the Opterons still have some use while the FPGA cards are not so easy to reuse.
Not that I play chess anymore
In this article it compares Spec 95 and Spec 2000 numbers , in one hand it acknowledge that both benchmarks are different, but it still says that the Spec2000 number of an Athon is xxx times that Spec95 of Deep Blue CPU.
In other word, this figure is meaningless but the author still use it because it’s impressive.. *sigh*