To main content

Machine learning in the loop

Control systems are in general based on the same structure (feedback loops) and physics-based models of the dynamic system regardless of application.

Based on model and parameter estimation, a control system is designed to give the system desired properties w.r.t stability, robustness etc. under reasonable assumptions. Developing such models and estimating the relevant parameters is time-consuming, expensive and in some cases infeasible due to unobservability and highly non-linear effects.

Today there is a gap between control theory developed in academia and what is currently applied in industry. Basic control methods such as PID-controllers are independent of system knowledge and are relatively straightforward to implement, but are often not sufficiently accurate for nonlinear processes. On the other hand, more complex approaches such as feedback linearizing and model predictive controllers are based on system identifications, modelling and full state knowledge, and provide analytical guarantees regarding stability, robustness, convergence, etc. However, the more complex methods are more challenging to implement due to unmeasurable states, highly nonlinear effects and unknown parameters.

The intersection between control theory and ML is an emerging research field with large potential for all types of control systems. ML consists of data-driven, flexible and adaptable methods that show improved performance and ability to recognize patterns and connections beyond that of traditional methods in areas such as image analysis and classification, natural language processing and medical diagnosis. We aim to benefit from the strengths of both approaches by combining ML and traditional control principles in a hybrid analytics control system which is both data-driven and physics-based

Explore research areas