Verification & Validation Methods for Complex AI-enabled Cyber-Physical Learning-Based Systems: A Systematic Literature Review

This systematic literature review (SLR) investigates how systems- and software engineers should approach verification and validation (V&V) of complex Artificial Intelligence (AI)-enabled cyber-physical learning-based systems. General trust in learning-based, complex, self-adaptive cyber-physical systems (CPSs) is low. For the technology to be accepted and trusted for more industrial and commercial use, proper V&V techniques, methods, frameworks, and theories are required to ensure desired outcomes. Principal findings show that traditional V&V methods could be classified as either inspection, demonstration, testing, or analysis. These conventional methods are highly nuanced for complex systems. They may even work for large CPSs where determinism and linear internal properties are assumed. However, the self-adaptive learning nature of AI-systems calls for adapted or new V&V methods. Primary complications encountered in V&V activities for AI-enabled systems relate to input data quality, model and requirements management, non-determinism, and AI black-box effect related to poor explainability, scale, and complexity issues. This review highlights the importance of a designated data officer with authority and capability to manage the relevant data for being relevant, available, bias-free, and adversarial sample free. Also, an issue list or complications may also provide caution for implementing AI-systems.