Research And Applications Of Information Entropy And Particle Swarm Optimization Based Extreme Learning Machine

Neural networks are a black-box model. There is no need to know the premise structure of the black box and only need to adjust the relationship between the network nodes, and then it is possible to obtain better performance in data processing. Therefore, neural network methods are widely used in various aspects among the fields of modeling. Extreme Learning Machine was proposed in2004. Due to its extremely fast training speed and better generalization performance, the extreme learning machine attracts much concern in recent years.When dealing with high-dimensional datasets, the ELM has some problems. The high-dimension of the data and the discrete data points would result in enlarging the model and reducing the model generalization performance. In order to solve this problem, this paper proposed a method using information entropy and particle swarm algorithm to optimize extreme learning machine. The proposed model was used to some regression problems and practical application, and the experimental results verify the feasibility of the algorithm. The main research contents are as follows: 1. Firstly, an overview of neural networks, entropy and swarm optimization was introduced. Then an information entropy based algorithm for enhancing extreme learning machine was proposed. In the proposed method, the mutual information was adopted to filter the input variables and delete the independent and weak correlation variables. Then the entropy method was used to optimize the weights of the input variables, avoiding the discrete data points to affect the accuracy of the network training. The feasibility of the method was proved by taking experiments on3UCI datasets.2. This paper presents an optimization particle swarm algorithm with sequential quadratic programming and tent chaotic map. The sequential quadratic programming has an effective optimization performance locally and the tent chaotic map owned a better ergodicity. So these two methods were combined to improve the performance of PSO. The improved PSO algorithm not only has better optimization ability, but also can avoid local optimum when the prematurity of groups occurs. Finally, the feasibility of the algorithm was validated by three common Benchmark functions.3. We used the improved PSO algorithm to optimize the extreme learning machine. The weights of extreme learning machine input layer were assigned randomly, which caused the model unstable. So the improved PSO was used to optimize the input weights for enhancing the performance of extreme learning machine. Finally, the feasibility of the algorithm was examined by taking experiments on3UCI datasets. 4. The improved method was applied to PTA chemical process modeling. This method not only reduces the data dimension, but also improves the modeling accuracy, which verified the feasibility of this proposed method. And the method can provide some guidance for actual chemical process modeling.