Network Learning from Best-Response Dynamics in LQ Games

In this paper, we focus on network structure inference problem for linear-quadratic (LQ) games from best-response dynamics. An adversary is considered to have no knowledge of the game network structure but have the ability to observe all players' best-response actions and manipulate some players' actions. This work presents a comprehensive framework for network learning from best-response dynamics in LQ games. First of all, we establish theoretic results that characterize network structure identifiability and provide numerical examples to demonstrate the usefulness of our theoretic results. Next, in the face of the inherent stability and sparsity constraints for the game network structure, we propose an information-theoretic stable and sparse system identification algorithm for learning the network structure. Finally, the effectiveness of the proposed learning algorithm is tested. The connection between network structure inference problem and classical system identification theory is covered by our work, which advances the literature.