Presentation is loading. Please wait.

Presentation is loading. Please wait.

Automated Test Data Generation Maili Markvardt. Outline Introduction Test data generation problem Black-box approach White-box approach.

Similar presentations


Presentation on theme: "Automated Test Data Generation Maili Markvardt. Outline Introduction Test data generation problem Black-box approach White-box approach."— Presentation transcript:

1 Automated Test Data Generation Maili Markvardt

2 Outline Introduction Test data generation problem Black-box approach White-box approach

3 Introduction Improtance of testing growing since mission- criticality of the software in our everyday life Software errors are more costly than ever Testing can be automated Test execution automation Test generation automation Test data generation automation

4 Problem: example User inputs three sides of a triangle (a, b, c). Which type is the triangle? Requirements: IF a input incorrect IF p*(p-a)*(p-b)*(p-c) sides not forming a triangle IF a==b || a==c || b ==c -> isoceles Kui a==b & b==c -> eqilateral Other -> scalene What strategy? -> what data?

5 Input validation automation The Concept side of a triangle equivalence partitions and boundary values Normal: ]0; [ Erroneous: ]- ; 0[, missing values Border values: {0} For testing the Input validation functionality, pick a random value from each equivalence partition for each side: P(-1, 2, 3), P(1, -2, 3), P(1, 2, -3) Same with boundary values P(0, 1, 2), P(1, 0, 2), P(1, 2, 0) Input validation with normal values P(1, 2, 3)

6 What about other requirements? If input values are dependent and that affects output, random values can not be used! – we may not be able to find needed values with random generation if p*(p-a)*(p-b)*(p-c) sides dont form a triangle We must use specification-based (Black-box) or program-based (White-box) test data generation

7 Black-Box approach Generating test data from formal specifications (ie. Z-notation) Classification Tree Method (CTM)...

8 Classification Tree Method 4 Based on equivalence partitions method: input and output properties are divided into equivalence partitions Equivalence partitions are combined into test cases The goal: minimal but sufficient amount of test cases 4 Dai, Z. R., Deussen, P. H. Automatic Test Data Generation for TTCN-3 Using Classification Tree Method (2005)

9 CTM Equivalence partitions form a tree structure Input dependencies are not resolved

10 White-Box approach White-box test data generation – based on program structure Test data generation problem: For program P and path u, find input x S, so that P(x) traverses path u, where S is the set of all input values Remember: white-box approach is based on program formalisation (graph, FSM, …)

11 Test data generator structure 2 2 Edvardsson, J. Contributions to Program- and Specification-based test data generation. (2002). www.ida.liu.se/~joned/papers/joned_lic.pdfwww.ida.liu.se/~joned/papers/joned_lic.pdf

12 Possible strategies (adequacy criteria) Statement coverage Branch coverage Condition coverage Multiple-condition coverage Path coverage …

13 Numerous methods for Constraint generator & Constraint solver Symbolic Execution Actual Execution Symbolic/Actual Execution hybrid Simulated Annealing Iterative Relaxation Technique Chaining Approach Genetic Algorithms MEA-Graph Planning...

14 Symbolic Execution 2 Popular static method for finding path constraints Path constraints are rewritten using input variables Not suitable for programs using pointers and arrays Not suitable for programs using precompiled units read(a,b) c=a+b; d=a-b; e=c*d; if (e>5) {...} a*a – b*b > 5 =>

15 Actual Execution 2 Program is executed several times On every execution: check, whether or not the desired path is executed If desired path is not executed, program is re- executed with slightly modified input values Program is re-executed until desired path is traversed or user-defined limit (time, execution count) is exceeded Solves some problems of symbolic execution since values of variables are available

16 Actual Execution For each path condition b i objective function is found: F i (x) {<|<=|=} 0 If F i (x) {<|<=|=} 0, then current path is executed F(x)= Σ F i (x), if branch consists of several conditions How to minimize objective function so that F i (x)=0 In other words, what input values are needed to execute desired path?

17 Simulated Annealing Simulated Annealing – generic probabilistic meta-algorithm for finding good approximation to the global optimum for a given function in a large search space Analogy from metallurgy: Process of annealing is used for reducing defects in material Metal is heated: atoms start to move Metal is cooled down slowly: greater probability that atoms find a suitable place

18 Simulated Annealing Goal: minimize the objective function -> desired path is executed Find a random solution for objective function Compare the solution with current solution of objective function Decide, whether or not the random solution is better than current solution

19 Simulated Annealing If random solution gives a better value (closer to 0) for objective function, the random value is always chosen (probability is 1) If random solution is not better than current solution – sometimes it is chosen – depends of the temperature The value of temperature is decreased In the beginning high temperature -> almost every solution is chosen When temperature is lowered, the probability of choosing worse solution is lowered until it is 0

20 Simulated Annealing: properties Choosing worse solutions in the beginning lowers the probability of getting stuck in a local optimum (drawback of Gradient Descent/Hill Climbing/Greedy algorithms) It is possible to show, that probability of finding global optimum is almost 1 Little use in practice, since finding the global optimum with sufficient significance by annealing takes more time than full search of the whole search space

21 Simulated Annealing Parameters for successful simulated annealing: Art rather than science How to find a random solution – how to minimize the count of iterations finding the optimum? How to determine, whether or not the worse solution is picked? Annealing schedule – from what temperature to start and how the temperature is lowered?

22 Genetic algorithms Imitates the process of natural selection Evaluation Choice Recombination and mutation Start with random set of solutions - population Solutions are evaluated for their fitness – ability to generate good offspring Chosen (good) solutions are recombined and mutated to generate a new generation of solutions

23 GA for test data generation 5 Algorithm is driven by control dependency graph of the program Graph nodes = program statements Graph edges = control dependencies between program statements Goal: find data for executing certain node (program sentence) Node X is post-dominated by node Y, if every directed path from X to the end of the program includes node Y 5 Pargas, R., Harrold, J.M., Peck, R.R. Test-Data generation Using genetic Algorithms (1999)

24 GA for test data generation Node Y is control dependent of Y, only if Exists a directed path from Y to X and all nodes on this path (except X,Y) are post- dominated by Y and X is not post-dominated by Y Control dependency predicate path (CDPP)– predicates that must be satisfied on acyclic path from initial node to some other node X

25 GA for TDG: algorithm Solution is set of test data Start with random set Evaluate fitness of data Execute the program with data, mark predicates on executed path Compare the found set of predicates with CDPP to desired node The more the found set of data allowed to execute CDPP, the better the data is

26 GA for TDG Best solutions are chosen, recombined and mutated to generate a new generation of solutions Non-typical application of GA – several possible solutions, depending on the test goal ie. find data for executing nodes A, C – more than one test may be needed if A and C are exclusive!

27 GA: example int i, j, k; 1: read i, j, k; 2: if (i<j) { 3: if (j<k) { 4: i=k; } else { 5: k=i; } 6: print i, j, k; 5 is test goal, CDGpath {ET, 2T, 3F}

28 GA: example Random population Fitness f{2, 2, 0, 0} Probability of choosing p i = f i /Σf j ({0.5,0.5,0,0}) One solution can be chosen more than once TestCaseInput(i,j,k)Path TC11, 6, 91,2,3,4,6 TC20, 1, 41,2,3,4,6 TC35, 0, 11, 2, 6 TC42, 2, 31, 2, 6

29 GA: Näide New population {(1, 6, 9), (0, 1, 4), (0, 1, 4), (0, 1, 4)} Recombination (one-point crossover): n first values form one parent and others from the other parent N=2: {(1, 6, 4), (0, 1, 9)} Mutation: Value in random position is replaced with a random number (0, 6, 4), (5, 1, 4)

30 Summary Numerous methods Black-Box White-Box Choice of methods depends on knowledge and preference of tester, Technology of SUT

31 Viited 1 Edvardsson, J. A survey on Automatic Test Data Generation. (1999). [WWW] www.ida.liu.se/~joned/papers/class_atdg.pdf www.ida.liu.se/~joned/papers/class_atdg.pdf 2 Edvardsson, J. Contributions to Program- and Specification-based test data generation. (2002). [WWW] www.ida.liu.se/~joned/papers/joned_lic.pdf www.ida.liu.se/~joned/papers/joned_lic.pdf 3 Gupta, N., M, Mathur, A., Soffa, M.L. Automated Test Data Generation Using an Iterative Relaxation Method (1999)

32 Viited 4 Dai, Z. R., Deussen, P. H. Automatic Test Data Generation for TTCN-3 Using Classification Tree Method (2005) 5 Pargas, R., Harrold, J.M., Peck, R.R. Test- Data generation Using genetic Algorithms (1999)


Download ppt "Automated Test Data Generation Maili Markvardt. Outline Introduction Test data generation problem Black-box approach White-box approach."

Similar presentations


Ads by Google