Test Data Generation For Path Coverage Based On Neural Network

The purpose of software testing is to improve the quality of software by methodof finding defects or even wrong in software. Existing statistical results showed that itcost more than50%of software development.The key of software testing isautomatic generation of test data and path coverage testing is an important tool togenerate test data automatically. The traditional test data generation for path coverageis transforming data generation problem as an optimization problem, for each test data,the program should obtain some information such as brunch distance and approachlevel by running instrumentation program so that we can calculate its fitness value,but it cost a lot of time. To reduce the computational cost and time cost, this papermade the following research.The third chapter of this paper put forward a method of test data generation forpath coverage based on BP neural network. In particular, BP neural network was usedto simulate the calculation process of the instrumentation program. First, randominput data should be running in the instrumentation program to obtain someinformation such as brunch distance and approach level so that we can calculate itsfitness value; then the input data and fitness value would be used to train BP neuralnetwork; at last, the test data could be generated by BP neural network and geneticalgorithm. The experiment proved that good test data could be generated by the BPneural network.Due to the characteristics of BP neural network, such as slower convergencespeed and the possibility of failure, the forth chapter of this paper put forward animproved method of test data generation for path coverage based on RBF neuralnetwork. The adding and deleting strategies were used to insure the high performanceof hidden layer in hidden layer structure. The experiment proved that the method ofthis chapter was better than that of last chapter and it can improve the efficiency ofautomatic generation of test data.The test data generation for many paths coverage has not been studied detailed,RBF neural network was used to simulate the fitness vaule of all target paths and animproved method of grouping strategy was used to group the target paths in the fifthchapter of this paper. The test data would be generated by delete the target pathconstantly. In experiment, this method has been improved effective by comparing with method ofAhmed, the method of random and the method of Gong.The results of this study enriched application of the neural network technology inthe field of software testing, it reduced the cost of time in running instrumentationprogram. Therefore, it has theoretical significance and practical value.