The evoluation of 'Boxes' to quantized inductive learning: a study in inductive learning

An inductive learning method is analyzed for use in on-line control. The controller has the benefit of being designed without a system model and is able to adapt itself to varying system parameters. Numerical experiments were performed with the Quantized Inductive Learning (QIL) algorithm, an extension of ‘Boxes’, on simple linear systems and a simulated simply supported aluminum beam.

Concurrent with the simulations, a theoretical analysis of the learning mechanism was generated. The evaluation of several issues with the algorithm (performance indices, sampling periods, and level of quantizations) were studied to validate the theory. A comparison with state feedback was used to compare the effectiveness of this method with traditional model-based approaches. The results indicate the method learns a control function which moves the system from an arbitrary initial condition to equilibrium or rejects a sinusoidal disturbance. In both cases, the control is learned in absence of an a priori system model.