Please create an account to participate in the Slashdot moderation system


Forgot your password?
IBM Hardware Games

IBM's Jeopardy Strategy 99

jfruhlinger writes "Developing a computer that could play chess once seemed like a worthy AI goal — but it turned out to be something of a dead end, as chess is very abstract and simple when compared to the real world. Will creating a game-show-playing computer lead to more interesting results? IBM hopes so, and its Watson machine will tackle problems in parallel processing, data searching, and natural language comprehension in an attempt to beat Jeopardy legends Ken Jennings and Brad Rutter. " IBM announced the man vs. machine competition last month.
This discussion has been archived. No new comments can be posted.

IBM's Jeopardy Strategy

Comments Filter:
  • by Haedrian ( 1676506 ) on Wednesday January 05, 2011 @04:40PM (#34769246)

    Once upon a time, a computer beating an expert at chess was amazing. Controversial.

    Even if they pull this off (which does happen to be a huge feat of AI if you think about it) - the general public won't care anymore. They think computers can do everything already.

    • actually, it may be more a feat of speed, since the computer will undoubtedly buzz in before Ken every time
      • It will probably be more a feat of statistical analysis of answers and finding the best ratio of certainty to uncertainty.

      • My understanding of the problem here is that really good human players often buzz in when they don't yet know the exact answer but they know that they know the answer. So they can get a jump on a computer that actually has to understand the question, process it and then access the answer before buzzing in. Essentially, humans get hunches and have intuition while (currently) computers do not.
        • by bberens ( 965711 )
          The computer can calculate that it has X seconds between the time it buzzes and the time it must answer. It can determine the most statistically likely correct answer it can find within X-1 seconds of "buzzing" and report that answer.
          • by timster ( 32400 )

            Jeopardy has a substantial penalty for incorrect answers, so blindly buzzing in requires a high threshold of confidence that you will be able to correctly answer any possible question (or, um, correctly question any possible answer).

            • I don't believe anyone was suggesting blindly buzzing in but rather calculating that you can calculate or finish calculating the correct question in 5 secs or less (to a high degree of confidence). A highly-parallel computer (read: multi-core) should be able to devote some number of threads of execution to actually calculating the answer and some other number to the calculation of the feasibility of calculating the right answer (and/or monitoring the other threads).
              • "calculating that you can calculate or finish calculating the correct question in 5 secs or less"

                Finding the right answer is a classical halting problem [], trying to calculate when it will find the answer is futile. All you can say is that it will take X amount of time to evaluate Y potential answers wich is insuffcient information to decide when to press the buzzer.
                • The halting problem would only come into play for a very specific class of queries (and those are surely not the sort of queries Jeopardy! typically involves). Doesn't this come down to what information you have in cache and what you have to look up (and how long it is projected to take to look it up)?
                  • No it's not about the cache. As I said it's trivial to predict how long it will take to evaluate X answers and halt, yet it's still impossible to predict if or when during that period it will find a "good" one and press the buzzer. It could be the first, last, or none of the answers evaluated. If you do manage this impossible feat of prediction then what you have done is created a more efficient evaluation algorithim and are back to square one trying to predict how long the new algorithim will take to find
                • In other words, the halting problem is only one of many decision problems (most of which are decidable [and in much less time than actually finding solutions to their counterparts in np-complete]).
                  • The halting problem is an example of an undecideable problem. All known search algorithims are examples of the halting problem since it takes the same or more time to compute when the algorithim will halt than it does to run the algorithim.
                    • Firstly, I doubt that a search algorithm would be the 'one true way' to question all Jeopardy! answers (and I think the rules of Jeopardy! are such that any non-trivial question will be accepted if it truly matches the answer given).

                      Secondly, [even if it were] the human brain is able to achieve this (presumably via some sort of holographic/striped storage) so it is possible in theory to do this in computers as well (assuming that there is no metaphysical explanation for the brain's capabilities)--i.e., e
          • by tlhIngan ( 30335 )

            The computer can calculate that it has X seconds between the time it buzzes and the time it must answer. It can determine the most statistically likely correct answer it can find within X-1 seconds of "buzzing" and report that answer.

            Which doesn't mean an instant win, since it's still a difficult problem to parse the answer (which may rely on puns and other trivial), analyze the context, and formulate the question. Get the question wrong, or fail to answer in 5 seconds (the maximum time you get after buzzin

            • by bws111 ( 1216812 )

              Players can not buzz in before the answer is completely read and the light comes on. Attempting to buzz in before that locks out your buzzer for a period of time, which is why sometimes you see players clicking wildly trying to get in.

    • Re:Meh... (Score:4, Funny)

      by zero.kalvin ( 1231372 ) on Wednesday January 05, 2011 @04:49PM (#34769362)
      Obligatory XKCD : [] though you know, i don't think we can even start to imagine what kind of application a real capable AI can have. Just leave it to the future!
      • I thought that the spammers where beating the CAPTCHA's by iframing them to free pron sites and assuming that whatever they enter is correct. User gets the pron and the spammer gets the CAPTCHA answer.
      • Only problem with any voting system is that spammers can get botnet votes. A comment like "SJDHIWH@IYG#" may have 4 million upvotes, none of which come from a human.

      • The algorithm is the same one used by slashdot. The weak point is in the moderation. Trolls will vote up other trolls and vote down actual constructive comments. Say do you smell something burning?
    • You hit that spot-on. The general public see things from movies from 20 years ago saying that in the year 2010 we would have this and that and expect those prophecies to be fulfilled. You also get those who see things in movies or on TV like CSI where you can just 'enhance' to a godly resolution and people just don't care about these kinds of advancements. When I was learning about internet frameworks such as "web 2.0 - 4.0" one of the things mentioned was that there will be hurdles to overcome and the a
    • Even if they pull this off (which does happen to be a huge feat of AI if you think about it) - the general public won't care anymore. They think computers can do everything already.

      Who the f*ck cares about the general public? Why does everything we do must be measured according to what "the general public" wants? If we acted that way, many things would not have happened, because the "general public" does not seem to be interested in. I care about this thing, that's enough for me.

      • . . .

        You think a marketting stunt like this - teaching a computer to play Jeopardy against a human isn't done for the general public?

        If it was for nerds, it'd do something else.

  • What is a possible new leap in interactive interfaces that would make a Bat-computer type UI possible?
  • The Penis Mightier for 200, Alex.

    IBM should give it a Scottish accent; that way even if it fails, it will still be funny.

    • What’s the difference between a mallard with a cold and you? One's a sick duck, and I can't remember the rest but your mother's a whore!

  • by $RANDOMLUSER ( 804576 ) on Wednesday January 05, 2011 @04:46PM (#34769320)
    I'll take Patent Trolls for 100, Alex.
  • I'm not trolling or anything, I'm honestly really curious what the value in dumping all this money into R&D for this issue is? Will we really gain deep insights into AI that we don't already have by doing this?

    • Re: (Score:3, Insightful)

      by SheeEttin ( 899897 )
      It's not dumping money into R&D. It's dumping money into marketing.
      • They are doing a lot of R&D to prepare for this specific challenge. It's not like the Jeopardy show is charging them to appear, they probably both benefit from the free publicity.

    • by Haedrian ( 1676506 ) on Wednesday January 05, 2011 @04:56PM (#34769454)

      Human Language Processing is still a weak point in getting computers to do what film computers do.

      If you can get a computer to understand what you mean, then it'd change UIs forever.

    • I'm not trolling or anything, I'm honestly really curious what the value in dumping all this money into R&D for this issue is? Will we really gain deep insights into AI that we don't already have by doing this?

      Cheaper than outsourcing support to India? (And potentially more accurate?)

    • by bws111 ( 1216812 )

      To the general public, it is just for the 'cool' factor. To IBM's customers, it says 'We figured out how to solve this difficult problem, and we can figure out your business problems too'.

    • Since IBM has retooled itself from a hardware/software company into an IT services company, I see a future where the nth-generation of their Jeopardy program mans the help desk, fielding technical questions and whatnot from IBM's very much "human" customers. If they could perfect this, then it could spell trouble to the millions of call center workers in India, the Philippines, or the emerging outsourcing powers in Africa.
    • "The effort of using machines to mimic the human mind has always struck me as rather silly: I'd rather use them to mimic something better." - E.W. Dijkstra
  • Let chatbot authors compete too!

    • Cleverbot buzzes in.

      Trebek: Yes, Cleverbot.

      Cleverbot: I'm not Cleverbot, you're Cleverbot.

      Trebek: Yes... Well then, Janice?

      Terrible, terrible idea.

  • they are tackling a much more difficult game of Jeopardy. For fun and profit.

    Hmm ... something doesn't seem to be right here ...

  • Nice move, there, I-beam.

  • by gurps_npc ( 621217 ) on Wednesday January 05, 2011 @05:15PM (#34769698) Homepage
    Unlike winning at chess, which has little if any real world possibilities (except allowing solitary chase playing for the grand masters). Answering jeopardy style vague questions is at the heart of many help desk applications, searching, and even reception work. This is a real product/service that can be sold.
    • by SamSim ( 630795 )
      And that's just considering the end result of this thing, namely Watson itself. Looking at the sheer amount of original research and work that had to be done to create it, it's unthinkable that there wouldn't be results in there that are worth spinning off and applying to other applications.
    • by Krneki ( 1192201 )

      Having a computer answering daily your simple questions is a big deal.

      Sure, it will take some time to get this CPU power in every home, but considering the hardware and software evolution it shouldn't be that far away.
  • And now we go to our Final Jeopardy category: "Famous Motherboards".
  • The hardest part of that is scoring and selecting the hits you get from the skein of database queries you make from the keywords in the clue.

    If you want to do something impressive, make it learn what it knows the same way Jeopardy contestants did: by reading books and organizing the data within them in content-addressable memory.

    • Sure it's impressive. It's never been within our reach to build a machine with such practical capabilities. Granted, it did not *learn* the information but does that really matter so long as it still gets the answers right? I'm sure Jeff Bezos and his ChaCha brethren are looking closely at this.
    • That "hardest part" is quite hard indeed. It implies huge advances in language processing, which, next to vision, is one of the "hard problems" in AI. Jeopardy answers involve trivia, yes, but many of them also take wordplay, puns, phonetics, analogies, non-sequiturs, etc. Jeopardy is not some cheesy show like Who Wants To Be A Millionaire; Google could win that one without trying hard... winning Jeopardy is going to require feats of AI that are leaps and bounds beyond anything else currently available.


  • *knock knock, door opens* "Yes?" "Who will give birth to the man who will lead the resistance against the machines after Judgement day ultimately overthrowing skynet and returning control of the planet to humans." "Umm, who is Sarah Connor?" *BLAM* "Why are you killing us?" "I'm sorry, please rephrase your statement in the form of an answer."
  • February Sweeps Gimmicks for 400, Alex
  • by Big_Breaker ( 190457 ) on Wednesday January 05, 2011 @05:34PM (#34769984)

    Most of the time players find the daily double while running through categories they know REALLY well. Then they only bet ~$2k out of their $12k stash.

    Even if they get the daily double right they will have to risk losing in final jeopardy b/c they haven't doubled the second place player's score. The smart play is to "make it a true daily double" and lap your opponents on a category you know well. Daily double questions are no harder than final jeopardy and I generally find them much easier. That's the time to risk it all. You not only increase your probability of winning but also your cash winnings.

    Imagine you are up 80% on your opponent and the final jeopardy category comes up as something you know NOTHING about. That's the time when you wish you bet more on that daily double.

  • Watson: What's the difference between you and a mallard with a cold? One's a sick duck and I can't remember how it ends, but your mother's a whore.
  • A computer that competes on "Dancing With The Stars".

    • by Curl E ( 226133 )
      "A computer once beat me at chess, but it was no match for me at kick boxing"
      • When I was a kid, that was the funniest thing ever.

        Then I went home, turned on my SNES, punched the difficulty in SF2 up to max and then cried a lot.

  • Jennings: I'll take CAPTCHAs for 1000, Alex...

    Big Blue: Damnit.

    • by treeves ( 963993 )

      Ah, but those people don't get on Jeopardy. And the Ken Jennings types don't mess with Wheel Of Fortune.
      That was a funny clip though.

  • Do they eliminate the buzzing in to claim the answer and have all players answer all questions? Because it would seem that the computer could always instantaneously buzz in and use the next few seconds to conjure up the answer, just as many humans do, save that 3-4 seconds of computational time can effect a pretty massive search / advantage. I would certainly run metrics on my program and figure out the optimal lead time to have it buzz prior to finding the answer. As has been noted, this is 75 percent m
  • I'm reminded of the webcomic Nine Planets Without Intelligent Life []. It's been a while since I reread the comic, but as I recall, what sparked the AI revolution was when scientists got bored with computers playing chess and the like and set about making a robot that could appreciate a theme restaurant.

Receiving a million dollars tax free will make you feel better than being flat broke and having a stomach ache. -- Dolph Sharp, "I'm O.K., You're Not So Hot"