A study conducted by Yuji Ito in collaboration with Kyoto University was published in IEEE Transactions on Automatic Control.
As robots and automated systems become increasingly sophisticated, technologies for controlling complex behavior under changing conditions are growing in importance. In recent years, researchers have explored the use of machine learning to learn system characteristics from real-world data and enable more advanced control. For such control systems to be used reliably, however, their stable operation must be theoretically guaranteed within a specified range of conditions.
A system can have an enormous number of possible operating states, making it impractical to verify stability for every state through direct computation. Another challenge is that machine learning models, including neural networks and Gaussian processes, have different structures, making it difficult to evaluate a wide variety of models using a common approach.
In this study, the researchers developed a method for guaranteeing stability across an entire target region based on calculations performed at a finite number of representative states. The method decomposes complex models into basic components, such as addition, multiplication, and nonlinear transformations, and combines the mathematical properties of these components. This framework enables control systems incorporating various machine learning models to be analyzed through a unified procedure. It can also be applied not only to stability analysis but also to controller design that improves control performance while guaranteeing stability.
The proposed method is expected to help improve the reliability of robots, automated systems, and other control systems that use machine learning to represent complex phenomena while also requiring stable operation.
Title: Second-Order Sampling-Based Stability Guarantee for Data-Driven Control Systems
Authors: Ito, Y., Fujimoto, K.
Journal Name: IEEE Transactions on Automatic Control
Published: August 31, 2026