Deep Structured Teams in Arbitrary-Size Linear Networks: Decentralized Estimation, Optimal Control and Separation Principle

TL;DR

This paper develops decentralized Kalman filters for large-scale linear networks with deep structured interactions, providing a scalable solution with a separation principle and mean-field approximation as the number of agents grows.

Contribution

It introduces a novel decentralized Kalman filtering approach for deep structured teams, proving linear optimal strategies and establishing a separation principle in this complex setting.

Findings

Optimal strategies are linear in local and deep state estimates.

Kalman gain converges to zero as the number of agents increases.

Proposes a scalable, two-scale Riccati equation-based solution.

Abstract

In this article, we introduce decentralized Kalman filters for linear quadratic deep structured teams. The agents in deep structured teams are coupled in dynamics, costs and measurements through a set of linear regressions of the states and actions (also called deep states and deep actions). The information structure is decentralized, where every agent observes a noisy measurement of its local state and the global deep state. Since the number of agents is often very large in deep structured teams, any naive approach to finding an optimal Kalman filter suffers from the curse of dimensionality. Moreover, due to the decentralized nature of information structure, the resultant optimization problem is non-convex, in general, where non-linear strategies can outperform linear ones. However, we prove that the optimal strategy is linear in the local state estimate as well as the deep state estimate and can be efficiently computed by two scale-free Riccati equations and Kalman filters. We propose a bi-level orthogonal approach across both space and time levels based on a gauge transformation technique to achieve the above result. We also establish a separation principle between optimal control and optimal estimation. Furthermore, we show that as the number of agents goes to infinity, the Kalman gain associated with the deep state estimate converges to zero at a rate inversely proportional to the number of agents. This leads to a fully decentralized approximate strategy where every agent predicts the deep state by its conditional and unconditional expected value, also known as the certainty equivalence approximation and (weighted) mean-field approximation, respectively.