Learning to extrapolate an optimal tracking control behavior towards new tracking tasks in a hierarchical primitive-based framework

The proposed hierarchical learning framework induces a generalized optimal tracking behavior for a control system. The L1 learning level ensures indirect closed-loop system (CLS) linearization using nonlinear state-feedback control with neural networks, based on a virtual state constructed from input-output samples of the assumed observable underlying controlled system. The linearized CLS behavior is learned to match a linear output reference model and prepares the subsequent L2 level learning of reference input–controlled output pairs called primitives, using an experiment-driven Iterative Learning Control. The learned primitives serve as a database of previous experiences which allow for the final L3 level to predict the optimal reference input which ensures that the CLS output perfectly tracks a new unseen trajectory, without relearning. With unknown dynamics of the underlying system, and being able to generalize a tracking behavior, the proposed hierarchical learning is validated on a multivariable hybrid electrical-software system. It shows traits of learning, memorization, generalizable behavior from past experience and adaptation for complex nonlinear systems, as some of the intelligent features required by Industry 4.0’s standards.