Machine-Learning and Game-Theoretic Inspired Strategies for Adversarial Freeze-Tag

In this thesis, we explore three machine-learning and game-theoretic inspired strategies to address the Freeze-Tag Problem with adversarial agents (FTPA). We develop a custom Gymnasium envi- ronment where unfrozen robots collaborate to “tag” and unfreeze immobile frozen robots while avoiding being tagged by the adversaries. The FTPA models scenarios such as restoring disabled relay towers after a cyberattack where different towers vary in priority, repair difficulty, and ad- versarial threats. We evaluate three strategies: a baseline heuristic where unfrozen robots target the nearest untargeted frozen robot, a hierarchical approach that incorporates adversary interac- tions and likelihood of success into decision-making, and a centralized multi-agent reinforcement learning strategy using Proximal Policy Optimization (PPO).