Название: Safety Assurance under Uncertainties: From Software to Cyber-Physical/Machine Learning Systems
Автор: Ichiro Hasuo, Fuyuki Ishikawa
Издательство: CRC Press
Год: 2025
Страниц: 348
Язык: английский
Формат: epub (true)
Размер: 17.2 MB
Safety assurance of software systems has never been as imminent a problem as it is today. Practitioners and researchers who work on the problem face a challenge unique to modern software systems: uncertainties. For one, the cyber-physical nature of modern software systems as exemplified by automated driving systems mandates environmental uncertainties to be addressed and the resulting hazards to be mitigated. Besides, the abundance of statistical Machine Learning components massive numerical computing units for statistical reasoning such as deep neural networks make systems hard to explain, understand, analyze or verify.
The book is the first to provide a comprehensive overview of such united and interdisciplinary efforts. Driven by automated driving systems as a leading example, the book describes diverse techniques to specify, model, test, analyze, and verify modern software systems. Coming out of a collaboration between industry and basic academic research, the book covers both practical analysis techniques (readily applicable to existing systems) and more long-range design techniques (that call for new designs but bring a greater degree of assurance).
Testing is an “activity in which a system or component is executed under specified conditions, the results are observed or recorded, and an evaluation is made of some aspect of the system or component”. Testing has been one of the key activities for quality assurance in software-intensive systems. The recent advance of Machine Learning (ML) techniques, especially Deep Learning, has led to active investigation towards their industrial applications. Beforehand, ML techniques were target of laboratory research and the primary concern was on the prediction performance. As the prediction performance has evolved rapidly, increasing demand is put more on the dependability and quality of the ML applications.
Machine Learning (ML) refers to techniques to build functionality of prediction by experience or data. The current trend from late 2010s for applications of ML techniques primarily focuses on supervised learning. In supervised learning, training data is given that contains input-output pairs so that the relationship between them can be learnt via training. Practical application of supervised learning has been actively investigated as its role is often clear: supervised learning allows to build software components by deriving their behaviors from training data even if we cannot rigorously specify the behaviors. Other types of ML techniques include unsupervised learning and reinforcement learning. Unsupervised learning allows for extracting mathematical relations from given data. An example of unsupervised learning techniques is clustering, which divides the given data into multiple clusters by some similarity metric. Unsupervised learning is considered as data analysis activities rather than software system development. Reinforcement learning allows us to build agents that take proper actions in an evolving environment.
The book provides high-level intuitions and use-cases of each technique, rather than technical details, with plenty of pointers for interested readers.
Скачать Safety Assurance under Uncertainties: From Software to Cyber-Physical/Machine Learning Systems
[related-news]
[/related-news]