Introduction
Classical stochastic team theory typically rests on the "common prior" assumption, where all agents agree on the probabilistic description of their environment. However, in decentralized systems, decision-makers (DMs) often possess subjective or disparate models of the world. This paper addresses the fundamental challenge of multi-person decision-making when agents disagree on the underlying probability space. By framing this discrepancy as a nonzero-sum stochastic game, the research establishes a rigorous equilibrium theory for quadratic objectives in general inner-product spaces.
Theoretical Framework
The study considers two decision-makers, each aiming to minimize a quadratic cost functional. The core innovation lies in handling the discrepancy between their subjective probability measures, $P_1$ and $P_2$.
- Hilbert Space Formulation: Decisions are treated as elements of general inner-product spaces, ensuring the theory applies to both discrete and continuous-time problems.
- Radon-Nikodym Derivatives: To reconcile the different models, the author utilizes Radon-Nikodym (R-N) derivatives. These derivatives act as "translation" factors, allowing expectations under one measure to be expressed in terms of the other.
- Decision Modes: The paper analyzes two primary modes of interaction: the Symmetric (Nash) mode, where DMs act simultaneously, and the Asymmetric (Stackelberg) mode, where one DM acts as a leader anticipating the follower's response.
Key Insights: The Gaussian Surprise
A significant finding of this work concerns the structure of optimal policies under Gaussian distributions. In standard Linear-Quadratic-Gaussian (LQG) problems, optimal policies are invariably linear. However, this research reveals a striking divergence:
- Nash Equilibrium: In the symmetric case, the unique stable equilibrium remains affine (linear plus a constant). This suggests that model disagreement does not fundamentally alter the structural simplicity of Nash solutions in Gaussian environments.
- Stackelberg Equilibrium: Conversely, in the asymmetric case, the leader's optimal policy is generically nonlinear. It typically involves exponential terms where the exponents are quadratic functions of the measurements. Linear solutions only emerge in highly specific, restricted scenarios.
Stability and Convergence
The paper provides sufficient conditions for the existence and uniqueness of these equilibria. A crucial practical contribution is the development of iterative schemes. For the symmetric case, the equilibrium is shown to be "stable," meaning it can be reached through a sequence of optimal responses. Notably, agents can converge to this joint strategy without needing to know each other's private probabilistic models—they only need to observe the other's previous policy.
Conclusion and Future Directions
This work extends the boundaries of stochastic game theory by formalizing decision-making under model uncertainty. While it provides a robust foundation for static information patterns, several avenues for future research remain. These include extending the theory to dynamic, multi-stage problems where agents may learn or share observations, and addressing the complexities that arise when the number of decision-makers exceeds two.
Summary updated