Immunised neurocontrol

摘要

This paper addresses the possibility of using artificial neural networks, together with concepts from the field of immunology, in the modeling and adaptive control of complex dynamic systems.Immunology is the science of the in-built defense mechanism that is present in all living beings to protect them against external attacks. A biological immune system can be thought of as a very robust, adaptive system that is capable of dealing with an enormous variety of disturbances and uncertainties. Biological immune systems use a finite number of discrete ‘building blocks’ to achieve this adaptiveness. These building blocks can be thought of as pieces of a puzzle, which must be put together in a specific way to neutralise, remove, or destroy each unique disturbance the system encounters. This paper outlines a technique which attempts to reproduce the adaptiveness of a biological immune system in an artificial neural network by identifying and processing artificial neural network building blocks. The neural network building blocks are identified off-line using back-propagation of error and are processed on-line using a modified genetic algorithm. The realisation of this new approach to adaptive neurocontrol [coined as immunised neurocontrol (INC)] is documented and results of the implementation of this procedure in adaptive control of an uncertain UH-1 longitudinal helicopter hover model are presented.