backtop


Print 53 comment(s) - last by blueboy09.. on May 3 at 2:32 AM

Skynet...ummm...Watson to compete on Jeopardy!

The producers of Jeopardy! and IBM are in discussions to allow an IBM supercomputer known as Watson to compete on the show against human competition. According to The New York Times, if Watson is able to beat the human competition the field of artificial intelligence will have made a leap forward.

IBM has had success in the past building super computers that were able to best human competitors. In 1997, IBM's Deep Blue supercomputer was able to defeat chess world champion Garry Kasparov in a match.

The difference between chess where all pieces have a known value and Jeopardy! is that the trivia questions asked in the game show have a wide and greatly varying range of relationships. These relationships are open to interpretation and the interpretations have to be made very quickly.

The IBM researchers who created Watson -- an homage to IBM founder Thomas J. Watson Sr. -- have said that they are not confident yet that their creation could compete well on the show. The New York Times reports that human champions are able to provide correct response 85% of the time to questions asked.

David A. Ferrucci, an AI researcher at IBM said, "The big goal is to get computers to be able to converse in human terms. And we’re not there yet.”

The contest is an effort by the IBM engineers to choose "grand challenges" that will help them make significant technical progress in AI. The rules proposed for the contest will force Watson to emulate all human qualities. Questions posed to Watson will be in text format while players will see text and hear the questions spoken by the show's host.

The computer will offer answers to the question via a synthesized voice and will choose its own follow up categories. IBM says that for the show, the computer would not be connected to the internet. How Watson will be presented and what gender the computer will be are under consideration. A screen and a projected avatar are one consideration.

Jeopardy! executive producer Harry Friedman said, "We’ve only begun to talk about it. We all agree that it shouldn’t look like Robby the Robot."

IBM will move a Blue Gene supercomputer to L.A. for the show.



Comments     Threshold


This article is over a month old, voting and posting comments is disabled

RE: Really AI?
By foolsgambit11 on 4/27/2009 9:46:25 PM , Rating: 3
Not exactly. Certain answers may be simple trivia ("This is the largest country in South America" - "What is Brazil?" - even this is hard to program - speech recognition, largest in what sense, what countries are part of the set of countries if South America, determining the word 'country' means 'nation-state', not 'rural area', etc.). But some categories have multiple decision criteria, linguistic puns, anagrams, etc. It's an incredible challenge, such that I'd recognize the program as possessing some rudimentary intelligence. Take for instance a category from this season, "Anagrams of Compass Directions", where the question is an anagram of a compass point, like, "To think over an issue long and hard" - "What is stew". The AI has to understand the question (answer), which has multiple answers (questions) that could fit. It has to figure out that it must look for a word or words whose definition is Trebek's answer, despite the fact that the definition it has for words will certainly not be worded the same way. Once it has a list of possible answers, it must form anagrams of all of them and compare the results with a list of all compass directions. All in 5 seconds from ringing in. It could perhaps work this problem in the opposite order, or from both sides, but it is certain that the programmers can't program a solution method for all kinds of Jeopardy categories, so the computer will have to come up with a method of arriving at a solution to complex problems on its own. It would be a huge accomplishment if the computer could even do poorly at Jeopardy.

Plus, the computer would have to have overall strategic planning, which would involve understanding of its own categoric strengths and weaknesses.


RE: Really AI?
By shin0bi272 on 4/28/2009 12:03:33 AM , Rating: 1
youre right but in your anagram example the computer would pull up all possible anagrams for all of the compass directions (assuming you are only talking about the cardinal directions) and their definitions and match a definition to the answer in miliseconds. Which wouldnt really be artificial intelligence it would be a well written program.
Artificial intelligence is where you walk up to a computer and ask it how it feels today and it makes a decision based on criteria that it came up with in its own order of importance. Like if it just said "im fine dave" every day you wouldnt think of it as "alive". But if you get a different response every day or it comes back with I feel sick I think I have a virus youd feel like it had a stronger awareness of its internal workings and its external stimuli.

Not knocking your post just pointing out that its not really intelligence if all its doing is systematically matching up definitions of words to possible answers and then phrasing the return in a form of a question. Granted that is how we think more or less but it doesnt mean its artificial intelligence when a pc does it.

If you want to see real ai look up "cog" its a robot that is breaking new ground in AI.


RE: Really AI?
By foolsgambit11 on 4/28/2009 4:48:09 PM , Rating: 2
You're right that it wouldn't be AI if it were programmed to solve the problem (I did mention that it could be solved either way), but the computer, assuming it had been in that game, wouldn't have been programmed to solve that specific problem class, since it hadn't been used. The point was that the computer would have to go about deciding how best to find a solution to the problem, which means in this case that it has to understand what form the answer will take (anagram), properly parse the clue, and understand synonymous definitions/phrases. It is by no means a trivial problem. Even if the computer were programmed to answer problems like this, it would have to recognize that that solution technique was called for in this category.

The problem isn't so simple as a series of string comparisons, because the computer is unlikely to have the definition of words phrased exactly the same as it is phrased in the clue.

It may be possible to program a computer to perform well on Jeopardy without any amount of intelligence (like it turned out to be possible for chess), but the amount of data the computer would have to store would be astronomical. Just feeding it the data would take years.


"If you can find a PS3 anywhere in North America that's been on shelves for more than five minutes, I'll give you 1,200 bucks for it." -- SCEA President Jack Tretton

Related Articles













botimage
Copyright 2014 DailyTech LLC. - RSS Feed | Advertise | About Us | Ethics | FAQ | Terms, Conditions & Privacy Information | Kristopher Kubicki