Multiagent Systems

If you are wondering what the connection is between multiagent systems, the semantic web, and Web 2.0 technologies, the whitepaper AI Meets Web 2.0 by Jay M. Tenenbaum of commerceNet provides a good overview. The paper also appears on the Winter 2006 issue of AI magazine.

He proposes that we will build smart systems incrementally, in a similar way to how we build large programs via encapsulation, except that the whole world will be involved.

For example, I wanted to buy a Nintendo Wii on the Sunday they came out so I visited Target hoping they would have some. As you might guess, they were all sold out when I got there because people had been waiting in line overnight. Undeterred, I then went to Ebay to check out their prices and noticed that the going rate was somewhere around $1000. This was too high so I kept checking over the next few days and noticed how clearing prices were dropping. I also noticed that it was a pain for me to do this as I had to keep opening new tabs on my browser and waiting for several auctions to clear in order to get the going rate.

Continues below the fold...

If I had some more free time (or better tools) I could have built a small agent which would keep track of the going rate for the Wiis (Wiies?) and provided the equivalent of a stock ticker. I could have then made this service available to all. Of course, once I had this service it would be a simple matter to extend it to any item that the user desires to track (PS3s, Books, Speakers, etc.). Unfortunately, the problem is not that easy as even if two auctions are for a Wii they could have different terms (delivery price, open/unponed box, seller's rating, extra wiimote, etc.) thus we need to perform semantic matchings between two items given the user's description of what is important. This is the kind of service that could already by provided by some other agent out there, after all, it would also be a useful service for vertical search engines. Thus, another agent would be written which uses both of these services to provide a much more customizable "ticker-tape" to any good in the world. Further down the line these services could be combines with delivery agents, reputation reporting agents, etc., in order to provide smarter services as we tangentially approach a truly artificially intelligent web.

That is the basic vision that Tenenbaum presents and one that I also deem very reasonable. We build smart systems by aggregating not-so-smart parts—semantic mashups. There is also an evolutionary aspect to this vision as unused agents will die off and useful agents will survive and produce mutated offsprings, some of which will survive. It is unlikely that this process will lead to a human-like intelligence but it will, almost by definition, lead to something much better: an artificial intelligence which perfectly complements our human intelligence.

To make this vision happen we need to keep building back-ends to available web services and slowly add semantics to the data they provide by using microformats and, later, ontologies. This is already happening and, since Amazon and flickr have shown that it is profitable, will probably continue. The Web is slowly transforming itself from a repository of static HTML documents to a multiagent systems inhabited by ever more sophisticated and inter-connected agents

Distributed eBay

Another example he presents is zBay, which is an implementation of the idea presented in this very prescient post. zBay is eBay without eBay. That is, zBay is just a protocol. When you want to sell something on zBay you merely post it on your blog using the given microformat. Market maker agents will then find your item for sale and will take care of letting buyers find your item. The seller and buyer agents can then negotiate directly with each other, without the need to pay a commission to the middle man (eBay). The system can then be augmented by the addition of trust agents which monitor user's reliability, search engines like google which point you to the appropriate market, and recommender agents which make reasonable recommendations on alternative items a buyer might want.

I have taken the distributed eBay idea a step further and proposed the implementation of distributed combinatorial auctions where the sellers solve the winner determination problem, or where the buyers solve the problem. These papers are but a start.