Abstract

Artificial intelligence systems are increasingly deployed as autonomous agents that interact directly with complex environments. Ensuring these agents remain reliable, adaptable, and useful presents an ongoing practical challenge. This dissertation studies agent-environment interactions through a progressive lens, examining behavior from foundational training stability and abstraction discovery to adaptation in unpredictable settings and open-ended creative search. The investigation begins by analyzing how hyperparameter settings affect early-stage Deep Q-Network performance, yielding practical guidelines for debugging reinforcement learning implementations. To address the challenge of long-horizon planning, the discovery of temporally extended options is formalized as the Minimum Shortcut Problem, demonstrating how agents can navigate environments more efficiently. As agents move beyond controlled settings, their ability to handle complexity and environmental shifts becomes critical. To evaluate model robustness against out-of-distribution data, a domain generalization benchmark for regression tasks is introduced, revealing that standard empirical risk minimization often performs as well as specialized robustness techniques. Furthermore, when environments contain other intelligent agents, reliable interaction requires robust internal modeling. By adapting metrics from representation learning, the stability and interpretability of multi-agent representations are evaluated beyond standard predictive task performance. Finally, the dissertation extends agent-environment interaction into the realm of computational creativity, exploring it as a formal search process. By characterizing creativity as an agent overcoming latent biases within its own search dynamics, a mathematical framework is introduced for evaluating transformational creative processes. Further, investigation into quality-diversity algorithms demonstrates how these search spaces can be navigated more effectively to discover collections of artifacts that are both diverse and high-quality. Together, these contributions provide a set of methodologies, benchmarks, and theoretical perspectives for understanding and improving how autonomous agents learn, adapt, and search within their environments.

Degree

PhD

College and Department

Computational, Mathematical, and Physical Sciences; Computer Science

Rights

https://lib.byu.edu/about/copyright/