Multiagent Systems

James Hendler posts on his IEEE Internet Computing article Embracing Web 3.0 (with Ora Lassila). The article is a nice quick summary of what has been happening in the semantic web and what the near future holds.

As they point out, RDF seems to be gaining acceptance in the business world (Oracle is proving support for it) and small companies are springing up which use RDF-based databases. An RDF triple store looks a lot like a relational database except that new associations can be added easily. For example, in a relational database you might have an Employee table with an ID and a Name column. If you later want to add a column for JobType then you have to either drop the table and create a new one or create a new table and link it to the old one via the ID, which will make all future queries much more complicated. An RDF store lets you deal with specific instances, so when a new employee that has a JobType appears you can simply store that new relation.

Just like relational databases have SQL, so do RDF databases have SPARQL, which does look a lot like SQL but with a few additions. Here a is a good tutorial on using sparql on Jena.

So, what next? Well, marking up data and writing a string-matching query language is the easy part. The hard part, the AI part, is to figure out how to use existing SPARQL services to dynamically put together value-added services. For example, imagine that all websites provide SPARQL backends to their data, and all use symbols (ontologies) from a limited but overlapping set (that is, 'foaf:friend' and 'facebook:friend' might not mean the same thing). How do write an agent that can provide good answers to all these questions:

• Find me all restaurants near my home which have been rated highly by my friends, are under $50, and have short wait lines.
• What is the most popular topic of conversation among 18 to 25 year old science savvy females?
• What is the going rate for a new Wii on the secondary market?
• Find me a good beach rental in the South Carolina coast for this summer with kid activities nearby.

One can probably see how to write an agent that specializes in one particular question, or question type, while a general question-answering agent is currently beyond our capabilities. Thus, I see that the near future will bring us more and more specialized agents that can answer questions that people care about. Over time these agents will merge into more generalized agents to the point where there will we will have a few general question-answering agents.

Will Google write these agents or will someone else be the next Google?