Creative Problem Solving in Intelligent Agents: Domain-Agnostic Methods for Novelty Resolution

Abstract

PhD Defense:

Humans and many other species have the remarkable ability to innovate and creatively problem solve on-the-fly. For example, when trying to remove a screw in the absence of appropriate tools, we may resort to using a quarter as a makeshift screw driver. Creative problem solving (CPS) is the process by which an agent discovers unknown information about itself and its environment, allowing it to accomplish a previously impossible goal. In contrast to humans, robust CPS in intelligent agents is currently beyond the scope of existing artificial intelligence (AI) systems. Developing CPS in intelligent systems would greatly improve their resourcefulness and adaptability, allowing for more seamless integration of such systems into everyday life. To bridge this gap, this dissertation introduces a collection of contributions toward the development of CPS, including theoretical formalizations of CPS, a framework for anomaly detection, a framework for fault diagnosis, two methods for action discovery in robots, and a method for life-long CPS. The results from several diverse experimental studies show that the tangible implementation of CPS is possible in the following contexts -- generalized anomaly detection, system-level fault diagnosis, action discovery in embodied agents, and in using a trained world model for life-long learning. The experimental domains of our studies include aviation (tested on a human-in-the-loop flight simulation environment), satellite operation (tested on a realistic flight hardware test-bed), and robotics (tested in 3D physics simulation environments). Future work should consider the development of CPS in more realistic real-world environments to help expedite/realize the beneficial integration of intelligent novelty resolution methods into everyday life.

Join the meeting in Tisch Library #304 or via Zoom:

https://tufts.zoom.us/my/evana.gizzi

Password: see colloquium email

Dial-in not an option for this event.