PhD Thesis Defense: Clement Nyanhongo

"Behavioral-Centric Team Evaluation via Consistent Rewards"

Abstract

Across human domains ranging from sports, business, and organizational settings, complex tasks are often solved by teams rather than individuals, leveraging benefits such as interaction, mutual support, complementary roles, cohesion, and task allocation. Evaluating team effectiveness, however, is inherently challenging due to the subjectivity of many existing techniques and the limitations of outcome-driven metrics, which primarily focus on performance results while overlooking the team processes that generate those results.

To address these challenges, this dissertation proposes a behavioral-centric, end-to-end framework for team evaluation using reward functions that model sequential team behavior. Reward functions provide compact and interpretable representations of team preferences and objectives, and can be learned via Inverse Reinforcement Learning (IRL) from observed behavior. Unlike manually hand-crafted rewards, IRL-derived rewards capture nuanced and implicit team goals embedded within historical behaviors. Nevertheless, learned reward functions can suffer from misspecification due to data sparsity and training instabilities, limiting their reliability. To improve robustness and quality, this dissertation introduces the concept of reward consistency as a formal measure of the stability and reliability of learned rewards under similar preferential conditions. In addition, canonicalization techniques are explored as preprocessing methods to standardize reward functions, thereby mitigating misspecification, enhancing interpretability, and potentially accelerating downstream Reinforcement Learning (RL) tasks. The refined reward representations enable richer forms of team evaluation, segmented into three key tasks: behavioral attribute extraction, team-impact rankings, and performance prediction.

Thesis Committee

• Eugene Santos Jr (Chair)
• George Cybenko
• Vikrant Vaze
• Raj Dasgupta (US Naval Research Lab)

Contact

For more information, contact Thayer Registrar at thayer.registrar@dartmouth.edu.