D3PI: Data-Driven Distributed Policy Iteration for Homogeneous Interconnected Systems

Abstract

Control of large-scale networked systems often necessitates the availability of complex models for the interactions amongst the agents. While building accurate models of these interactions could become prohibitive in many applications, data-driven control methods can circumvent model complexities by directly synthesizing a controller from the observed data. In this paper, we propose the Data-Driven Distributed Policy Iteration (D3PI) algorithm to design a feedback mechanism for a potentially large system that enjoys an underlying graph structure characterizing communications among the agents. Rather than having access to system parameters, our algorithm requires temporary "auxiliary" links to boost information exchange of a small portion of the graph during the learning phase. Therein, the costs are partitioned for learning and non-learning agents in order to ensure consistent control of the entire network. After the termination of the learning process, a distributed policy is proposed for the entire networked system by leveraging estimated components obtained in the learning phase. We provide extensive stability and convergence guarantees of the proposed distributed controller throughout the learning phase by exploiting the structure of the system parameters that occur due to the graph topology and existence of the temporary links. The practicality of our method is then illustrated with a simulation.

Data-Driven Distributed and Localized Model Predictive Control

“…However, the amount of data that is required by the algorithm scales proportionally with the size of the network. In [16], it assumed that "auxiliary" links among subsystems exist, which are then exploited to solve the LQR problem with unknown dynamics and local communication; this assumption however makes its extension to an online approach (e.g., MPC) computionally unappealing. Moreover, while there exist works that guarantee stability in centralized data-driven MPC settings [9], it is unclear how such techniques can be adapted to provide theoretical guarantees in the distributed setting.…”

Online Adversarial Stabilization of Unknown Networked Systems

“…Among these, Fattahi et al (2020) demonstrates a first use case of SLS theory in learning-based distributed controller learning. However, most prior work that considers distributed learning and control schemes use the stochastic noise or no-noise model, assume a known stabilizing distributed controller is given, and cannot handle general communication delay during learning, e.g., Li et al (2021b); Alonso et al (2021); Jing et al (2021); Alemzadeh and Mesbahi (2019); Talebi et al (2021a); Alemzadeh et al (2021). Ho and Doyle (2019) presents an adaptive SLS controller but requires a known stabilizing controller and does not have guaranteed stability for large uncertainties.…”