Robust Intelligent Control

For artificial intelligence-based control systems to be entrusted with making real-time decisions in an industrial setting, AI methods must become more robust. Robustness is the ability to perform reliably under uncertainty, when disturbances occur, when the model is inaccurate, and when unusual events take place. The robustness of AI-based control methods is less critical in virtual settings where the cost of failure is low, compared to physical systems where failures are associated with risks to human life, property, and the environment. One way to achieve robust optimization is to rely on probabilistic forecasts, for example for the power production of wind farms. The uncertainty of such forecasts is high, mainly due to the uncertainty in weather predictions, but different forecasting models capture this uncertainty with varying degrees of accuracy. In addition, scientific research in this field lacks good benchmarks for the performance of different models, since performance tests of new models are often conducted on closed datasets, and the software code for the models is typically not shared. This makes it more difficult for third parties to compare their models against those proposed in the literature. To enable such a shared understanding of the state of the art, the project will create such benchmarks for wind power forecasting.

Robust is one of the three pillars of trustworthy AI according to EU’s High Level Expert Group on AI. The Norwegian government and the OECD also mention robustness as a key property for AI to be deployed in safety and business-critical applications. Robustness in this setting is the ability to keep performance and stability in the presence of uncertainty, disturbances, modelling errors, and out-of-distribution examples. The issue of robustness is particularly problematic in control of physical systems as it is associated with risks to human safety, assets, and the operating environment. Control systems manages and commands other devices or systems and are used in safety and business-critical applications. Intelligent control systems is the application of AI in control systems to achieve intelligent characteristics such as adaptation, learning and planning under uncertainty. Even though intelligent controllers have been shown to outperform traditional controllers in several domains, the industrial uptake is limited by the lack of robustness in the underlying methodology. Currently there is a gap between the research on very advanced methods tested in a controllable and predictable environment and the industrial settings where the environment is neither completely known nor controllable and information can be scarce, uncertain and of low quality. This gap can be closed by developing novel methods for robust intelligent control that are tested in both controlled environments and in industrial systems. RICO is a collaboration between researchers from the NTNU and SINTEF and the industry represented by TrønderEnergi, Ohmia Retail and Solution Seeker. New and robust methods for intelligent control and methods for evaluating their robustness will be developed and evaluated in industrial settings to meet the society’s and industry’s standards for deployment of trustworthy AI systems in industry.