"Behavioral-centric team evaluation via consistent rewards"

Abstract

The process of team evaluation is essential for understanding team dynamics, which can be useful in predicting team performance and formulating strategies for future improvement. However, team evaluation presents significant challenges due to agent interactions that complicate the collective team behavior. To ensure robustness, team evaluation demands approaches that are behavioral-driven rather than simply outcome focused, so that interactions, motivations, and complex nonlinear processes that dictate the overall behavior, are fully incorporated. Several works in multiagent systems have adopted reward-based team evaluation approaches, which aim to compute rewards via Inverse Reinforcement Learning (IRL) to represent complex team behavior. However, reward functions are often inconsistent due to sparsity, noise and convergence instabilities which can occur during IRL training.

This thesis explores the property of "reward consistency," as a means to measure the stability, and replicability of IRL algorithms in representing agent behavior under the same preferential conditions. It proposes corrective solutions to improve reward consistency in IRL algorithms by converting rewards into standardized forms that are deterministic and stable. These solutions help to represent complex team preferences, which significantly improves the task of team evaluation.

Thesis Committee

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

Contact

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