Research Interests

As a broad generalisation it seems that directors of companies are keen to explore the benefits of artificial intelligence (AI). This is mainly because of the possibility of improving productivity or reducing staff numbers, by exploiting the abundance of data a company gathers that in the past was not used very effectively. It is very likely that the general management and processes of a business can be advanced by using such data at little extra cost. However, for control engineers the focus is more on improving the control of machinery or processes at a plant level. The question is therefore whether machine learning is likely to improve existing or future systems.

Going back in time we did of course live through the era of advances in adaptive control through techniques like model reference adaptive control and self-tuning control systems. The latter was deceptively simple to understand and involved the logical steps of identifying plant models and then using them repeatedly for control design. However, from a practical viewpoint only “auto-tuners” made a real impact, and adaptive control has mainly remained a topic of academic research, particularly when solutions for general control applications were sought. It is true that specialised tailored adaptive control solutions have been applied successfully, but the reliability of general algorithms has been questionable. Safe and robust adaptive control has remained an elusive concept. This is even though there is now a tidal wave of academic papers that use neural networks, deep learning, digital twins and data-driven control approaches. This suggests AI based methods will have an impact in control solutions, even if many contributions are based more on persuasive intuitive arguments rather than a rigorous theory.

We may now return to the question of whether machine learning can help. One of the great benefits of model based adaptive systems is that the likely behaviour and performance can be predicted. It is therefore possible to gain some confidence in how such algorithms will perform, admittedly often under quite restrictive assumptions. Unfortunately, an AI inspired black box data driven control solution does not provide the same intuitive understanding of operation. It does not provide engineers with the same confidence model-based methods provide to use such a solution since it is difficult to satisfy safety and other concerns. Machine learning solutions are not therefore acceptable in circumstances where valuable product could be lost, accidents could happen or erratic behaviour result. However, there is now some consensus in the control community that a combination of model-based, and data-driven ideas could be valuable and importantly provide a new impetus to the development of adaptive controls.

It seems a very reasonable assumption that by exploiting data using machine learning it should be possible to improve controller tuning and gain performance benefits. The concept applies to both classical controls, where say PID tuning of gains can be used, or more advanced methods such as predictive controls where improved prediction capabilities in nonlinear systems can be exploited. The conclusion at present is therefore that AI and machine learning methods do have something to offer the control engineer. Small cautious steps to improve existing systems using machine learning seems a sensible way forward. We may then get to the point where there are reliable data-driven adaptive controllers that can improve the quality of control and be used for a much wider range of applications, reducing design and tuning effort and limiting the deleterious effects of uncertainties on performance.

— Mike J Grimble