Model-Based Reflection and Self-Adaptation

Autonomous agents operate in complex and dynamic environments. Therefore, an intelligent agent must be able to learn new capabilities throughout its lifetime or it risks becoming obsolete and useless. Ideally an intelligent agent would be capable of both retrospective and proactive self-adaptation. Retrospective adaptation occurs when an agent fails in its assigned task, and proactive adaptation occurs when an agent is presented with a task different from the task(s) for which it was originally designed. In this talk, I will describe research in my group on retrospective and proactive self -adaptation in autonomous agents. We endow the autonomous agent with a teleological model of its reasoning process. The model enables the agent to introspect/reflect on its reasoning and to l ocalize the needed modifications. I will present experime ntal results from the domains of reactive control, route planning, and device assembly. I will also describe an integration of model-based reflection with reinforcement learning which appears quite promising.