The Big Picture
Representing time-stepped decision problems as a network of linked optimizations makes it possible to compute exact equilibrium strategies for linear–quadratic multi-agent problems where each agent only observes some other agents’ states.
ON THIS PAGE
The Evidence
Model each agent’s time-indexed decision problem as a node in a network that captures who sees what. For games with linear dynamics and quadratic costs, the usual optimality conditions (first-order KKT conditions) collapse into Riccati-like matrix equations (a recursive matrix formula used in control) that fully characterize Nash equilibria under arbitrary partial-observation patterns. The paper gives a step-by-step recipe—build the network, write node-level optimality conditions, then solve the coupled Riccati-like system—and demonstrates it on a three-player circular observation example. A2A Protocol Pattern
Data Highlights
1Any T-step dynamic game can be turned into an MPN with T nodes per agent (one node per decision time).
2Open-loop construction creates T−1 temporal edges per agent (connections from each time node to the next).
3Illustrative example: solved a 3-agent game with horizon T=3 and a cyclic observation pattern to produce explicit Riccati-like equations for the equilibrium.
What This Means
Engineers designing or testing multi-agent systems where agents have limited or asymmetric visibility—use this to predict strategic behavior and design interaction rules. Control and robotics teams working with models that are linear in dynamics and quadratic in cost can get exact equilibrium policies rather than relying on heuristics. Researchers studying multi-agent evaluation or orchestration can use the framework to analyze how partial visibility changes outcomes. Mutual Verification Pattern
Not sure where to start?Get personalized recommendations
Ready to evaluate your AI agents?
Learn how ReputAgent helps teams build trustworthy AI through systematic evaluation.
Learn MoreKeep in Mind
Results apply exactly only to linear dynamics and quadratic costs; nonlinear problems will need approximations or new theory. The method assumes the information pattern (who observes whom and when) and model dynamics/costs are known ahead of time; unknown or evolving observation patterns are not handled here. Computational complexity grows with the number of agents and time steps because the KKT-derived system couples many node-level problems into a larger matrix system. Context Drift
Methodology & More
Real-world groups of decision-makers rarely fall into the two textbook extremes of either seeing the entire system at every step or seeing only the initial condition. More often, each agent can observe a different subset of other agents’ states at each time. Represent those time-indexed, partially visible decision problems as a Mathematical Program Network (MPN): nodes are agents’ optimization problems at each time step, and directed edges encode how one node’s decisions influence reachable nodes later in the network. That representation makes the cross-dependencies explicit and systematic to analyze. For games with linear state updates and quadratic costs (linear–quadratic games), first-order optimality conditions for each node (the KKT conditions, which are the necessary equations for constrained optimality) remain valid and, because of convexity, are sufficient. Concatenating those node-level KKT conditions yields a coupled system that can be written as Riccati-like equations—a recursive matrix relation familiar from optimal control. Solve the resulting system to get Nash equilibrium policies under any interleaved observation pattern. The paper walks through building the MPN, forming node Lagrangians, deriving stationarity multipliers, and eliminating them to produce the final Riccati-like system, then illustrates the pipeline on a three-player, three-step cyclic observation case. The approach gives practitioners a concrete, exact tool for predicting strategic interactions in partially observed multi-agent settings and a foundation for extending to learning or non-linear scenarios. Human-in-the-Loop Pattern Consensus-Based Decision Pattern
Avoid common pitfallsLearn what failures to watch for
Credibility Assessment:
Mixed author signals: two authors with low h-index but contains established researchers (Ufuk Topcu, David Fridovich-Keil) — recognized researchers though affiliations not listed; solid but not top-tier venue.