Browsing by Author "Kästner, Max"

Now showing 1 - 1 of 1

Open Access
Analyzing software architectures of open-source AI-based software
(2024) Kästner, Max
Recently, software systems that include artificial intelligence (AI) components (also called AI-based software systems) have become increasingly popular. However, the integration of AI into software systems brings new challenges compared to traditional software systems. In particular, the software architecture of AI-based software systems needs more research. In this thesis, we investigate the software architecture of open-source AI-based software applications, with an explicit focus on the design patterns and architectural tactics used in the architectures. For this purpose, we have mined open-source AI-based software systems from GitHub to analyze the software architecture of the software systems. To identify the design patterns, we use a design pattern catalog specifically created for AI-based software systems. We analyzed a total of 3 open-source AI-based software systems and were able to identify a total of 19 unique design patterns, with a balanced number of found traditional adapted design patterns and new AI-specific design patterns. We could see that design patterns from the categories Architecture and Deployment were more frequently identified in the systems than in the other categories Implementation, Process, Security & Safety and Testing & Quality Assurance. It was noticeable that many design patterns in the Architecture category were adapted from traditional design patterns, while several new AI-specific design patterns were found for Deployment, solving the problem of how to deploy the machine learning (ML) models. While a lot of new AI-specific design patterns for Security & Safety have been established, we could only identify very few. In addition, testing AI components is very unpopular in the analyzed systems and difficult as the correctness of an ML model is hard to test.