RI: Small: A Study of Agent's Expectations for Nondeterministic and Dynamic Domains

点击翻译按钮获取中文摘要

Society is increasing relying on both (1) fully autonomous systems physically-embodied such as self-driving cars and automated cashiers as well as virtually-embodied such as automated stock traders and characters in virtual environments; and (2) semi-autonomous systems in physically-embodied platforms such as the break system in vehicles and virtual environments such as trip advisors. This project addresses the following question: when these agents are executing a course of action, how to compute the expectations of the agent's actions? This is a fundamental question as expectations can be used to detect when the agent's behavior is anomalous; that is, when their course of action deviates from the expectations. They can also be used to explain the agent's behavior by indicating what to expect from the agent's own execution. This project also includes a concrete educational activity targeting high school students that is tightly integrated with the particularities of the research program. The project will develop a plan for an educational unit providing an experience centered around a hands-on exercise experiencing AI by evaluating if an automated agent is performing its assigned tasks. Through that educational unit, students will gain an appreciation of the challenges in assessing if automated AI systems are behaving as intended.This project studies the problem of computing the expectations of course of action generated by an agent. Specifically, this project will investigate this problem (1) when the course of action is a policy that accounts for non-determinism in the outcome of the actions in the domain; (2) when the course of action is a plan and the domain includes numeric fluents bounded by margins of error; (3) when the course of action is a policy and the domain includes numeric fluents bounded by margins of error. The project is agnostic with regard to the particular planning paradigm used to generate the course of action and with regard to the paradigm used for correction of the course of action when a discrepancy is detected. The study is based on two representations of the courses of action: (1) sequential plans with numeric fluents valued within margins of errors and (2) policies for non-deterministic domains. Both of these representations account for uncertainty in the execution of the actions.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.