Learning Robust Communication by Adversarial Training in Networked System Control

Effective communication is paramount to achieving efficient cooperation in networked system control (NSC). Nonetheless, real-world challenges such as node failures and channel noise impede the generalization capabilities of networked agents. In this study, we seek to enhance the robustness of networked control strategies by introducing \(\text {RoComm}\) . An adaptive adversary is incorporated during the training phase to emulate potential node failures. This adversary strategically targets and attacks vulnerable nodes in the system, prompting protagonist agents to adapt and become less susceptible to node failures over time. When both policies reach convergence, it indicates that the protagonist policy can sustain robust performance, even amidst the most detrimental noise. Empirical results from two distinct NSC scenarios—traffic signal control and adaptive cruise control—demonstrate that RoComm not only augments the resilience of networked control policies against node failures but also amplifies their generalization capacities across varying environmental conditions, subsequently elevating their overall performance.