Genetic programming (GP) is an evolutionary algorithm that attempts to evolve solutions to a problem by using concepts taken from the naturally occurring evolutionary process. This thesis introduces the concepts of GP model development by applying the technique to steady-state model evolution. A variation of the algorithm known as the multiple basis function GP (MBF-GP) algorithm is described and its performance compared with the standard algorithm. Results show that the MBF-GP algorithm requires significantly less computational effort to evolve models of comparable accuracy to the standard algorithm. The steady-state algorithm is then modified to enable the evolution of dynamic process models. Three case studies are used to demonstrate algorithm performance and show how the MBF-GP algorithm produces performance benefits similar to those observed in the steady-state modelling work. A comparison with neural networks reveals that GP is able to match the accuracy of the network predictions but is more expensive computationally. However, a significant advantage of the GP algorithm is that it can automatically evolve the time history of model terms required to account for process characteristics such as the system time delay.The model development process is not simply a case of reducing the error between the predicted and actual process output. The parallel nature of GP means that it is ideally suited to solving multi-objective problems. The MBF-GP algorithm is modified to incorporate a Pareto based ranking scheme that allows models to be compared using multiple performance criteria. The ranking scheme allows preference information in the form of goals and priorities to be specified in order to guide the search towards the desired region of the search space. Two case studies are used to demonstrate the performance of this technique. The first example uses the multi-objective algorithm to improve the parsimony of the evolved model structures. The second example demonstrates how a set residual correlation tests can be combined and used as an additional performance measure. In each case, the multi-objective algorithm performs significantly better than the single objective version. In addition, the inclusion of preference information overcomes some of the difficulties associated with conventional Pareto ranking and produces a greater number of acceptable solutions. Prefaceii

No abstract

No abstract

No abstract

No abstract