Abbas Rahimi

Abbas Rahimi


Start: Sep. 2019
Finish: Sep. 2021
Thesis Title: A Framework Based on Deep Learning for Model Generation in Model-Driven Engineering
Supervisor: Dr. Shekoufeh Kolahdouz-Rahimi     Advisor: Dr. Massimo Tisi
Current Position: Ph.D. Student at Johannes Kepler University of Linz (JKU)

Abbas Rahimi received his B.Sc. in Computer Engineering (Software) from Yazd University, Yazd, Iran, in 2015.


Thesis Abstract:
In model-driven engineering (MDE), models are the main artifacts. Models are used in various activities, for instance, model transformation testing. In the real world, we faced a large range of domains and some limitations (such as the necessity of realistic models for evaluating critical systems) that make it difficult to access appropriate model repositories; Therefore, model generation has been proposed as one of the most important challenges in MDE. Researchers have developed various tools and methods to generate new models using different approaches (such as graph grammar, partitioning, and random). However, they rarely presented an approach that is able to produce new models with considering realistic features. In this dissertation, the concepts of machine learning and in particular deep learning have been used to present an intelligent framework for generating new models. In the represented framework, the advantages of generative adversarial networks (GANs) are used to generate the model. This framework is able to produce new models in four main steps by getting a metamodel and only one big instance model of it as inputs. Inspired by the structural similarity of the model and the graph, in the first step, an encoder maps the input model to a graph using metamodel’s structural information. The neural network parameters are then adjusted, and the graph structure is learned by the network. Next, the trained network is asked to generate new graphs using its learned information. Finally, in the last step, the new graphs are transformed to the model domain by a graph-to-model decoder. In other words, in the final stage, graphs are mapped to models. The represented framework is able to generate new artificial models. The most important feature of the generated models is that they are adequately realistic, so they can be used for various activities in the context of MDE. In this research, we developed a model-to-graph encoder and a graph-to-model decoder to change the problem space and use smart graph-based facilities. Graph-based criteria (such as Multiplex participation coefficient) have been used to evaluate the presented framework. The statistical results of the evaluation, illustrate that make use of generative deep learning algorithms such as GANs can improve the realistic nature of the artificially generated models.‎


Papers in English