The Cyc Project
I finally started reading a book I got from the library a few weeks ago called "Understanding Artificial Intelligence". The book is a compilation of articles about AI published in Scientific American over the past ten or so years. One of the articles I read last night was by a guy named Doug Lenat, who talked about an AI system developed at Stanford in 1984 called Cyc (as in en-cyc-lopedia). Lenat's basic premise is that a fundamental requirement for a computer program that is designed to exhibit human-like reasoning is that the program have a solid understanding of everything that humans would loosely group under the term "common sense". Otherwise, the program will not be able to make realistic, believable decisions. Cyc, therefore, is designed to be a vast knowledge base of concepts and relations that humans generally take to be common sense.
In the mid-90's the Cyc project was spun off into a private company called Cycorp. Lenat has continued to oversee the development and expansion of Cyc; the knowledge base now contains about 300,000 concepts and over 3 million "assertions", or relations between concepts. I found a ton of information about Lenat and Cyc by googling their names, including an hour-long Google Tech Talk that Lenat presented a few months ago where he drew parallels between Cyc and the Google search engine, showing how despite Google's power, it cannot interpret simple phrases or provide answers to questions. Combining the technology behind Google with the knowledge base behind Cyc, however, could lead to a radically more powerful organization, utility, and accessibility of information in the future. It sounds like this capability could both complement, and in some cases surpass, the advantages of the Semantic Web, assuming the latter ever actually materializes.
Cycorp has an open source version of the knowledge base available, as well as an expanded version specifically for use by research institutions. The company is currently working a number of national security-related contracts, an area where this kind of system could be put to effective use, culling through millions of pieces of information and trying to estimate probabilities of future events based on a common sense understanding of the world.
I think this approach to creating an intelligent system that can exhibit believable reasoning is fascinating, if a bit over-ambitious (after 20 years of work, the developers believe that it only contains about 2% of the total information it would need to operate with human-level intelligence) and narrowly focused. Having a working system like Cyc, even with the limitations it currently has in terms of comprehensiveness, helping control the responses of non-player characters in games could revolutionize the way players interact with them. Combined with a stab at natural language processing, I could see the days of static dialogue tree conversations coming to an end, with the player being given the ability to query NPCs on subjects of interest and receiving realistic answers. Imagine a Half-Life where you could hound Barney about that beer he owes you? Or a Deus Ex where you could coax more tips out of Tracer Tong? This kind of interactivity would completely change the way we even play games, throwing all the interactive game mechanics conventions right out the window.
In the mid-90's the Cyc project was spun off into a private company called Cycorp. Lenat has continued to oversee the development and expansion of Cyc; the knowledge base now contains about 300,000 concepts and over 3 million "assertions", or relations between concepts. I found a ton of information about Lenat and Cyc by googling their names, including an hour-long Google Tech Talk that Lenat presented a few months ago where he drew parallels between Cyc and the Google search engine, showing how despite Google's power, it cannot interpret simple phrases or provide answers to questions. Combining the technology behind Google with the knowledge base behind Cyc, however, could lead to a radically more powerful organization, utility, and accessibility of information in the future. It sounds like this capability could both complement, and in some cases surpass, the advantages of the Semantic Web, assuming the latter ever actually materializes.
Cycorp has an open source version of the knowledge base available, as well as an expanded version specifically for use by research institutions. The company is currently working a number of national security-related contracts, an area where this kind of system could be put to effective use, culling through millions of pieces of information and trying to estimate probabilities of future events based on a common sense understanding of the world.
I think this approach to creating an intelligent system that can exhibit believable reasoning is fascinating, if a bit over-ambitious (after 20 years of work, the developers believe that it only contains about 2% of the total information it would need to operate with human-level intelligence) and narrowly focused. Having a working system like Cyc, even with the limitations it currently has in terms of comprehensiveness, helping control the responses of non-player characters in games could revolutionize the way players interact with them. Combined with a stab at natural language processing, I could see the days of static dialogue tree conversations coming to an end, with the player being given the ability to query NPCs on subjects of interest and receiving realistic answers. Imagine a Half-Life where you could hound Barney about that beer he owes you? Or a Deus Ex where you could coax more tips out of Tracer Tong? This kind of interactivity would completely change the way we even play games, throwing all the interactive game mechanics conventions right out the window.
posted by GregP at 10:21 PM
0 Comments:
Post a Comment
<< Home