Automatic Generation of Specification-Based Test Cases by Applying Genetic Algorithms in Reinforcement Learning

Abstract

Automatic test pattern generation by applying genetic algorithms in reinforcement learning is proposed and the results of evaluation with an ATM system which assuming the existence of a bug related to a global variable are reported. For software development based on formal specifications, a number of methods for the automatic generation of test cases have previously been proposed. However, most of those methods are for testing whether or not the specifications are correctly implemented, and the problem of automatically generating test cases for which all of the paths can be traversed, including paths that were not anticipated, remains unsolved. We have previously demonstrated the feasibility of using genetic algorithms as an effective approach to that problem by evaluation with a simple test problem that assumes a single-variable input and evaluation assuming that the total number of paths is known. Here, we use a genetic algorithm in which the test cases are used as the gene locus values in the chromosome with redundant gene length and the difference vector is used for performing mutation. We compare this approach to the pairwise method and the vibration method, which are leading research areas, in a more realistic testing environment such as evaluation of ATM system programs. We show that the proposed method can provide higher program path coverage than the previous methods and effectively generates test cases for a bug related to a global variable.