Learning how to plan and instantiate a plan in multi-agent coalition

Abstract

We propose an innovative two-step learning approach to planning-instantiation for multi-agent coalition formation in dynamic, uncertain, real-time, and noisy environments. The first step learns about the planning of a coalition to improve its quality, adapting to the real-time and environmental requirements. The second step learns about the instantiation of the plan to improve the formation process, taking into account uncertain and dynamic behaviors of the peer agents. Decomposing the approach into two steps allows for modularity and flexibility in learning: learning how to plan a coalition is strategic while learning how to instantiate a plan is tactical. Our approach employs a case-based reinforcement learning (CBRL) framework.