As software becomes ever more complex and computing power becomes more distributed in autonomous components, components which are becoming ubiquitous, a key challenge in computer science is to understand the behaviour of these autonomous components and the interactions between them.
One promising approach is to treat the components as independent agents, and to build interaction mechanisms that can be used to coordinate such agents. My research centres around exactly this.
Within this broad area, the main topics that I am currently working on are:
A fundamental property of agents is their autonomy, which implies that what is in the best interests of one agent is not necessarily in the best interests of another. In such competitive environments, agents may deceive one another, and trust becomes an issue.
In interactions between autonomous entities, one cannot just tell another what to do. Rather agents have to persuade one another. A good account of how to do this is provided by formal systems of argumentation, in which agents exhange their reasons (arguments) for their beliefs (and desires and intended actions), and we have been working on argumentation-based interaction mechanisms for many years. Argumentation also provides us with an approach to decision making, and we have also been looking at using argumentation for handling trust.
Automated mechanism design
In recent years, techniques from game theory have become widely used in computer science, and the area of multiagent systems is no exception. From a game theory perpsective, designing a multiagent system is an exercise in mechanism design --- setting up the payoffs so that agents interact in the way desired by the system designer. We have been looking at the use of machine learning to automate the aquisition of new mechanisms, especially where the mechanisms are two-sided.
Robots are prototypical agents, having to interact with a noisy, dynamic environment. Now that robots are relatively cheap, and are becoming more capable, teams of robots are an attractive testbed for techniques from multiagent systems. We are looking at the effectiveness of interaction mechansism in organising heterogeneous teams of robots engaged in search and rescue tasks.
I also have an interest in datamining, and collaborate on this with George Hripsak at the Department of Biomedical Informatics at Columbia University.
More information on my work in these areas is given in the details of current projects and in the descriptions of the projects carried out by the Agents Lab.