Data-driven Identification Method Of Likely Metamorphic Relation For Numerical Calculation Programs

The development of numerical calculation software is to better solve mathematical problems faced in scientific research and engineering technology.Due to factors such as huge number of calculations or difficulty in collecting operating data,it is often difficult for testers to construct the correct expected value of the program to verify,that is the test Oracle problem which makes testing work become very difficult.As the numerical calculation program is a very important program,a little of subtle errors may also lead to huge consequences,so it is urgent to solve the problem of test Oracle.Metamorphic testing is one of the effective methods to solve this problem.It judges whether there are errors in the program under test by checking the relation(metamorphic relation)between multiple executions of the program.The current research on metamorphic testing focuses on the identification methods of metamorphic relation,including static analysis and data-driven methods.The data-driven method can discover metamorphic relation that are not easy to find in the static analysis method,reduce the difficulty of identifying process,and has obvious advantages and development prospects.Aiming at the problem of data-driven identification methods that rely on preset relation for data mining while ignoring domain knowledge,this topic proposes a new data-driven likely metamorphic relation identification method that combines domain knowledge to generate input relation.The main research content of the thesis includes the following three aspects:(1)Established the recognition rules of input relation based on domain knowledge.From the algorithm level background knowledge of the numerical calculation program: the nature of the numerical solution of the algorithm,the numerical characteristics of the program input data,as well as the data mutation and composition setting input relation are analyzed,and four input relation recognition rules are established.Through manual static analysis combined with the domain knowledge of the program to be tested,a batch of meaningful input relation can be derived.(2)Proposed a new data-driven identification method of likely metamorphic relation.Under the guidance of the input relation,a certain test case generation strategy is used to generate test data pairs.Before identifying the output relation,the relation pattern of six comparison types and ten function types are preset,and then the statistical analysis tool SPSS is used for data fitting.The test data pair is generated from the input relation,which prepares for the output relation mining to reduce data redundancy,and avoids blindness.(3)Realized the method application of the ordinary differential equations numerical method program and the trigonometric function program,and compare and analyze with the identification method of search polynomial.This paper derives 31 likely metamorphic relations of Runge Kutta programs and 22 likely metamorphic relations of trigonometric function sine programs.Comparing and analyzing the likely metamorphic relations discovered by the sine program with manual derivation and search methods,it is found that the metamorphic relation identified in this paper is effective and exceeds the above two methods in type and quantity.The above research shows that the data-driven likely metamorphic relation identification method proposed in this paper is effective.It combines the advantages of static analysis and data-driven identification methods,reduces the difficulty of finding metamorphic relation.In turn,it provides reference for the in-depth study of the metamorphic relation and metamorphic testing.