Safe and Guaranteed Machine Learning for Control

Modern autonomous systems such as autonomous vehicles, robotic manipulators, surgical robots, and power networks, increasingly rely on machine learning to handle complex dynamics and adapt to uncertain environments. However, learning-based components are typically “black boxes,” making it challenging to provide the formal guarantees required in safety-critical applications. Our research develops verified learning-based control methods that integrate machine learning, control theory and formal methods. We build data-driven symbolic abstractions with correctness guarantees, learn safe controllers and design scalable tools that ensure robustness even in the presence of uncertainty, disturbances, and model mismatch. Ultimately, our goal is to make learning-enabled controllers certifiable, trustworthy, and deployable in real engineering systems.

• Learning the System Dynamics with Guarantees

In many engineering systems, the underlying dynamics are partially unknown, highly nonlinear, or too complex to model accurately. Our work uses machine learning to learn or approximate these dynamics while preserving formal safety guarantees. We develop verified, data-driven symbolic abstractions with statistical correctness bounds, learn safe model approximations using tools such as Gaussian processes, and design scenario-based methods that remain robust to modeling uncertainty. Altogether, this research builds a unified framework for learning system dynamics safely, ensuring that every data-driven component comes with quantifiable guarantees for control design.

• Learning the Controller with Guarantees

When controllers are learned, such as through reinforcement learning or neural network training, we develop methods to ensure that the resulting closed-loop system remains safe and robust. Our work provides formal verification tools for neural network controllers, integrates symbolic control with reinforcement learning to handle continuous state–action spaces safely, and develops learning-based abstractions that support certified policy synthesis for nonlinear systems. This research establishes a principled foundation for safe, trustworthy learning-based controllers, enabling the deployment of RL and neural policies in safety-critical cyber-physical systems.

• K. Hashimoto, A. Saoud, M. Kishida, T. Ushio and D. V. Dimarogonas, “Learning-based symbolic abstractions for nonlinear control systems ,” Automatica, 2022 Dec 1. vol. 146, p. 110646. 

• L. Jouret, A. Saoud and S. Olaru, “Safety Verification of Neural-Network-based Controllers: A Set Invariance Approach,” IEEE Control Systems Letters, 2024.

• A. Devonport, A. Saoud and M. Arcak, “Symbolic Abstractions From Data: A PAC Learning Approach,” IEEE Conference on Decision and Control (CDC), 2021 (pp. 599-604).

• S.B. Alaoui, A. Saoud, “How to discretize continuous-state action spaces in Q-learning: A symbolic control approach,” IEEE Conference on Decision and Control (CDC) 2024 (pp. 8314-8319). 

• A. Saoud and R. G. Sanfelice, “Computation of Controlled Invariants for Nonlinear Systems: Application to Safe Neural Networks Approximation and Control,”  IFAC Analysis and Design of Hybrid Systems (IFAC ADHS), IFAC-PapersOnLine, 2021 Jan. vol. 54, no. 5, pp.91-96. 

• B. Altiner, A. Saoud, A. Caldas, M. Makarov, “Scenario Convex Programs for Dexterous Manipulation under Modeling Uncertainties,” IEEE International Conference on Automation Science and Engineering (CASE) 2024 (pp. 3490-3497). 

• B. Jeloka, F. Nicolaau, A. Saoud, R. N. Banavar, “Data-Driven Control of Adaptive Cruise Control Systems Using Differential Flatness and Gaussian Processes,” IEEE American Control Conference (ACC) 2024 (pp. 5094-5099). 

• Z. Benhamidouch, S.B. Alaoui, A. Abbou and A. Saoud, “A Q-Learning Approach to Model-Free Infinite Horizon Control for Linear Time Delay Systems,” IEEE Conference on Decision and Control (CDC) 2024 (pp. 3335-3340).