On the Use of Machine Learning and Search-Based Software Engineering for Ill-Defined Fitness Function: A Case Study on Software Refactoring

The most challenging step when adapting a search-based technique for a software engineering problem is the definition of the fitness function. For several software engineering problems, a fitness function is ill-defined, subjective, or difficult to quantify. For example, the evaluation of a software design is subjective. This paper introduces the use of a neural network-based fitness function for the problem of software refactoring. The software engineers evaluate manually the suggested refactoring solutions by a Genetic Algorithm (GA) for few iterations then an Artificial Neural Network (ANN) uses these training examples to evaluate the refactoring solutions for the remaining iterations. We evaluate the efficiency of our approach using six different open-source systems through an empirical study and compare the performance of our technique with several existing refactoring studies.