Evolutionary computation for model structure selection in system identification

Abstract

System identification is a field of study involving the derivation of a mathematical model to explain the dynamical behaviour of a system. One of the steps in system identification is model structure selection which involves the selection of variables and terms of a model. Several important criteria for a desirable model structure include its accuracy in future prediction and model parsimony. A parsimonious model structure is desirable in enabling easy control design. This research explores the use of Evolutionary Computation (EC) in model structure selection. The effectiveness of penalty function in the objective function of EC is investigated. The results show that a suitable penalty function parameter can be achieved by its relation to the smallest estimated and tolerable parameter value. Using this function, an algorithm named Modified Genetic Algorithm (MGA) is proposed as it is able to reduce the possibility of premature convergence. MGA is proven to be more efficient than the original genetic algorithm where it is able to find a parsimonious model within a fixed or even shorter evolution period. Another algorithm, named Deterministic Mutation Algorithm (DMA) is proposed to reduce computational burden and reliance on optimum algorithm parameter setting. DMA is a simpler procedure that is able to assist user to obtain a parsimonious model within a shorter time. All of these system identification techniques are carried out by applying the algorithms to a number of simulated and real-life systems, namely gas furnace, Wölfer sunspot and hairdryer, using discrete-time models. Validations of the model structures are made using correlation tests and cross-validation.