Local distributed agent matchmaking

Abstract

This paper explores through simulation an abstract model of distributed matchmaking within multi-agent systems. We show that under certain conditions agents can find matches for cooperative tasks without the help of a predefined organization or central facilitator. We achieve this by having each agent search for partners among a small changing set of neighbors. We work with a system where agents look for any one of a number of identical task matches, and where the number of categories of tasks can be as large as 100. Agents dynamically form clusters 10 to 100 agents in size within which agents cooperate by exchanging addresses of non-matching neighbors. We find that control of these clusters cannot be easily distributed, but that distribution in the system as a whole can be maintained by limiting cluster size. We further show that in a dynamic system where tasks end and clusters change matchmaking can continue indefinitely organizing into newsets of clusters, as long as some agents are willing to be flexible and abandon tasks they cannot find matches for. We showthat in this case unmatched tasks can have a probability as lowas. 00005 of being changed per turn. © Springer-Verlag Berlin Heidelberg 2001.