Evolutionary modelling of industrial systems with genetic programming

Knowledge, experience, and intuition are integral parts of decision making. However, these alone are not sufficient to manage today’s industrial systems. Often predictive models are required to weigh options and determine potential changes which provide the best outcome for a system. In this respect, the dissertation develops approximate models, metamodels, of industrial systems to facilitate a means to quantify system performance when the trade-off between approximation error and efficiency (time and effort spent on model development, validation, maintenance and execution) is appropriate. Discrete-event simulation (DES) is widely used to assist decision makers in the management of systems. DES facilitates analysis with high fidelity models as a consequence of its flexibility. However, this descriptiveness introduces an overhead to model building and maintenance. Furthermore, due to stochastic elements and the size of the systems modelled, model execution times can be computationally demanding. Hence, its use in operational tasks such as design, sensitivity analysis and optimisation can be significantly undermined when efficiency is a concern. In this thesis, these shortcomings are addressed through research into the use of genetic programming for metamodelling. Genetic programming is a branch of evolutionary algorithms which emulate the natural evolution of species. It can evolve programs of a domain via symbolic regression. These programs can be interpreted as logic instructions, analytical functions etc. Furthermore, genetic programming develops the models without prior assumptions about the underlying function of the training data. This can provide significant advantage for modelling of complex systems with non-linear and multimodal response characteristics. Exploiting these properties, the dissertation presents research towards developing metamodels of manufacturing systems (or their DES models) via genetic programming in the context of symbolic regression. In particular, it contributes to; (i) exploration of an appropriate experimental design method suitable to use with genetic programming, (ii) to a comparison of the performance of genetic programming with neural networks, using three different stochastic industrial problems to identify its robustness; (iii) research into an improved genetic programming and dynamic flow time estimation.