Automated Model Design Using Genetic Algorithms and Model Checking

In recent years there has been a growing interest in applying metaheuristic search algorithms in model-checking. On the other hand, model checking has been used far less in other software engineering activities, such as model design and software testing. In this paper we propose an automated model design strategy, by integrating genetic algorithms (used for model generation) with model checking (used to evaluate the fitness, which takes into account the satisfied/unsatisfied specifications). Genetic programming is the process of evolving computer programs, by using a fitness value determined by the program's ability to perform a given computational task. This evaluation is based on the output produced by the program for a set of training input samples. The consequence is that the evolved program can function well for the sample set used for training, but there is no guarantee that the program will behave properly for every possible input. Instead of training samples, in this paper we use a model checker, which verifies if the generated model satisfies the specifications. This approach is empirically evaluated for the generation of finite state-based models. Furthermore, the previous fitness function proposed in the literature, that takes into account only the number of unsatisfied specifications, presents plateaux and so does not offer a good guidance for the search. This paper proposes and evaluates the performance of a number of new fitness functions, which, by taking also into account the counterexamples provided by the model checker, improve the success rate of the genetic algorithm.