Classification by induction: application to modelling and control of non-linear dynamical systems

Abstract

The modelling and identification of non-linear dynamical systems are considered in this paper. The emulation of an existing controller, a skilled human for example, is a special case of this general treatment. A technique is sought, capable of developing general black-box non-linear models with both numerical and symbolic data. The models themselves are expressed in a high-level human-understandable format and are induced from examples of past behaviour. In the case of human controllers, this approach removes reliance on the articulation of skilled behaviour. The studied approach is based on the automatic induction of decision trees and production rules from examples; these are particular cases of classifiers. The algorithms used are a product of the machine learning sub-field of artifical intelligence research. A formalism is developed whereby the modelling and control of general dynamical systems are transformed to classification problems, and therefore become amenable to processing by the induction algorithms mentioned above. Experimental results are presented describing the induction of executable models, both of skilled human control behaviour and of an exising automatic controller. Experiments were performed in simulations and on physical laboratory apparatus.