Economically Founded Multiagent System

Tuomas Sandholm

Course Description

Multiagent systems research, a subfield of AI, studies the interactions of computational agents. These agents can represent different real world parties, and they can have different preference structures. Important applications include manufacturing planning and scheduling among multiple agile enterprises, markets for electricity, allocating and pricing bandwidth in multi-provider multi- consumer computer networks, network management, multiagent information gathering on the web, distributed vehicle routing among independent dispatch centers, electronic commerce, resource allocation in distributed operating systems, meeting scheduling, scheduling of patient treatments across hospitals, classroom scheduling, and planning and scheduling of multi- contractor software projects, to name just a few. Multiagent systems can save users' time, but they may also achieve better solutions (e.g. by enhanced negotiation and coalition formation) than human agents can in combinatorially and strategically complex domains.

A key research goal is to design open distributed systems in a principled way that leads to globally desirable outcomes even though every participating agent only considers its own good and may act insincerely. The tutorial covers relevant results in AI, game theory, market mechanisms, voting, auctions, coalition formation, and contract nets. Emphasis is given to rigorous concepts, results, and algorithms - both classic ones from microeconomics and recent ones from the distributed AI community - that have direct applications to computational multiagent systems. Implementation experiences will be shared. Effects of different computational limitations (agents' bounded rationality) are discussed as a key feature that has not received adequate attention. Examples of real-world applications will be presented.

Prerequisite Knowledge

The tutorial equips the serious systems builder with rigorous techniques for making multiple self- interested agents cooperate efficiently. It also serves to familiarize newcomers and executive level participants with the issues involved. No prior knowledge is assumed in economics or multiagent systems. A general familiarity with computer science will be helpful.

About the Lecturers

Tuomas Sandholm is assistant professor of computer science at Washington University. He received the M.S. (B.S. included) with distinction in Industrial Engineering and Management Science from the Helsinki University of Technology, Finland, in 1991. From 1988 to 1992, he worked as a research scientist in the software industry. He earned the M.S. and Ph.D. degrees in computer science from the University of Massachusetts at Amherst in 1994 and 1996 respectively. He has published 23 refereed articles in AI Journal, IJCAI, AAAI and other forums - as well as numerous book chapters, technical reports and other papers. He has been a program committee member for five major conferences, and a reviewer for seven journals and numerous conferences. He has six years of experience designing efficient multiagent systems. This work has focused both on theory and implementations. He has also been involved in developing two fielded AI systems: a pension law expert system and a large- scale transportation optimization application.
Last modified: Thu Feb 20 13:21:13 JST 1997