Model-driven development methodology and domain-specific languages for the design of artificial intelligence in cyber-physical systems

• Modellgetriebene Entwicklungsmethodik und Domänenspezifische Sprachen für den Entwurf Künstlicher Intelligenz in Cyber-Physischen Systemen

Kusmenko, Evgeny; Rumpe, Bernhard (Thesis advisor); Aßmann, Uwe (Thesis advisor)

Düren : Shaker Verlag (2021, 2022)
Book, Dissertation / PhD Thesis

In: Aachener Informatik Berichte Software Engineering 49
Page(s)/Article-Nr.: xiv, 324 Seiten : Illustrationen

Dissertation, RWTH Aachen University, 2021

Abstract

The development of cyber-physical systems poses a multitude of challenges requiring experts from different fields. Such systems cannot be developed successfully without the support of appropriate processes, languages, and tools. Model-driven software engineering is an important approach which helps development teams to cope with the increasing complexity of today's cyber-physical systems. The aim of this thesis is to develop a model-driven engineering methodology with a particular focus on interconnected intelligent cyber-physical systems such as cooperative vehicles. The basis of the proposed methodology is a component-and-connector architecture description language focusing on the decomposition and integration of cyber-physical system software. It features a strong, math-oriented type system abstracting away from the technical realization and incorporating physical units. To facilitate the development of highly-interconnected self-adaptive systems, the language enables its users to model component and connector arrays and supports architectural runtime-reconfiguration. Architectural elements can be altered, added, and removed dynamically upon the occurrence of trigger events. In order to fully cover the development process, the proposed methodology, in addition to structural modeling, provides means for behavior specification and its seamless integration into the components of the architecture. A matrix-oriented scripting language enables the developer to specify algorithms using a syntax close to the mathematical domain. What is more, a dedicated deep learning modeling language is provided for the development and training of neural networks as directed acyclic graphs of neuron layers. The framework supports different learning methods including supervised, reinforcement, and generative adversarial learning, covering a broad range of applications from image and natural language processing to decision making and test data generation. The presented toolchain enables an automated generation of fully functional C++ code together with the corresponding build and training scripts based on the architectural models and behavior specifications. Finally, to facilitate the integration and deployment of the modeled software in distributed environments, we use a tagging approach to model the middleware and to control a middleware generation toolchain.