Adaptive Task and Resource Allocation in Multi-Agent Systems

ABSTRACT In this paper, we present an adaptive organizational poli y for multi-agent systems alled tra e. tra e allows a olle tion of multi-agent organizations to dynami ally allo ate tasks and resour es between themselves in order to eÆ iently pro ess an in oming stream of task requests. tra e is intended to ope with environments in whi h tasks have time onstraints, and environments that are subje t to load variations. tra e is made up of two key elements: the task allo ation proto ol (tap) and the resour e allo ation proto ol (rap). The tap allows agents to ooperatively allo ate their tasks to other agents with the apability and opportunity to su essfully arry them out. As requests arrive arbitrarily, at any instant, some organizations ould have surplus resour es while others ould be ome overloaded. In order to minimize the number of lost requests aused by an overload, the allo ation of resour es to organizations is hanged dynami ally by the resour e allo ation proto ol (rap), whi h uses ideas from omputational market systems to allo ate resour es (in the form of problem solving agents) to organizations. We begin by formally de ning the task allo ation problem, and show that it is npomplete, and hen e that entralized solutions to the problem are unlikely to be feasible. We then introdu e the task and resour e allo ation proto ols, fo ussing on the way in whi h resour es are allo ated by the rap. We then present some experimental results, whi h show that tra e exhibits high performan e despite unanti ipated hanges in the environment.