Feature Selection and Function Approximation Using Adaptive Algorithms

Abstract

Multiobjective heterogeneous flexible neural tree (HFNT) and multiobjective hierarchical fuzzy inference tree (HFIT) are two novel adaptive algorithms, which were proposed for the feature selection and function approximation after comprehensive literature reviews of the neural network and fuzzy inference system paradigms, respectively. The proposed algorithms were designed as a tree-like model, and the best tree-structure was selected from a topological space by applying a multiobjective evolutionary algorithm that simultaneously minimized both approximation error and tree complexity. Further, the parameter vector of the selected tree, from the Pareto front, was tuned by using a metaheuristic algorithm. For HFNT, the dynamics of natural selection was exploited to introduce functional heterogeneity in the HFNT nodes, and a diversity index was introduced for creating diverse HFNTs during its tree optimization phase. Subsequently, an evolutionary ensemble of HFNTs was proposed for making use of the final population. On the other hand, the HFIT nodes were low-dimensional type-1 or type-2 fuzzy inference systems, and the tree-like model was a hierarchical arrangement of such nodes. The performance of both HFNT and HFIT on benchmark datasets was better than the performance of the algorithms in the literature. Additionally, both HFNT and HFIT was used for the predictive modeling of the industrial problems, in which the feature selection was a crucial challenge in addition to the prediction. High approximation ability with the simple model generation is the vital contribution of the proposed algorithms for predictive modeling of complex problems.

Multiobjective heterogeneous flexible neural tree (HFNT) and multiobjective hierarchical fuzzy inference tree (HFIT) are two novel adaptive algorithms, which were proposed for the feature selection and function approximation after comprehensive literature reviews of the neural network and fuzzy inference system paradigms, respectively. The proposed algorithms were designed as a tree-like model, and the best tree-structure was selected from a topological space by applying a multiobjective evolutionary algorithm that simultaneously minimized both approximation error and tree complexity. Further, the parameter vector of the selected tree, from the Pareto front, was tuned by using a metaheuristic algorithm. For HFNT, the dynamics of natural selection was exploited to introduce functional heterogeneity in the HFNT nodes, and a diversity index was introduced for creating diverse HFNTs during its tree optimization phase. Subsequently, an evolutionary ensemble of HFNTs was proposed for making use of the final population. On the other hand, the HFIT nodes were low-dimensional type-1 or type-2 fuzzy inference systems, and the tree-like model was a hierarchical arrangement of such nodes. The performance of both HFNT and HFIT on benchmark datasets was better than the performance of the algorithms in the literature. Additionally, both HFNT and HFIT was used for the predictive modeling of the industrial problems, in which the feature selection was a crucial challenge in addition to the prediction. High approximation ability with the simple model generation is the vital contribution of the proposed algorithms for predictive modeling of complex problems.