Development and testing of meta heuristic regression tool


7 th International Conference on Advanced Technologies (ICAT’18), 28 April - 01 May 2018

  • Publication Type: Conference Paper / Full Text


One of the drawbacks of artificial intelligence (AI) based modelling approaches is that they do not have an open and simple mathematical model. So it is difficult to understand and apply AI-based models in real-world applications. In addition, AI-based models often complicate the problems that can be modelled with simple equations. In some studies in the literature, it is seen that regression-based models produce much faster and more effective solutions than AI-based models. Moreover, it is possible to produce equations that can be applied simply and quickly for the problems with regression. Although regression studies offer fast and simple models, there are some limitations to AI-based models. For example, as the number of parameters and the level of complexity of a problem increase, it becomes difficult to obtain a regression equation suitable for this problem. Because the majority of regression tools are based on gradient descent-based approaches. This approach is unsuccessful against local minimum traps in situations where search space is complex. Only simple regression equations can be produced with regression-software tools used for this task. In this paper, a meta-heuristic based regression tool is designed and developed. The designed tool minimizes the problems encountered in gradient descent-based approaches. It becomes possible to obtain effective regression equations for problems with a high number of parameters and a high level of complexity. In the meta-heuristic curve fitting process, four different regression equations are optimized for each problem. The optimum values of the parameters of these equations are discovered by two modern meta-heuristic search methods such as symbiotic organism search and lightning search algorithm. The developed regression tool has been applied to different problems and tested. For the test, the regression problems in the UCI Machine Learning Data Repository are used. The results from experimental studies show that the designed software tool explores optimum equations for regression problems.