Greg Detre
@ 16.15 on Tuesday, 06 March, 2001
IBM Hursley
started off as Navy weapons engineering, then Grand Prix, then Star Wars (AI discrimination architecture), then IBM assimilated Data Sciences
chess - deep computational problem, applicable elsewhere (e.g. Blue Gene, protein folding)
it�s like when they set up two labs in Beijing and India on natural language
major web event - 4 million viewers in 1997
Kasparov vs Deep Blue � solved the Turing test for chess, but did not come close to the problem of real intelligence � didn�t even try
Kasparov � can remember move in every game, play 30 people blindfolded, see 14 moves ahead
Murray Campbell � Master level chess player, greatest compliment of his life was when Kasparov accused him of forcing one of Deep Blue�s moves
deep computing � shopping patterns, travel schedules, protein folding, fraud detection
3 approaches
improved algorithms
Kasparov is working at the meta-data level, meta-concepts, generating hypothesising and applying to test data � he can�t do brute force, like Deep Blue (they had more chips to add in the third game)
architectural
special-purpose hardware
games are good as the fruit fly of AI
well-defined rules + goals
easy to measure progress
large pool of experts
too complex to solve
looking for general methods for solving complex problems
deep thought 2 � massive parallelism + super-computing, specifically-designed architecture and software for that architecture
some Grandmasters claim that their style would be better suited at playing computers than Kasparov
every single move of Deep Blue�s is fresh � it doesn�t take history into account
Feb 1996 Kasparov won a 6-game match
what did they learn:
search was acceptable
chess knowledge was insufficient
needed better opening and endgame preparation
more flexibility during the match
they changed the weighting in the algorithms between but not during games
May 1997
Kasparov unveils anti-computer strategy - exploit his human long-range foresight
Deep Blue � half-way through, made a random move � software bug
Kasparov won the game
analysed the code, found the bug, decided that it was a random move
Kasparov concluded that the computer had worked its way all the way through the tree, and knowing it couldn�t win, made a random move
Kasparov then chose not to accept a draw in the second game and lost
40 moves in 2 hours � 30 supercomputer nodes, 500 processors = 200m posible positions, 2m/second
Deep Blue in Bedfont
alpha beta minimax algorithm
master node � worker nodes � chess accelerators (90% of the processing, 80% of the delegation by workers at software level???)
typically, Grandmaster easily looks 10 moves ahead � same with Deep Blue
each move takes an extra 5 times processing power
add quiescence � determines vulnerable areas with critical moves, and goes really deep with early critical moves � consistently 14 moves (move = move by white and a move by black)
extended book of 600,000 Grandmaster games
evaluation
complex feature detector on accelerator chip
8000 different features
each feature weight can be set individually � and is worked out and tweaked painstakingly with pen and paper with a Grandmaster player, rather than neural network or reinforcement learning
Deep Blue junior� - 1 second of processing on 1 processor on 1 node � still beats most people
contempt rating � largely to do with whether it accepts/offers draws
What spinoffs are you considering???
Blue gene � pharmaceutical industry
meteorological work
atomic weapons research
Specialist processor that knew about the domain � is that done elsewhere???
yes, e.g. in telecommunications for price-demand models, meteorology
What made you stop at the number of processors???
probably cost and convenience rather than science
Could a Grandmaster tell that it was Deep Blue rather than a human playing � that would be the real chess Turing test???
doesn�t think that there�s anything unique to Deep Blue�s technique
The emphasis in Deep Blue was on architecture over algorithms???
suspects not � it�s the least risk approach
but it�s less interesting
seems as though processing power is creeping towards intelligence, but that would be disappointing
it comes back to intelligence vs the solving of a complex problem
Turing test as appropriate as asking a submarine whether it can swim
necessary but not sufficient??? � NO, sufficient but not necessary
Are you working on Go playing??? � branching
IBM aren�t dong anything more with chess
still use chess for algorithm research
not yet, but interested in Go
management think it�s probably not worth their while
Did you ever feel guilty about what you did to Kasparov???
What about the Turing test for speech???
What decisions did you take to keep it strictly chess???