Many real-world systems possess a hierarchical structure where a strategic plan is forwarded and implemented in a top-down manner. Examples include business activities in large companies or policy making for reducing the spread during pandemics. We introduce a novel class of games that we call structured hierarchical games (SHGs) to capture these strategic interactions. In an SHG, each player is represented as a vertex in a multi-layer decision tree and controls a real-valued action vector reacting to orders from its predecessors and influencing its descendants behaviors strategically based on its own subjective utility. SHGs generalize extensive form games as well as Stackelberg games. For general SHGs with (possibly) nonconvex payoffs and high-dimensional action spaces, we propose a new solution concept which we call local subgame perfect equilibrium. By exploiting the hierarchical structure and strategic dependencies in payoffs, we derive a back propagation-style gradient-based algorithm which we call Differential Backward Induction to compute an equilibrium. We theoretically characterize the convergence properties of DBI and empirically demonstrate a large overlap between the stable points reached by DBI and equilibrium solutions. Finally, we demonstrate the effectiveness of our algorithm in finding emph{globally} stable solutions and its scalability for a recently introduced class of SHGs for pandemic policy making.
Download