General Question

LostInParadise's avatar

Any insights on how a computer can win at Jeopardy?

Asked by LostInParadise (24776points) January 14th, 2011

In a trial round of against former high winning champions of the television quiz show Jeopardy, a computer designed by IBM came out on top. To me this seems extraordinary. I can understand how it can have huge amounts of information stored. What I don’t understand is how it can interpret the questions and use them to organize its information search. Are call center workers in danger of being replaced by computers?

Observing members: 0 Composing members: 0

6 Answers

Vincentt's avatar

Hehe, long term perhaps, but if they get any more helpful from that…

It is impossible to explain quickly how this works. Presumably they wrote some research papers on it, and it would probably take quite some years to reach the level of expertise required to comprehend those. The Wikipedia page(artificial_intelligence_software) mentions the use of a lot of algorithms concurrently to verify answers.

You might find True Knowledge a bit better (but still difficult) to understand.

janbb's avatar

Jeopardy is being filmed at IBM headquarters today. Human competitors are playing against Watson – IBM’s contestant. In trial rounds as you say, the computer won. I assume the search algorithms are similar to Natural Language searches.

janbb's avatar

Thinking about it more, I can see what would be problematic about it in terms of the word play and the categories. My son works at IBM so I asked him but he doesn’t know specifics about what they did.

funkdaddy's avatar

It’s an amazing achievement, this is way up the difficulty scale from playing chess (or tic-tac-toe for those that remember Wargames).

I think a few things are probably working for them here though

1) Jeopardy questions are fairly short, they have to fit in those small boxes so there’s unlikely to be a ton of extra words or descriptors. Questions are going to be straight forward for the most part.

2) They have a huge number of questions to run the program against from all the previous shows. So Watson’s “team” can run their program against thousands of questions, see what types of questions they get right or wrong, and then add to their solving algorithm as needed and run everything again very quickly. At the start they were probably making huge changes every day to try and get the percentage of right answers up. Eventually they cover a wide variety of not only topics and categories but styles of questions.

3) There’s not much penalty for not knowing the answer, worst case scenario is someone else gets it, they don’t have to guess at questions they aren’t sure about.

4) Fast computers = fast answers and fast buzzing in. If Watson “knows” the answer, it probably has it way before us poor meatbags.

flutherother's avatar

I am surprised by this as I didn’t know computers were so advanced but then I was surprised when computers began beating us humans at chess. Watson appears to understand the questions as it gives the correct answer but it can only achieve this through brute force and massive processing power. I read that it has 15 terabytes of RAM and uses multiple algorithms simultaneously. It has learned through trial and error the probability of an answer being correct and is now beating human contestants at the game.

Quite astonishing however and we may soon be speaking to computers more than other humans.

janbb's avatar

The show is being taped today and will be broadcast at some point. I am eager to see it.

I remember a long time ago reading that one of the distinguishing facets of AI was that the computer would be able to learn from its mistakes and auto-correct so I assume that is built in to Watson as well as a knowledge base. But how the hell does it press the button?

Answer this question




to answer.

This question is in the General Section. Responses must be helpful and on-topic.

Your answer will be saved while you login or join.

Have a question? Ask Fluther!

What do you know more about?
Knowledge Networking @ Fluther