Data-based Receding Horizon Control of Linear Network Systems

Ahmed Allibhoy, Jorge Cortés

Abstract

We propose a distributed data-based predictive control scheme to stabilize a network system described by linear dynamics. Agents cooperate to predict the future system evolution without knowledge of the dynamics, relying instead on learning a data-based representation from a single sample trajectory. We employ this representation to reformulate the finite-horizon Linear Quadratic Regulator problem as a network optimization with separable objective functions and locally expressible constraints. We show that the controller resulting from approximately solving this problem using a distributed optimization algorithm in a receding horizon manner is stabilizing. We validate our results through numerical simulations.

Introduction

In traditional control theory, designing an effective controller typically requires a precise mathematical model of the system, often represented by state-space matrices. However, for large-scale modern networks such as power grids, traffic management systems, or swarms of autonomous robots, identifying these global models is frequently computationally prohibitive or practically impossible.

Recent advancements have introduced "Data-enabled Predictive Control" (DeePC), which bypasses model identification by using raw historical data. However, DeePC was originally formulated for centralized systems. This paper addresses the critical challenge of distributed data-based control: enabling individual agents, who only possess local observations of their own and their neighbors' states, to cooperatively stabilize a global system without a central coordinator or a known model.

Methodology

The proposed approach leverages the Fundamental Lemma of behavioral systems theory. This lemma posits that a single trajectory of a linear system, if "persistently exciting," contains sufficient information to characterize any other possible trajectory.

Data-Based Representation

Instead of differential equations, the authors utilize Hankel matrices—structured arrangements of historical state and input data—to represent the system's future behavior. This allows the controller to predict future states directly from past data.

Distributed Optimization and Control

The standard Linear Quadratic Regulator (LQR) problem is reformulated into a distributed optimization framework. Key features include:

  • Separable Objectives: The global cost function is decomposed so that each node only manages its local data and constraints.
  • Primal-Dual Flow Algorithm: Agents solve the control problem online by exchanging small messages with immediate neighbors, iteratively reaching a consensus on optimal control actions.
  • Receding Horizon Control: Following the principles of Model Predictive Control (MPC), the system solves a finite-horizon optimization at each time step, implements the first control action, and repeats the process.

Experimental Evaluation

The framework was validated using synthetic numerical simulations on two distinct network topologies:

  1. Newman-Watts-Strogatz Network: A model for "small-world" networks typical of social and biological systems.
  2. Star Network: A topology featuring a central hub connected to multiple peripheral nodes.

Trajectories were generated using random, marginally stable inputs to ensure the historical data was sufficiently informative for the controller to learn the underlying dynamics.

Key Results and Discussion

The simulations demonstrated that the distributed, data-based controller effectively drives the network state to the origin.

Simulation results for Newman-Watts-Strogatz and Star networks showing state convergence and iteration counts.

Performance and Stability

As the optimization tolerance ($\rho$) is decreased, the performance of the distributed controller converges to that of a traditional MPC with full model knowledge. A significant theoretical contribution is the proof that the system remains stable even with approximate optimization solutions, provided the error remains within a specific bound. This is crucial for real-time applications where perfect convergence may be too slow. Notably, the simulations confirmed that the controller remains effective even when the time horizon is shorter than the system's dimension ($T < n$).

Communication Efficiency

The "iterations per step" analysis (shown in the right-hand plots of Figure 1) indicates that the algorithm converges rapidly. Furthermore, the authors introduced a local "suboptimality certificate," allowing agents to determine when they have communicated enough to guarantee stability, thereby conserving bandwidth.

Limitations and Future Directions

While promising, the current framework is limited to linear, noise-free systems. Real-world applications will require extending these methods to handle nonlinear dynamics and sensor noise, likely through regularization techniques. Additionally, the data volume required for "persistence of excitation" scales quadratically with network size ($O(n^2)$), presenting a potential bottleneck for extremely large systems. Future work will also focus on incorporating safety constraints on states and inputs.

Summary updated

Abstract

We propose a distributed data-based predictive control scheme to stabilize a network system described by linear dynamics. Agents cooperate to predict the future system evolution without knowledge of the dynamics, relying instead on learning a data-based representation from a single sample trajectory. We employ this representation to reformulate the finite-horizon Linear Quadratic Regulator problem as a network optimization with separable objective functions and locally expressible constraints. We show that the controller resulting from approximately solving this problem using a distributed optimization algorithm in a receding horizon manner is stabilizing. We validate our results through numerical simulations.
BibTeX
@inproceedings{allibhoy2021data,
  author = {Ahmed Allibhoy and Jorge Cortés},
  title = {{Data-based Receding Horizon Control of Linear Network Systems}},
  booktitle = {IEEE Control. Syst. Lett},
  year = {2021},
  url = {https://openreview.net/forum?id=TYaGl90AS6},
}
Figure