A Robust Approach to Belief Mismatch in Non-cooperative Dynamic Games

Ryan Park, Li-huan Shen, Jesse Milzman, Negar Mehr, David Fridovich-KeilarXiv 2026 · arXiv

TL;DR

This paper introduces a robust planning framework for multi-agent dynamic games that accounts for "belief mismatch"—situations where agents have different, private estimates of the world state. By modeling a hypothetical adversary ("nature") that perturbs the predicted beliefs of other agents, the authors develop a polynomial-time solver that reduces performance degradation by over 48% compared to standard methods that assume all agents share a common belief.

What Problem Does This Paper Address?

In robotics and autonomous systems, game theory is a powerful tool for predicting how multiple agents (like self-driving cars and pedestrians) will interact. However, most existing solvers rely on a "common belief" assumption: they assume every agent sees the same sensor data and maintains the same estimate of the world.

In reality, interactions are often non-cooperative and private. A pedestrian might misjudge a car's speed, or a drone might have a noisier GPS than its neighbor. As shown in the figure below, if an autonomous car assumes a pedestrian has perfectly estimated its distance, it might fail to brake when the pedestrian—operating on a different, incorrect belief—decides to cross the street.

A car and pedestrian interaction where mismatched distance estimates lead to a collision.

When agents plan based on the fiction of a shared belief, their strategies become brittle. The authors aim to create a planner that is robust to these inevitable discrepancies in how different agents perceive their environment.

Methods

The authors formulate the problem as a Robust Belief Space Feedback Nash Equilibrium (FNE). Their approach involves three key technical shifts:

  1. Augmented Belief Space: Instead of one joint belief, the ego agent tracks a distribution for every agent, allowing for divergent state estimates.
  2. Adversarial "Nature": Inspired by $H_\infty$ robust control, they introduce a hypothetical $(N+1)$-th player called "nature." During the ego agent's planning phase, nature perturbs the predicted future beliefs of all other agents in a way that maximizes the ego agent's cost. This forces the ego agent to choose strategies that remain safe even if others' beliefs deviate from expectations.
  3. Decoupled Gradients: Standard solvers couple every agent's cost to the entire joint belief. This paper's algorithm evaluates each agent's cost only through their own belief, further mitigating the "shared belief" bias.

The resulting optimization problem is solved using a modified version of the iterative Linear-Quadratic-Gaussian (iLQG) algorithm, which maintains polynomial-time complexity relative to the number of agents and state dimensions.

Datasets

The authors evaluated their approach using two simulated experimental setups:

  • Activism (Opinion Dynamics): Two activist groups attempt to influence the opinions of three lawmakers. This tests robustness against epistemic uncertainty regarding how other agents gather information.
  • Hockey (Attacker vs. Defender): A defender (ego agent) must block an attacker from scoring. This tests robustness against misspecified parameters, such as the attacker's internal "cost" for expending effort.

Key Insights and Ideas

  • Belief vs. State Perturbation: Unlike classical robust control which perturbs the physical state dynamics, this method perturbs the belief dynamics. This allows the agent to reason specifically about the uncertainty of others' internal mental states.
  • The Cost of Robustness: There is a clear trade-off. Being robust helps significantly in "antagonistic" environments (where others act more aggressively than expected), but it can lead to higher costs in "collaborative" environments because the ego agent is being unnecessarily cautious.
  • Mutual Robustness is Pareto Inefficient: In the activism experiment, the authors found that if both agents choose to be robust, they both end up with higher costs than if they were both non-robust. This suggests a "prisoner's dilemma" dynamic in robust planning.

Key Results

The robust approach demonstrated significant improvements in handling model mismatch:

  • 48% Reduction in Cost: In the activism scenario, the robust planner reduced the additional cost incurred due to misspecified belief dynamics by 48% compared to the non-robust baseline.
  • Graceful Degradation: As shown in Figure 4, as the environment becomes more "adversarial" (positive drift), the robust strategies (lower $c$ values) maintain much lower costs than the non-robust baseline (black dashed line).
Total cost vs. assumed drift. Robust strategies (colored lines) outperform the non-robust baseline in adversarial settings.
  • Bounded Behavior: In the hockey experiment (Figure 6), the defender's robust strategy converged toward the non-robust baseline as the "nature" penalty increased, showing that the adversarial influence is well-behaved and leads to realistic, careful behavior rather than erratic movements.
Defender cost in the hockey experiment. Robust defenders (lower c) incur lower costs as the attacker becomes more aggressive.

Limitations and Future Work

  • Computational Complexity: While polynomial, the complexity is $O((N+1)(\eta_x)^6)$, where $\eta_x$ is the state dimension. This high-degree polynomial may limit real-time applications for systems with very large state spaces.
  • Gaussian Assumption: The current solver assumes beliefs are Gaussian and uses an Extended Kalman Filter (EKF) for updates. This may not capture highly non-linear or multi-modal uncertainties well.
  • Exploiting Mismatch: The authors note that while they focus on mitigating mismatch, a future direction could involve an ego agent proactively inducing or exploiting belief mismatch in others to achieve its goals.

Summary updated

Abstract

Game-theoretic models are increasingly used to describe multi-agent interactions in control and robotics applications. However, theoretical and computational complexity makes these models intractable in many dynamic, incomplete information scenarios. Recent work makes strong assumptions to reduce computational complexity: agents share sensor measurements with one another and maintain a common estimate of state uncertainty. Unfortunately, these assumptions enforce a degree of cooperation in generically noncooperative interactions. This paper takes the perspective of an ego agent seeking both computational tractability and robustness to unanticipated behavior of other agents due to mismatched state estimates from private---rather than shared---sensor measurements. To that end, we formulate the game theoretic planning problem in an augmented belief space with perturbed Bayesian dynamics that allows the ego to predict all agents' potentially divergent state estimates. These perturbations are chosen by a hypothetical adversary, ``nature,'' to reflect how non-ego agents' state estimates---based on measurements the ego agent will not see at runtime---may deviate from the ego's own belief. Ultimately, our approach can be understood in the spirit of classical H-infinity robust control, extended to noncooperative, stochastic dynamic games. We present a solver with polynomial-time iterations and conduct two experiments which show how robust belief space reasoning leads ego agents to proactively limit their future vulnerability to model misspecification of other agents' beliefs and actions, reducing performance degradation by over 48%.
BibTeX
@misc{park2026robust,
  author = {Ryan Park and Li-huan Shen and Jesse Milzman and Negar Mehr and David Fridovich-Keil},
  title = {{A Robust Approach to Belief Mismatch in Non-cooperative Dynamic Games}},
  year = {2026},
  eprint = {2610.40999},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2610.40999},
}
Figure