Publication Date

2025

Document Type

Thesis

Committee Members

Xiaodong Zhang, Ph.D. (Advisor); Weisong Wang, Ph.D. (Committee Member); Trevor Bihl, Ph.D. (Committee Member)

Degree Name

Master of Science in Electrical Engineering (MSEE)

Abstract

The development of intelligent and competitive agents in AI versus AI adversarial environments was explored through the utilization of reinforcement learning techniques with sensing modalities. A Hide-and-Seek simulation environment was developed using the Unity game engine along with the ML-Agents Toolkit. An engagement test campaign with a set of performance metrics was designed. Four AI versus AI adversarial scenarios were considered using the Proximal Policy Optimization (PPO) and Soft Actor-Critic (SAC) multi-agent reinforcement learning algorithms. Furthermore, the impact of sensing modalities on competing agents’ learning performance was investigated by varying the sensing capabilities of the hider and seeker, respectively. Experiments across RL algorithms and observation models show that the seeker has the highest competitive advantage in the scenario wherein the seeker is trained with SAC versus a hider trained with PPO; while the hider demonstrated the best performance when the seeker was trained with PPO versus a hider trained with SAC. In addition, comparative studies show that extra sensing modalities generally result in improved agent learning performance, with the employment of a far-field sensor, even in the presence of measurement noise, outperforming a near-field sensor. Moreover, results suggest that when an agent has a competitive advantage from powerful AI algorithms, its performance is more robust to variations in sensing modalities.

Page Count

78

Department or Program

Department of Electrical Engineering

Year Degree Awarded

2025

ORCID ID

0009-0000-1468-865X


Share

COinS