Search

In this thesis a heuristic controller for dynamic processes is presented whose control policy is able to develop and improve automatically. The controller's heuristics take the form of a set of linguistic statements or rules which can be expressed quantitavely by using the theory of fuzzy sets. This type of controller has particular application in the control of complex processes which has traditionally been achieved by using a human operator or implementing his heuristic control protocol on a digital computer. Because of the inconsistency of a human operator's decision process and the difficulty in obtaining an adequate control policy a controller is developed which is able to generate and modify the control heuristics automatically as a result of monitoring the process' performance. In this way the controller is able to organize itself with respect to the process it is trying to control so as to eventually produce a stable or converged control policy. The validity of such a scheme has already been shown although its characteristics had not been fully investigated. A general theory of such a controller is developed and its practical implementation on a digital computer described in relation to single-input single-output processes. The theory is then generalized to encompass multivariable processes A series of experiments are described and carried out which set out to study the controller's properties especially with a view to its sensitivity, convergence qualities and range of application. The conclusions drawn from the results of these experiments vindicate such a general approach to machine learning of heuristics and demonstrate the scheme's robustness and limits of application. The behaviour of the controller with multivariable processes which the results revealed were of particular interest. The results are analyzed and discussed and further avenues of research emanating from this work are considered.