Hybride AI – Less data, better models
Two keywords are currently on everyone’s lips: Artificial Intelligence (AI) and the idea of the hybrid, which unites different things. Bringing the two together describes a hitherto largely unknown field of research: Hybrid AI. It helps to find answers to questions for which solutions have been difficult to find so far. IAV combines physical systems with artificial neural networks (KNN) and achieves outstanding results – not only in the automotive sector.
«We enrich our models with expert knowledge, in this way, we achieve better models with the same amount of data or very good models with less data.»
— Research Engineer at IAV-Lab and DFKI
Better to generalize
“AI has the core problem, however, that data can only represent reality under certain conditions,” explains Kruschel. “As a result, the data situation is often not sufficient to generalize well.” Generalize means that the network receives completely new data from the same application area, but that describes a different scenario. Hybrid AI provides a solution for this. The term “hybrid” means “crossed, mixed.” More specifically, this means: “We enrich our models with expert knowledge,” explains Ferdinand Küsters, Research Engineer at IAV-Lab and DFKI. “In this way, we achieve better models with the same amount of data or very good models with less data.” If the network can access structural information or, in other words, physics-based data preprocessing, the model also performs better on new data.
Best of both worlds: Control engineering and Artificial Intelligence
A pre-selection of input variables is the easiest way to integrate expert knowledge into a model. This procedure has long been standard in data-based modeling. The hybrid AI approach goes beyond this: Here, physical models are integrated directly into the data-driven model and adapted with AI support. This allows new components to be developed quickly, for example for the Digital Twin. In controller design, IAV combines neural networks and differential equations in a multi-stage process – with the result that the controllers designed with it generalize better than classical neural networks. Such an approach is also helpful, for example, in computationally complex flow simulations. Küsters explains: “The problem was solved with NeuralODEs, a quite new form of neural networks, which internally use differential equations for modeling.” Therefore they are well suited for physical processes.
With hybrid AI, IAV specialists are finding the best solutions to problems in almost all areas ever more quickly, as Kruschel explains. Küsters describes it in more detail: “If, for example, you bring in the structure of a robot, i.e. information about its rough design, you can achieve greater accuracy more quickly.” Of course, this also applies to the automotive sector, where IAV’s core expertise lies: The problem with predicted cylinder pressure mentioned at the beginning has long since been solved – thanks to hybrid AI.
The article was published in automotion 03/2020, the automotive engineering magazine of IAV. Here you can order the automotion free of charge.