Enhancing a GA-based BPN forecasting model by employing the Taguchi method

Back-propagation networks (BPNs) are now the most commonly employed form of artificial neural network (ANN) and are used to solve a large number of practical problems. However, BPNs have a tendency to become trapped at local minima when applied to forecasting problems. This tendency causes the forecasting results of the BPN to be inconsistent and unpredictable. The use of heuristic techniques such as genetic algorithms (GAs) to optimise the BPN has been proposed as a potential means to address this problem. GAs have been successfully employed to establish network parameters and topology settings which optimise BPN performance. However, BPNs and the GAs themselves have many undetermined network parameters and topology settings, and the impact and interactions of these controllable factors are not yet fully understood. A greater understanding of these issues would enable decision-makers to establish a feasible solution-searching region with an appropriate network topology, thereby increasing the quality of their decisions. Therefore, a requirement exists to develop a methodical approach for identifying these parameter values. Accordingly, this study adopts the Taguchi method to calibrate the controllable factors of a GA-based BPN forecasting model. An L8(27) inner orthogonal array is constructed for the controllable factors relating to the BPN's parameters and topology. Similarly, an L4(23) outer orthogonal array is used to arrange the GA parameters, which represent noise factors in the present study. The solutions obtained from the proposed GA-based BPN model for chaotic time series problems are compared with those presented by a previous study for both in-sample and interpolation data, and are found to be in good agreement.