Table 1 - uploaded by Edouard Batot
Content may be subject to copyright.
Statistical comparison of results between random search and our approach on three WFR learning scenarios.
![Fig. 4: Non Sorting Genetic Algorithm NSGA-II [18]](profile/Edouard-Batot/publication/327146164/figure/fig4/AS:672258225868802@1537290259541/Non-Sorting-Genetic-Algorithm-NSGA-II-18_Q320.jpg)
Context in source publication
Context 1
... -Sanity Check As can be seen in Table 1, the RND configuration gives very poor results in comparison with an OBJ execution for the two most complex metamod- els (average accuracy on test bench is 0.5 vs 0.76 for Project Manager and 0.53 vs. 0.94 for Statemachine). The difference in both cases is statistically significant (p-value<0,001) and the effect size is large (Cohen's d > 5). ...
Citations
... In addition, EAs are known to give more power to good solutions, which can cause converging issues due to loss of diversity, a problem known as single fitness peak. Using behavior specifications such as test cases to guide the search in EAs can exacerbate these limitations [5,41]. ...
... In this paper, we extend our EA-based approach from [47] to automatically find patches to correct transformation with a greater number of semantic errors. In particular, to improve the efficiency and effectiveness of EAs using test cases, our improved approach leverages the notion of social diversity [5]. This metric promotes patches which tackle errors that are less covered by the other patches of the population. ...
... On the other hand, understanding the impact of syntactic diversity on the behavior of a program is quite complex [27]. We thus focus on semantic diversity, which is also known to be more efficient to prevent single fitness peak [5,46]. ...
Model transformations play an essential role in the model-driven engineering paradigm. However, writing a correct transformation requires the user to understand both what the transformation should do and how to enact that change in the transformation. This easily leads to syntactic and semantic errors in transformations which are time-consuming to locate and fix. In this article, we extend our evolutionary algorithm (EA) approach to automatically repair transformations containing multiple semantic errors. To prevent the fitness plateaus and the single fitness peak limitations from our previous work, we include the notion of social diversity as an objective for our EA to promote repair patches tackling errors that are less covered by the other patches of the population. We evaluate our approach on four ATL transformations, which have been mutated to contain up to five semantic errors simultaneously. Our evaluation shows that integrating social diversity when searching for repair patches improves the quality of those patches and speeds up the convergence even when up to five semantic errors are involved.
... In addition, EAs are known to give more power to good solutions, which can cause converging issues due to loss of diversity, a problem known as single fitness peak. Using behavior specifications such as test cases to guide the search in EAs can exacerbate these limitations [4,11]. ...
... On the other hand, understanding the impact of syntactic diversity on the programs' behaviors is quite complex [23]. We thus focus on semantic diversity, which is also known to be more efficient to prevent single fitness peak [4,36]. ...
... Batot et. al [4] proposed a social diversity measure giving higher scores to solutions which pass test cases frequently failed by the other solutions. A solution passing numerous test cases that the majority of the population also pass will receive a lower score than a solution passing less test cases but which are failed by a majority of the population. ...
Model transformations play an essential role in the Model-Driven Engineering paradigm. Writing a correct transformation program requires to be proficient with the source and target modeling languages, to have a clear understanding of the mapping between the elements of the two, as well as to master the transformation language to properly describe the transformation. Transformation programs are thus complex and error-prone, and finding and fixing errors in such programs typically involve a tedious and time-consuming effort by developers. In this paper, we propose a novel search-based approach to automatically repair transformation programs containing many semantic errors. To prevent the fitness plateaus and the single fitness peak limitations, we leverage the notion of social diversity to promote repair patches tackling errors that are less covered by the other patches of the population. We evaluate our approach on 71 semantically incorrect transformation programs written in ATL, and containing up to five semantic errors simultaneously. The evaluation shows that integrating social diversity when searching for repair patches allows to improve the quality of those patches and to speed up the convergence even when up to five semantic errors are involved.
Traceability is the capability to represent, understand and analyze the
relationships between software artefacts. Traceability is at the core of many software
engineering activities. This is a blessing in disguise as traceability research
is scattered among various research subfields, which impairs a global view and
integration of the different innovations around the recording, identification, evaluation
and management of traces. This also limits the adoption of traceability
solutions in industry.
In this sense, the goal of this paper is to present a characterization of the traceability
mechanism as a feature model depicting the shared and variable elements
in any traceability proposal. The features in the model are derived from a survey
of papers related to traceability published in the literature. We believe this
feature model is useful to assess and compare different proposals and provide a
common terminology and background. Beyond the feature model, the survey we
conducted also help us to identify a number of challenges to be solved in order
to move traceability forward, especially in a context where, due to the increasing
importance of AI techniques in Software Engineering, traces are more important
than ever in order to be able to reproduce and explain AI decisions.
