The reason for the present upsurge of interest in intelligent control is that the present generation of control systems are incapable, to a greater or lesser extent, of dealing with problems of a certain complexity. Fortunately, human operators (HO) are often experts in keeping the complex control systems on the right track. In this paper a method for controller design has been investigated, based on a concept of developing a mathematical model for the behaviour of the HO of the process. This method treats the HO's behaviour as a dynamic process in itself, transformed from the dynamics of the unknown process to be controlled. The following three phases are distinguished in controller design: (i) Observation phase, during which the HO controls the process by himself for a specific time interval. During this time the data representing the state of the process and the corresponding human controller actions are registered. (ii) Modelling phase, during which the mapping between the inputs and the outputs of the controller is learned. The model used is a neural network with radial basis functions, and the estimation of the model parameters is conducted by using a multiparameter leastsquare estimation. (iii) Testing phase, during which the human operator behaviour model or the derived neural controller is used to control the process. The controller performance is evaluated by analysing its behaviour under both the same conditions used in learning phase, and completely different conditions to study the controller reliability.
The method has been successfully applied to control a nonlinear level-control process using computer simulation.