Interleaved Information Structures in Dynamic Games: A General Framework with Application to the Linear-Quadratic Case

A team of researchers from UT Austin and other institutions has published a paper introducing a novel framework for solving a critical problem in multi-agent AI and game theory: dynamic games with 'interleaved information structures.' Traditional models assume agents either see the full system state (feedback) or only the initial state (open-loop), which are often unrealistic. In real-world scenarios like autonomous vehicle coordination or robotic teams, agents typically have partial, overlapping views. This new work provides the first systematic method to model and solve games where each agent observes only a specific subset of other agents at every timestep.

Their first major contribution is a modeling technique that frames these complex games as Mathematical Program Networks (MPNs), where the network's structure directly encodes the informational dependencies between agents. Their second contribution is a solution method: for the common and tractable linear-quadratic (LQ) dynamic game, they leverage the MPN formulation to derive a set of Riccati-like equations. These equations systematically characterize Nash equilibria—the stable outcomes where no agent can benefit by changing strategy alone. The researchers demonstrated their framework with a three-agent example featuring a cyclic information structure, proving its practical applicability for designing and analyzing more realistic multi-agent AI systems.