An improved learning algorithm for random neural networks based on particle swarm optimization and input-to-output sensitivity. (January 2019)

• Record Type: Journal Article
• Authors: Ling, Qing-Hua; Song, Yu-Qing; Han, Fei; Zhou, Cong-Hua; Lu, Hu
• Is Part Of: Cognitive systems research. Volume 53(2019)
• Page Start: 51
• Page End: 60
• Publication Date: 2019-01
• Subjects: Random neural networks -- Particle swarm optimization -- Extreme learning machine -- Input-to-output sensitivity; Cognition -- Periodicals; Cognitive engineering (System design) -- Periodicals; Artificial intelligence -- Periodicals; 153.05
• DOI: 10.1016/j.cogsys.2018.01.001
• Languages: English
• ISSNs: 1389-0417

Abstract

Abstract: Extreme learning machine (ELM) for random single-hidden-layer feedforward neural networks (RSLFN) has been widely applied in many fields in the past ten years because of its fast learning speed and good generalization performance. But because traditional ELM randomly selects the input weights and hidden biases, it typically requires high number of hidden neurons and thus decreases its convergence performance. It is necessary to select optimal input weights and hidden biases to improve the convergence performance of the traditional ELM. Generally, the single-hidden-layer feedforward neural networks (SLFN) with low input-to-output sensitivity will cause good robustness of the network, which may further lead into good generalization performance. Moreover, particle swarm optimization (PSO) has no complicated evolutionary operators and fewer parameters need to adjust, and is easy to implement. In this study, an improved ELM based on PSO and input-to-output sensitivity information is proposed to improve RSLFN's convergence performance. In the improved ELM, PSO encoding the input to output sensitivity information of the SLFN is used to optimize the input weights and hidden biases. The improved ELM could obtain better generalization performance as well as improve the conditioning of the SLFN by decreasing the input-to-output sensitivity of the network. Finally, experiment results on the regression and classification problems verify the improved performance the proposed ELM.