Cooperative Agent System for Quantifying Link Robustness in Tactical Networks

This paper presents a cooperative agent system employing Reinforcement Learning (RL) to quantify path robustness in tactical networks. We implement an environment builder agent generating network graphs with dynamic topologies, constrained by tactical network characteristics and conditions derived from experimental results using real military radios (VHF). Subsequently, the RL agent interacts with this environment, learning to quantify the robustness of feasible transmission paths by solving an optimal control problem. Furthermore, we propose an adversary to evaluate the RL agent's robustness against adversarial perturbations. Experiments across a large set of network graphs reveal that policy optimization algorithms (PPO and TRPO) outperform others (ARS, DQN and A2C), demonstrating high rewards (similar to 105) in the absence of adversaries. However, when assessing performance in scenarios with adversarial noise injection, experimental results indicate that iterative, and momentum boosting attack techniques can substantially degrade the RL model's performance by up to 300% lower rewards.