Scalable Synthesis and Verification of String Stable Neural Certificates for Interconnected Systems footnoteinfo

Jingyuan Zhou, Haoze Wu, Haokun Yu, Kaidi Yang · 2025

Read Scalable Synthesis and Verification of String Stable Neural Certificates for Interconnected Systems footnoteinfo by Jingyuan Zhou, Haoze Wu, Haokun Yu, Kaidi Yang (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2509-10118.txt.

Introduction

Jingyuan Zhou and 3 co-authors build stability proofs for networks of learned controllers in a 2025 control paper. Linked systems such as platoons need string stability. That is the promise that a wobble at the front fades down the line. Neural controllers perform well in such settings. They arrive as black boxes, though. So the promise had no formal backing.

Their frame joins two tools. One is a discrete-time form of scalable input-to-state stability. The other is neural network checking. Together they build certificates for the learned system. They then verify those certificates in a robust way. The check allows for gaps between true dynamics and the learned copy. Theory and methods stretch the pipeline to large systems. Fleet operators gain more than a controller. They gain a proof-shaped pledge that shocks die out.

Catalog record

CREATORJingyuan Zhou, Haoze Wu, Haokun Yu, Kaidi Yang
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE79 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2509-10118

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Scalable Synthesis and Verification of String Stable Neural Certificates for Interconnected Systems footnoteinfo Jingyuan Zhou, Haoze Wu, Haokun Yu, Kaidi Yang Abstract Ensuring string stability is critical for the safety and efficiency of large-scale interconnected systems. Although learning-based controllers (e.g., ones based on reinforcement learning) have demonstrated strong performance in complex control scenarios, their black-box nature hinders formal guarantees of string stability. To address this gap, we propose a novel verification and synthesis framework that integrates discrete-time scalable input-to-state stability (sISS) with neural network verification to formally guarantee string stability in interconnected systems. Our contributions are four-fold. First, we establish a formal framework for synthesizing and robustly verifying discrete-time scalable input-to-state stability (sISS) certificates for neural network–based interconnected systems. Specifically, our approach extends the notion of sISS to discrete-time settings, constructs neural sISS certificates, and introduces a verification procedure that ensures string stability while explicitly accounting for discrepancies between the true dynamics and their neural approximations. Second, we establish theoretical foundations and algorithms to scale the training and verification pipeline to large‑scale interconnected systems. Third, we extend the framework to handle systems with external control inputs, thereby allowing the joint synthesis and verification of neural certificates and controllers. Fourth, we validate our approach in scenarios of mixed-autonomy platoons, drone formations, and microgrids. Numerical simulations show that the proposed framework not only guarantees sISS with minimal degradation in control performance but also efficiently trains and verifies controllers for large‑scale interconnected systems under specific practical conditions. keywords Neural Certificates; Neural Network Verification; String Stability; Scalability. † † thanks: [ † † address: Department of Civil and Environmental Engineering, National University of Singapore † † address: Department of Computer Science, Amherst College † † address: Institute of Operations Research and Analytics, National University of Singapore footnoteinfo] Corresponding author Kaidi Yang. , , , 1 Introduction Large-scale interconnected systems haber2014subspace ; lyu2022small ; zhang2025privacy ; janssen2024modular arise in many real-world applications, such as power grids silva2021string , smart transportation wang2022optimal ; zhou2024parameter , and industrial process control zhu2002model . These systems have become increasingly complex and tightly coupled due to advances in sensing and communication technologies. A critical challenge in controlling interconnected systems is to ensure string stability , which prevents local disturbances from amplifying when propagating through the network feng2019string . In vehicle platoon systems, for example, the loss of string stability can cause minor speed fluctuations to escalate into stop-and-go waves, thereby degrading efficiency and compromising safety. Control strategies for enhancing string stability in interconnected systems can be divided into model-based and learning-based approaches. Model-based controllers, such as linear feedback control wang2021leading , model predictive control gratzer2022string ; aboudonia2025adaptive , and sliding mode control guo2016distributed , enable direct analysis of string stability using time-domain (Lyapunov-based) or frequency-domain (transfer function-based) techniques feng2019string . However, these model-based methods often rely on accurate system models and can become less effective or overly conservative when dealing with complex or uncertain dynamics in large-scale interconnected systems. To address these limitations, learning-based controllers, in particular, neural network-based controllers, have gained increasing popularity thanks to their ability to handle complex dynamics and uncertain environments in interconnected systems cheng2019end ; li2021reinforcement ; zhou2024enhancing . However, due to the black-box nature of neural networks, it is challenging to provide formal guarantees for string stability. For existing neural network-based controllers, string stability is often treated as a soft constraint in the training process (e.g., incorporated into the loss or reward function) without offering a rigorous theoretical guarantee. To the best of our knowledge, few studies have addressed the challenge of verifying and ensuring string stability in neural network-based controllers. The only notable attempt is the Estimation-Approximation-Derivation-Calculation framework proposed by Zhang et al. zhang2024string that approximates learning-based car-following controllers as linear models and then uses transfer function analysis to assess string stability. However, this approach suffers from significant approximation errors, making it unsuitable for rigorous string stability verification. It is worth noting that several works seek to provide formal guarantees for local stability rather than string stability. For instance, Dai et al. dai2021lyapunov propose a Lyapunov-based approach to ensure local stability by designing neural network controllers with a verifiable Lyapunov certificate. Building on this, Yang et al. yanglyapunov introduce a framework for verifying neural control under both state and output feedback, further extending Lyapunov-based guarantees. Moreover, Mandal et al. mandal2024safe ; mandal2024formally leverage Lyapunov barrier certificates to formally verify deep reinforcement learning controllers, demonstrating safe and reliable training for aerospace applications. Zhang et al. zhang2023compositional train compositional Lyapunov certificates for networked systems and introduce strategies to simplify the training process. However, these works on local stability cannot be readily extended to ensure string stability in interconnected systems. First, local stability guarantees that small perturbations around an equilibrium for an individual agent decay, whereas string stability ensures that disturbances do not amplify as they propagate along a chain of agents. Thus, local stability alone is insufficient for analyzing the cumulative effects of inter-agent disturbance propagation inherent in string stability. Second, local stability analyses typically focus on small, isolated systems, whereas string stability requires analysis of large-scale networks, presenting scalability challenges. Consequently, formally guaranteeing string stability within interconnected systems with neural network controllers remains an open and challenging problem. Statement of Contribution . To bridge the research gap, we propose a scalable verification and synthesis framework for learning a neural controller with a formal string stability guarantee. Our contributions are four-fold. First, we establish a formal framework for the synthesis and robust verification of discrete-time scalable input-to-state stability (sISS) certificates for neural network-based interconnected systems. Specifically, we extend the notion of sISS silva2024scalable to discrete-time systems, construct neural sISS certificates, and develop a robust verification procedure that guarantees string stability while explicitly accounting for discrepancies between the true dynamics and their neural representations. Second, we address the scalability challenge by developing theoretical foundations and algorithms that reuse neural sISS certificates across smaller or structurally equivalent systems, thereby significantly improving the efficiency of verification and training for large-scale interconnected systems. Third, we generalize the framework to systems with external control inputs, enabling the simultaneous synthesis and verification of neural certificates and controllers. Fourth, we further validate the approach in multiple interconnected systems, including mixed-autonomy platoons, drone formations, and microgrids. Simulation results demonstrate that our framework ensures discrete-time sISS in large-scale systems while preserving control performance. 2 Related Work This work contributes to two fields of study: (i) string stable control for interconnected systems and (ii) neural certificates and neural network verification. String Stable Control. String stability is an important property in the analysis of interconnected systems, which ensures that external disturbances do not amplify as they propagate through successive agents feng2019string ; rodonyi2019heterogeneous ; silva2024scalable . This concept has been extensively studied in various applications, such as vehicular platoons wang2021leading ; silva2021string2 ; vargas2025stochastic , drone formations riehl2022string , and microgrids silva2021string . Most existing works focus on analyzing string stability for model-based controllers feng2019string , such as linear control wang2021leading , model predictive control gratzer2022string , and H_{\infty} control kayacan2017multiobjective . In contrast, few works have investigated string stability for model-free controllers, particularly neural network-based controllers zhou2024enhancing . Among the few existing works, Zhang et al. zhang2024string relies on linearization and may make inaccurate conclusions,

Read the complete text