Learning dynamical models using motifs

Automatically creating dynamical system models, M, from data is an active research area for a range of real-world applications, such as systems biology and engineering. However, the overall inference complexity increases exponentially in terms of the number of variables in M. We solve this exponential growth by using canonical representations of system motifs (building blocks) to constrain the model search during automated model generation. The motifs provide a good prior set of building blocks from which we can generate system-level models, and the canonical representation provides a theoretically sound framework for modifying the equations to improve the initial models. We present an automated method for learning dynamical models from motifs, such that the models optimize a domain-specific performance metric. We demonstrate our approach on hydraulic systems models.