The Impact of Local Information on the Performance of Multiagent Systems
share
How does the control architecture of a multiagent system impact the achievable performance guarantees? In this paper we answer this question relative to a class of resource allocation problems, termed multiagent set covering problems. More specifically, we demonstrate how the degree of information available to each agent imposes constraints on the achievable efficiency guarantees associated with the emergent collective behavior. We measure these efficiency guarantees using the well-known metrics of price of anarchy and price of stability, which seek to bound the performance of the worst and best performing equilibria, respectively. Our first result demonstrates that when agents only know local information about the resources they can select, a system operator is forced to manage an inherent trade-off between the price of anarchy and price of stability, i.e., improving the quality of the worst equilibria comes at the expenses of the best equilibria. Our second result demonstrates that a system operator can move well beyond this price of anarchy / price of stability frontier by designing a communication protocol that provides agents with a limited amount of system-level information beyond these local quantities.
READ FULL TEXT