Adaptive Agent Allocation for Massively Multi-agent Applications

Abstract

Although distributed computing is necessary to execute massively multi-agent applications, the distribution of agents is challenging especially when the communication patterns among agents are continuously changing. This paper proposes two adaptive agent allocation mechanisms for massively multi-agent applications: one mechanism aims at minimizing agent communication cost, while the other mechanism attempts to prevent overloaded computer nodes from negatively affecting overall performance. We synthesize these two mechanisms in a multi-agent framework called Adaptive Actor Architecture (AAA). In AAA, each agent platform monitors the workload of its computer node and the communication patterns of agents executing on it. An agent platform periodically reallocates agents according to their communication localities. When an agent platform is overloaded, the platform migrates a set of agents, which have more intra-group communication than inter-group or inter-node communication, to a lightly loaded agent platform. These adaptive agent allocation mechanisms are developed as fully distributed algorithms, and they move the selected agents as a group. In order to evaluate these mechanisms, preliminary experimental results with large-scale micro UAV (Unmanned Aerial Vehicle) simulations are described.