An Evolutionary Approach to Translate Operational Specifications into Declarative Specifications

EvoSpex: An Evolutionary Algorithm for Learning Postconditions

“…Our representation of candidate assertions is based on the encoding used in [24], where chromosomes represent conjunctions of assertions (each gene in a chromosome represents an assertion). That is, given a chromosome c, the candidate postcondition ipc represented by c is defined as follows: c = (g1, g2, .…”

“…Let us describe how we build the initial population, to start our genetic algorithm. In order to create individuals representing "meaningful" postconditions, i.e., assertions stating properties of objects that are reachable at the end of the method executions, we take into account typing information, as in [24]. We consider a type graph built automatically from the class under analysis: nodes represent types, and each field f of type B in class A will produce an arc in the graph going from the node representing A to the node representing B.…”

An evolutionary approach to translating operational specifications into declarative specifications

“…It is worth mentioning that, despite the fact that we use a semantics-preserving translation to take, for a very small scope, the operational specification Φ op into the declarative context, this is not actually a requirement of our approach. We may, as we did for a preliminary version of our technique [36], use the operational specification Φ op to produce a set of positive and negative examples, i.e., instances that satisfy and do not satisfy Φ op , respectively, and use these to evaluate the fitness of a candidate c; this approach simply evaluates how many of the positive and negative instances satisfy the specification Φ c , and defines the fitness of c as described above. To generate the instances, any test input generation mechanism that requires an operational specification, e.g.…”

“…Other related works are, of course, the preliminary version of our technique presented in [36], the works on specification inference as put forward in techniques like Daikon [12] and Deryaft [34] (that we compared with in this paper), and our recent work in [37]. As opposed to the work presented in this paper, the approach in [37], while it also uses a genetic algorithm, is less powerful since it only learns invariants composed of properties from a given catalog, in the style of [34] but considering both positive and negative instances.…”

“…This paper is a journal extension of [36]. The additional contributions, with respect to the original conference submission, are the following: (i) an improved genetic algorithm, in particular in relation to how the fitness of candidate specifications is measured, (ii) additional data structures considered for the evaluation of efficiency, effectiveness and precision of our technique to translate operational specifications into declarative ones, (iii) an experimental evaluation of the impact of various parameters of the genetic algorithm, including our approach to producing the initial population, the population size, and the mutation rate, and (iv) an evaluation of the impact of our learned declarative specifications in aiding symbolic execution through tight bounds (whose computation requires declarative logical specifications).…”