Multi-Agent Architectures that Facilitate Apprenticeship Learning for Real-Time Decision Making: Minerva and Gerona

This paper describes the Minerva and Gerona agent architectures, which have been designed to facilitate apprenticeship learning in real-time decision making domains. Apprenticeship is a form of learning by watching, which is particularly useful in multi-agent knowledge-intensive domains. In this form of situated learning, human and synthetic agents refine their knowledge in the process of critiquing the observed actions of each other, and resolving underlying knowledge differences. A major design feature of Minerva and Gerona is their method of knowledge representation of domain and control knowledge, both static and dynamic. Their representations facilitates reasoning over domain and control knowledge for the purpose of apprenticeship learning. This ability to reason over domain and control knowledge plays a central role in solving the global and local credit assignment problems that confront an apprenticeship learner.