An Improved Ensemble Extreme Learning Machine Based on ARPSO and Tournament-Selection
Extreme learning machine (ELM) performs more effectively than other learning algorithm in many cases, it has fast learning speed, good generalization performance and simple setting. However, how to select and cluster the candidate are still the most impor
- PDF / 269,711 Bytes
- 8 Pages / 439.37 x 666.142 pts Page_size
- 17 Downloads / 199 Views
School of Computer Science and Communication Engineering, Jiangsu University, Zhenjiang, Jiangsu, China [email protected] 2 School of Computer Science and Engineering, Jiangsu University of Science and Technology, Zhenjiang, Jiangsu, China
Abstract. Extreme learning machine (ELM) performs more effectively than other learning algorithm in many cases, it has fast learning speed, good generalization performance and simple setting. However, how to select and cluster the candidate are still the most important issues. In this paper, KGA-ARPSOELM, an improved ensemble of ELMs based on K-means, tournament-selection and attractive and repulsive particle swarm optimization (ARPSO) strategy is proposed to obtain better candidates of the ensemble system. To improve classification and selection ability in the ensemble system, K-means is applied to cluster the ELMs efficiently while tournament- selection is used to choose the optimal base ELMs with higher fitness value in proposed method. Moreover, experiment results verify that the proposed method has the advantage of being more convenient to get better convergence performance than the traditional algorithms. Keywords: Extreme learning machine Attractive and repulsive particle swarm optimization K-means Tournament-selection
1 Introduction Extreme learning machine (ELM) [1, 2] is proposed for single-hidden layer feedforward neural networks (SLFNs). Different from BP algorithm and other learning algorithm, it can simply set number and type of hidden layer neurons randomly without the iterative solution. Moreover, ELM does not need to adjust input parameters in the network, fast learning speed and better generalization performance are features and advantages of the ELM. Thus ELM has received increasing attention in recent years. Some studies show that extreme learning machine optimization using particle swarm algorithm (PSO) is very effective [3]. However, traditional PSO still has the disadvantages of premature convergence and easily fall into local minima [4]. To reduce the impact of this problem, attractive and repulsive particle swarm optimization (ARPSO) [5] was proposed to ensure the diversity of swarm effectively in the search process, so it could get better convergence accuracy. In the E-ARPSOELM [6], by considering the classification accuracy and diversity of the ensemble system, ARPSO was applied to select the base ELMs from the initial ELM pool. © Springer International Publishing Switzerland 2016 Y. Tan et al. (Eds.): ICSI 2016, Part II, LNCS 9713, pp. 89–96, 2016. DOI: 10.1007/978-3-319-41009-8_9
90
Y.-Q. Wu et al.
How to find suitable ensemble ELMs with more concise and compact methods are the jobs of this study. K-means algorithm [7] is still the one of the most widely used partitioning clustering algorithm in different areas of science. In this paper, K-means method is used to cluster the base ELMs in the first phase, which greatly increases the diversity between the different classes of samples. Then, tournament-selection is used to choose better members in
Data Loading...
