|Developer Diary · You Heard It Here First · 19 November 2003|
|Chess, Go and Artificial Intelligence|
Yesterday afternoon Garry Kasparov drew game 4 of his match against computer chess heavyweight Fritz at the New York Athletic Club to tie the match at 2 points each. With this result it seems humans and computers are running neck and neck for the chess crown. Last year Kasparov, rated #1 in the world, drew a match against another computer, Deep Junior and World Champion Kramnik drew a similar match.
Twenty-five years ago chess was often cited as a test of "artificial intelligence" for a computer but nowadays the mystique of a chess-playing computer has worn off and programs like Fritz are considered to be just big number crunchers, not intelligent beings. A better test of "intelligence" for a game would be the game of Go. Go relies more on pattern recognition than chess does. In go even an average player can easily beat the strongest computers. An interesting question is whether it will be possible to create a strong go-playing computer that does not have other artificial intelligence characteristics or whether a strong go-playing computer will imply intelligence. In other words, is pattern recognition and use the essence of intelligence?
Artificial intelligence itself remains as elusive as ever. Currently the world championship of artificial intelligence is the Loebner Prize. You get a $100,000 and a gold medal if you write a program that can pass the Turing Test, ie consistently fool people into thinking it is another person, and a machine vision test. There is also a one-time prize in the form of a silver medal and $25,000 for a program that can outperform at least two humans in a Turing test. Every year they also offer a bronze medal for the most convincing computer. Judging from the competition this year the computers still have a long way to go. Here is an excerpt from this year's winner, Jabberwock:
2003-10-18-15-57-18 JABBERWOCK: Salutation, roger. JUDGE: thanks - is this a south american greeting? JABBERWOCK: Hello there! JUDGE: let us talk about something eles JABBERWOCK: Hey, great idea! Go for it! JUDGE: what about football JABBERWOCK: Football sucks JUDGE: i agree with you - what is your favourite subject JABBERWOCK: You want to butter up me, roger? JUDGE: i certainly do if we share a common interest JABBERWOCK: Certainly not! JUDGE: are you becoming defensive? JABBERWOCK: Heck! I won't say "yes" or "no" right now JUDGE: you must come off the fence JABBERWOCK: Whatever you say, roger JUDGE: lets get to know each other better JABBERWOCK: Let's not. I'd like you to explore this topic more deeply ...
Apparently the best way to win the contest is to have the most variety of non-committal phrases like "Whatever you say" and "I won't say yes or no right now". The logs I read gave no indication that any of the programs had much ability to understand and relate to what the judge was saying. When I was a student in artificial intelligence 10 years ago we deeply studied subjects like knowledge representation, reasoning and logic, pattern matching, natural language and so on, but all this work seems to have gone for nought when you read transcripts like the above.
We endlessly tinker with the cogs and symbols of intelligent life but never can we create it.
|Developer Diary · firstname.lastname@example.org · bio · Revised 19 November 2003 · Pure Content|