Date

2017

Document Type

Dissertation

Degree

Doctor of Philosophy

Department

Computer Science

First Adviser

Munoz-Avila, Hector

Other advisers/committee members

Heflin, Jeff; Cox, Michael T.; Davison, Brian D.

Abstract

The growing abundance of autonomous systems is driving the need for robust performance. Most current systems are not fully autonomous and often fail when placed in real environments. Via self-monitoring, agents can identify when their own, or externally given, boundaries are violated, thereby increasing their performance and reliability. Specifically, self-monitoring is the identification of unexpected situations that either (1) prohibit the agent from reaching its goal(s) or (2) result in the agent acting outside of its boundaries. Increasingly complex and open environments warrant the use of such robust autonomy (e.g., self-driving cars, delivery drones, and all types of future digital and physical assistants). The techniques presented herein advance the current state of the art in self-monitoring, demonstrating improved performance in a variety of challenging domains. In the aforementioned domains, there is an inability to plan for all possible situations. In many cases all aspects of a domain are not known beforehand, and, even if they were, the cost of encoding them is high. Self-monitoring agents are able to identify and then respond to previously unexpected situations, or never-before-encountered situations. When dealing with unknown situations, one must start with what is expected behavior and use that to derive unexpected behavior. The representation of expectations will vary among domains; in a real-time strategy game like Starcraft, it could be logically inferred concepts; in a mars rover domain, it could be an accumulation of actions' effects. Nonetheless, explicit expectations are necessary to identify the unexpected. This thesis lays the foundation for self-monitoring in goal driven autonomy agents in both rich and expressive domains and in partially observable domains. We introduce multiple techniques for handling such environments. We show how inferred expectations are needed to enable high level planning in real-time strategy games. We show how a hierarchical structure of Goal-driven Autonomy (GDA) enables agents to operate within large state spaces. Within Hierarchical Task Network planning, we show how informed expectations identify states that are likely to prevent an agent from reaching its goals in dynamic domains. Finally, we give a model of expectations for self-monitoring at the meta-cognitive level, and empirical results of agents equipped with and without metacognitive expectations.

Share

COinS