Theses and Dissertations

Advisor

Bethel, Cindy

Committee Member

Archibald, Christopher

Committee Member

Chen, Jingdao

Committee Member

Ball, John

Committee Member

Rahimi, Shahram

Date of Degree

8-7-2025

Original embargo terms

Immediate Worldwide Access

Document Type

Dissertation - Open Access

Major

Computer Science

Degree Name

Doctor of Philosophy (Ph.D.)

College

James Worth Bagley College of Engineering

Department

Department of Computer Science and Engineering

Abstract

In many real-world domains with continuous action spaces, an agent’s performance is driven by two key factors: its decision-making skill, or ability to select high-quality actions, and its execution skill, or precision in carrying out these actions. However, accurately evaluating these skills, particularly when direct information about the agent’s intentions is unavailable, remains a challenging problem. This research addresses this need by developing a theoretical framework and proposing novel estimation methods to assess both decision-making and execution skills based solely on observable data: executed actions and their associated rewards. The work consists of six core studies focused on developing and evaluating methods for estimating agents’ decision- making and execution skills in these types of settings with different types of agents. The first study introduces a reward-based method that estimates execution skills under specified assumptions about decision-making skills. The second study relaxes these assumptions by reasoning about agent actions within a Bayesian network, leading to enhanced execution skill estimation. The third study extends this network to jointly reason about actions and rewards. This approach facilitates the estimation of decision-making and execution skills, further improving the accuracy of the estimates. The fourth study adapts the approach from the third study to estimate execution skill in high-dimensional spaces and track changes over time by employing a particle filter within a dynamic Bayesian network. The effectiveness of the estimation methods is evaluated experimentally across various agent types and domains, showcasing their strengths and limitations. The results indicate that accounting for both decision-making and execution skills significantly improves the accuracy of the estimation compared to isolated approaches. To further evaluate the applicability of the methods, the fifth study presents a real-world application of the methods: evaluating the accuracy of Major League Baseball pitchers. Lastly, the sixth study presents a preliminary investigation of the impact and practical value of estimation methods in real-world settings. These studies collectively establish a robust framework for skill estimation, supporting the real-time assessment of agent performance and offering potential applications in player development and decision- making processes across diverse fields where agents interact in continuous action spaces.

Share

COinS