Hybrid Machine Learning Control for Safe System Exploration

The model is segmented into two distinct components: a physics-based model that ensures system safety by respecting physical constraints, and a data-driven model that maximizes information gain from measurements. This segmentation allows each component to specialize - the physics model prevents unsafe states while the data model explores new information, resolving the contradiction between information gain and safety

The method changes the parameter representation by learning parameters of the physics-based model from measurement data while maintaining the structural integrity of physical laws. This allows the model to adapt to new information without violating safety constraints, enabling exploration within safe boundaries

The physics-based model serves as an intermediary between noisy measurements and the learned parameters. Instead of directly learning from noisy data, the method uses the physics model as a mediator that filters and structures the information, allowing adaptation to new data while maintaining precision through physical constraints

The method performs preliminary action by incorporating physical knowledge and constraints into the model structure before learning from data. This pre-structuring with physics-based relationships prepares the model to handle noisy measurements more effectively, improving both adaptability and precision