Search

Feature selection plays an important role in data preprocessing. The aim of feature selection is to recognize and remove redundant or irrelevant features. The key issue is to use as few features as possible to achieve the lowest classification error rate. This paper formulates feature selection as a multi-objective problem. In order to address feature selection problem, this paper uses the multi-objective bacterial foraging optimization algorithm to select the feature subsets and k-nearest neighbor algorithm as the evaluation algorithm. The wheel roulette mechanism is further introduced to remove duplicated features. Four information exchange mechanisms are integrated into the bacteria-inspired algorithm to avoid the individuals getting trapped into the local optima so as to achieve better results in solving high-dimensional feature selection problem. On six small datasets and ten high-dimensional datasets, comparative experiments with different conventional wrapper methods and several evolutionary algorithms demonstrate the superiority of the proposed bacteria-inspired based feature selection method. [ABSTRACT FROM AUTHOR]

Abstract: This paper describes a novel Bacterial Foraging Optimization (BFO) approach to multi-objective optimization, called Multi-objective Bacterial Foraging Optimization (MBFO). The objectives in the Multi-objective Bacterial Foraging Optimization are maintained by a fitness survive mechanism. Bacteria with the smaller health values have the better chance to survive. Meanwhile, the main goal of multi-objective optimization problems is to obtain a superior non-dominated front which is closed to the true Pareto front. With identification of such features, the idea of integration between health sorting approach and pareto dominance mechanism are developed to search for Pareto-optimal set of problems. Moreover, strategy keeping a certain unfeasible border solutions based on a given probability is considered to improve the diversity of individuals. In addition, two different performance metrics: Diversity and Generational Distance are introduced as well to evaluate multi-objective optimization problems. Compared to two other multi-objective optimization evolutionary algorithms MOPSO and NSGA-II, simulation results show that in most cases, the proposed MBFO is able to find a much better spread of solutions and convergence to the true Pareto-optimal front faster. It suggests that MBFO is very promising in dealing with ordinary multi-objective optimization problems. [Copyright &y& Elsevier]