Abstract
Modern large-scale interconnected systems, such as power grids and robot swarms, present significant challenges for traditional control methods due to the curse of dimensionality and unknown underlying physics. This paper introduces a scalable framework for designing distributed controllers for these systems. By integrating System Level Synthesis (SLS) with Lasso-based system identification, the authors demonstrate that near-optimal controllers can be learned with a sample complexity that scales logarithmically with system size. Furthermore, the synthesis process is computationally efficient, operating in near-linear time.
Introduction
Infrastructure in the modern era—ranging from smart cities to autonomous transportation networks—comprises thousands of interconnected subsystems. Controlling such systems is notoriously difficult for two primary reasons:
- The Curse of Dimensionality: Standard Linear-Quadratic Regulator (LQR) methods typically require data samples and computational resources that scale cubically with the number of components, making city-scale applications infeasible.
- Unknown Dynamics and Safety Requirements: The exact mathematical models governing interactions between subsystems are often unknown. While black-box machine learning offers a potential solution, it frequently lacks the stability and safety guarantees essential for critical infrastructure.
This research proposes a method to learn a distributed controller—where individual sub-controllers communicate only with their immediate neighbors—that is both robust to model inaccuracies and computationally efficient.
Methodology
The authors employ a two-stage "Identify-then-Control" pipeline to address these challenges.
Sparse System Identification
Physical systems often exhibit "sparsity," where a component's state is only influenced by its local neighborhood. Exploiting this structure, the authors use a Lasso-based M-estimator to learn the system matrices ($A$ and $B$). This approach reduces the required sample complexity to $O(\log n)$, where $n$ is the system dimension, allowing for accurate model recovery from minimal data.
Robust Distributed Synthesis via SLS
Rather than optimizing the controller directly, which is often a non-convex and difficult problem, the framework utilizes System Level Synthesis (SLS). SLS reparameterizes the problem to optimize "system responses"—the closed-loop maps from disturbances to states and actions.
As shown in the figure above, the SLS framework allows for a structured realization of the controller using the learned system responses $\Phi_x$ and $\Phi_u$. This structure ensures that the resulting controller adheres to the desired communication constraints and stability requirements.
Tractable Surrogate Optimization
To make the optimization problem computationally feasible, the authors introduce several refinements:
- Finite Impulse Response (FIR) Restriction: The controller is restricted to a finite history window, transforming an infinite-dimensional problem into a finite-dimensional one.
- Robustness Terms: A robustness component is added to the objective function to account for worst-case errors in the identified model.
- Problem Decomposition: The large-scale optimization is decomposed into $n$ independent, small-scale Quadratic Programs (QPs) that can be solved in parallel, ensuring scalability.
Experimental Evaluation
The framework was evaluated using synthetic Graph Laplacian systems with chain structures, where each node's state is influenced only by its immediate neighbors. The experiments compared the proposed "Robust Distributed Controller" against "Nominal Distributed" (ignoring uncertainty) and "Nominal Centralized" (standard LQR) benchmarks across systems ranging from 20 to 150 states.
Key Results
- Stability under Uncertainty: In simulations with 10% model error, the robust controller maintained a 100% stability ratio. In contrast, nominal controllers often failed to stabilize the system as the number of control inputs decreased.
- Sample Efficiency: Near-optimal performance was achieved even with very short learning trajectories (e.g., 150 steps for a 40-state system), roughly proportional to the number of non-zero elements in the system matrices.
- Computational Scalability: The algorithm's runtime scaled near-linearly ($O(n^{1.004})$), confirming its suitability for very large-scale applications where traditional solvers fail.
Discussion and Future Work
The study highlights that sparsity is a critical lever for overcoming the dimensionality challenges in control. By assuming sparsity, the sample complexity is reduced from linear to logarithmic. The authors also provide a theoretical decomposition of the "optimality gap" into Uncertainty Error (from model identification) and Truncation Error (from the FIR approximation), proving that the former decays with model accuracy while the latter decays exponentially with controller length.
Future research directions include:
- Initial Stability: Addressing the requirement for an initial stabilizing controller to collect data safely.
- Weak Sparsity: Extending the theoretical proofs to systems where interactions decay over distance but are not strictly zero.
- Output Feedback: Adapting the framework for scenarios where only partial, noisy sensor data is available, which is common in practical robotics applications.
Summary updated