Data-Driven Cooperative Output Regulation via Distributed Internal Model
Liquan Lin, Jie Huang
TL;DR
This work tackles cooperative output regulation for unknown linear MAS with MIMO followers over a general static connected digraph. It introduces a data-driven distributed internal model leveraging a normalized virtual tracking error to decouple follower dynamics and enable learning of the Riccati solution via off-policy RL/ADP. The authors develop PI- and VI-based learning strategies and then propose two refinements that further reduce problem dimensionality and relax solvability conditions, leading to substantial computational savings. The resulting framework is scalable to cyclic graphs and unknown system matrices, offering a model-free approach to cooperative regulation with practical online applicability.
Abstract
The existing result on the cooperative output regulation problem for unknown linear multi-agent systems using a data-driven distributed internal model approach is limited to the case where each follower is a single-input and single-output system and the communication network among all agents is an acyclic static digraph. In this paper, we further address the same problem for unknown linear multi-agent systems with multi-input and multi-output followers over a general static and connected digraph. Further we make two main improvements over the existing result. First, we derive a set of much simplified linear systems to be applied by the integral reinforcement learning technique. Thus, the number of the unknown variables governed by a sequence of linear algebraic equations is much smaller than that of the existing approach. Second, we show that the sequence of linear algebraic equations can be further decoupled to two sequences of linear algebraic equations. As a result of these two improvements, our approach not only drastically reduces the computational cost, but also significantly weakens the solvability conditions in terms of the the full column rank requirements for these equations.