Improved optimization parameters prediction using the modified mega trend diffusion function for a small dataset problem

Abstract

This paper proposes a modified mega trend diffusion (MTD) function based on the K-means clustering algorithm to generate artificial samples for a training dataset. This would improve the prediction accuracy in a backpropagation neural network (BPNN) algorithm used in small dataset problems. The main contribution of this paper is in solving the attributes redundancy problem in an mega trend diffusion (MTD) function construction when there are two and three overlapped regions in the functions, using the K-means clustering algorithm. When used in predicting the parameters of an optimization algorithm, significant improvements in the prediction errors were observed, compared to the previous MTD method. The improvements were achieved by clustering the membership function (MF) for each attribute from the overlapped regions in the MF triangle. In this work, this algorithm is used to predict the control parameters of the artificial bee colony optimization (ABCO) (Ni and Li), which was then used in finding the optimal exit door locations of building layouts. For a case study, six samples of multi-room building layouts were considered. Each layout consists of information on the number of rooms (ni), room sizes (si) and corridor width (wi). The performance of the model was evaluated against the conventional MTD method. The superiority of the proposed method over the conventional MTD was confirmed by the 17.67% and 28.68% improvements in the prediction error for twofold and threefold cross-validations, respectively. It is envisaged that the method can be very useful in improving the prediction error of data samples of various scales and with different sizes of artificial data.