Metamorphosis Relationship Generation Based on Fixed Memory Step Gradient Descent Method with Noise

Metamorphosis relationship generation is a state-of-the-art practical applied testing method in the field of software testing industry, and its main idea is to detect whether there are errors caused by unsatisfied programs in two or more output results. The most important task is of course to generate the metamorphosis relationship, and such process is a kind of supervised learning with huge amount of calculation. As metamorphosis relationship generation is a kind of supervised learning whose specific loss function is based on absolute value and has poor derivability, it not only depends on huge amount of calculation, but also requires complex algorithms to get the optimal solution, as the traditional gradient descent (GD) method cannot work well in solving its loss function. In this paper, two improvements for metamorphosis relationship generation has been made: firstly, Huber loss function is introduced to make the main loss function derivable; secondly, fixed memory step gradient descent method with noise is introduced to solve the sawtooth effect and falling into local optimum problems. Via numerical analysis, it has been verified that the improved metamorphosis relationship generation has better performance.