The Role of Meta-Reasoning in Achieving Effective Multi-Agent Coordination

Abstract

In this paper, we consider the role of meta-reasoning in achieving effective coordination among multiple agents in maintaining and executing joint plans in an uncertain environment. We assume that each agent has responsibility for performing a particular set of activities in the plan over time. Each agent is provided with an initial schedule and a set of alternative activities, but no agent has a global view of either the problem or solution. Individual agents must maintain their local schedules as execution events deviate from expectations. In circumstances where there are inter-dependencies among the activities of multiple agents, agents must coordinate. We describe an agent architecture designed to provide this distributed schedule management capability and to frame the meta-level control problem that it presents. We identify several degrees of freedom in configuring the agent's core computational components, each of which affects the proportion of computational cy-cles given to local scheduling and inter-agent coordination processes. We motivate the need for on-line reasoning by considering how aspects of the current control state impact the utility of different configurations. We present experimental results that demonstrate the potential benefit of adapting configurations to match the current control state. Finally, we sketch a framework for learning this mapping from experience as multi-agent schedule management problems are solved.