An adaptive fuzzy ant colony optimization for feature selection

作者：Liu Yuanning; Wang Gang; Chen Huiling; Zhao Zhengdong; Zhu Xiaodong^*; Liu Zhen

来源：Journal of Computational Information Systems, 2011, 7(4): 1206-1213.

• Adaptive fuzzy
• Adaptive parameters
• Ant-colony optimization
• Benchmark data
• Classification performance
• Data feature
• Evaporation rate
• Exploration and exploitation
• Feature selection
• Feature selection problem
• Feature subset
• Fuzzy controllers
• Little research
• Local optima
• Machine-learning
• Optimization problems
• Prediction accuracy
• Premature convergence
• Problem size
• Running time
• Search spaces
• Searching spaces
• Standard pattern
• Static parameters
• Statlog

Standard pattern classifiers perform on all data features. Whereas, some of the features are redundant or irrelevant, which reduce prediction accuracy, and increase running time of classifier. The purpose of this study is to search optimal feature subset, in order to increase the classification performance. The feature selection problem differs from traditional optimization problem on the problem size, namely the searching space or the problem dimension. Thus it is easy to converge at a local optimum by means of the ant colony optimization (ACO) based method with static parameter settings, but the adaptive parameter control strategy can balance between exploration and exploitation in the search space to intelligently improve solutions and avoid premature convergence. Only little research has been reported to adjust parameters of ACO adaptively. This paper proposes a novel adaptive fuzzy ant colony optimization (AFACO) based method for feature selection. We design two fuzzy controllers to adjust parameters of ACO, namely pheromone evaporation rate and the number of ants. The performance of the proposed method is compared with that of the Standard ACO based, Standard particle swarm optimization (PSO) based and genetic algorithm (GA) based methods on 10 benchmark data sets, taken from UCI machine learning and StatLog databases. The experimental results show that the proposed method outperforms the other methods in prediction accuracy, and is efficient and effective for feature selection problem.

• 出版日期2011