A study on Ant Miner parameters

Ant Miner is the application of Ant Colony Optimization algorithm in data classification problem. Since it was proposed by Parpinelli et al., in 2001, a lot of data classification studies were performed with the aid of this algorithm. Also a lot of comparisons were performed between the results obtained with this algorithm and the results obtained by improved Ant Miner algorithms. Usually, in different studies different values were chosen for the input parameters such as the number of ants, the minimum number of instances covered by each rule, and so on. So we ask the question how to choose the values of input parameters in order to obtain good results for output parameters like prediction accuracy, the number of discovered rules and the execution time? In this paper we run on 15 datasets obtained from UCI Machine Learning Repository a number of 32 combinations of input parameters, and we try to find connections between the values of the input parameters and the obtained results. Finally we wrote a set of remarks regarding Ant Miner parameters, which may be useful in choosing input parameters in different studies in order to obtain overall good results. Before starting the experiments we performed a small modification to the Ant Miner open source code in order to assure that each run will work with the same subsets of data, for a fair comparison of results.