Decentralized reactive adjustment of agent teamwork organization to changing environment

An agent team with differentiated skill profiles of individual members, that performs a task in a dynamic environment which exhibits significant unpredictable changes, may need to adjust its work organization in order to maintain satisfactory performance. The uncertainty of long-term effects of such modifications favors reactive, incremental, and decentralized approaches. This paper explores the idea that such approaches should be based on flexible, dynamic combinations of complementary adjustment strategies. We investigate the effects of several strategies and their combinations on the performance of teamwork in a microworld setting. The strategies are represented by simple abstract models, whose impacts on team performance are studied experimentally using the Agent Interaction Model Simulator (AIMS) framework. The specific strategies include: independent replanning of individual subtasks; direct mutual help among teammates; and bilateral swapping of subtasks in progress among the team members. The results demonstrate that some of the component strategies are more adaptable and valuable than others, but that a well-formed combination regularly outperforms any single method. We also identify and discuss the challenges arising in the design and implementation of combined interaction protocols, formulate an improved version of the Requester-Initiated Action Mutual Assistance Protocol (RIAMAP), and introduce an interaction protocol for subtask swapping.