With the rapidly development of DNA sequencing and bioinformatics technique,a huge amount of high-dimensional single-nucleotide polymorphism (SNP) data is produced.Although many approaches for detection of the complex associations between SNPs and common human diseases in genome-wide association studies have been proposed,some of them only confine to explore on single genetic markers.Recently,an increasing number of studies have found that one of the most important factors for emergence and development of complex diseases is the interactions between SNPs,that is to say,epistasis or epistatic interactions.As a consequence,more and more studies caught great attention on epistatic interactions.In this study,a method named epiACO based on ant colony optimization algorithm is proposed for identifying epistatic interactions.In epiACO,a fitness function Svalue which combined mutual information with Bayesian networks is introduced for detecting epistatic interactions.Svalue has effectively solved the unicity of one evaluation measure and greatly improved the detection power of epiACO.Furthermore,a self-adaption adjustment parameter is designed to improve the processing capacity of models displaying no marginal effects in epiACO.Unlike the traditional process way of the identified solutions,a memory based strategy is designed to dispose the optimal solutions of epiACO,which records the optimal solutions to compare with the solutions that identified at current iteration and get final suspected solutions.The memory based strategy effectively improves the computational efficiency,enhances the processing ability of all optimal solutions and generates a more accurate way for detecting epistasis.Furthermore,a post-processing tactics is also employed to improve the power of detecting pure epistasis.Experiments of epiACO are compared with some other representative methods which are AntMiner,IACO,AntEpiSeeker and MACOED in both simulated and real age-related macular degeneration datasets.Results show that epiACO outperforms others in detection power for a large scale SNP datasets and might provides some significant clues on heuristics for inferring epistatic interactions.