Multiple Parameter Control for Ant Colony Optimization Applied to Feature Selection Problem
Gang Wang,HaiCheng Eric Chu,Yuxuan Zhang,Huiling Chen,Weitong Hu,Ying Li,XuJun Peng
DOI: https://doi.org/10.1007/s00521-015-1829-8
2015-01-01
Neural Computing and Applications
Abstract:The ant colony optimization algorithm (ACO) was initially developed to be a metaheuristic for combinatorial optimization problem. In scores of experiments, it is confirmed that the parameter settings in ACO have direct effects on the performance of the algorithm. However, few studies have specially reported the parameter control for ACO. The aim of this paper was to put forward some strategies to adaptively adjust the parameter in ACO and further provide a deeper understanding of ACO parameter control, including static and dynamic parameters. We choose well-known ant system (AS) and ant colony system (ACS) to be controlled by our proposed strategies. The parameters in AS and ACS include β , pheromone evaporation rate ( ρ ), exploration probability factor ( q 0 ) and number of ants ( m ). We have proposed three adaptive parameter control strategies (SI, SII and SIII) based on fuzzy logic control which adjusts ρ , q 0 and m , respectively. The feature selection problem is considered for evaluating the parameter control strategies. In addition, because AS and ACS are not intrinsically fit for feature selection problem, we have modified the AS and ACS, which are named as fuzzy adaptive ant system (FAAS) and fuzzy adaptive ant colony system (FAACS), to make them more suitable for feature selection problem. Because only one parameter is allowed to be dynamically adjusted in FAAS or FAACS, the remaining parameters should be statically specified. Thus, we have developed parametric guidelines for proper combination of static parameter settings. The performance of FAAS and FAACS is compared with that of the AS-based, ACS-based, particle swarm optimization-based and genetic algorithm-based methods on a comprehensive set of 10 benchmark data sets, which are taken from UCI machine learning and StatLog databases. The numerical results and statistical analysis show that the proposed algorithms outperform significantly than other methods in terms of prediction accuracy with smaller subset of features.