A hierarchical learning framework for generalizing tracking control behavior of a laboratory electrical system

A hierarchical learning framework (HLF) is validated on a rheostatic brake emulator called Electrical Braking System (EBS). The three-level learning starts with indirect closed-loop feedback linearization at level L1 by using input-output data samples collected from the EBS under exploration settings. A virtual state representation serves for feedback control in order to reach a linear model reference tracking, the solution being learned with a value iteration reinforcement learning approach. On top of the linearized closed-loop control system (CLCS), a secondary level L2 learning phase takes place, with an experiment-driven Iterative Learning Control (EDILC) used for learning reference input-controlled output pairs called primitives. The intent is to make the CLCS’s output have a shape with adequate approximation capacity. Learning is done by repetitions here, however, the final learning level L3 uses the level L2 learned primitives to predict the reference input ensuring optimal tracking of an unseen before desired trajectory, this time without repetitions. The proposed HLF displays features that are specific to intelligent organisms: memorization, learning and generalization of previously learned behavior.