A Framework for the Verification and Validation of Artificial Intelligence Machine Learning Systems
Date of Award
Kari, J, Lippert, D.Sc.
An effective verification and validation (V&V) process framework for the white-box and black-box testing of artificial intelligence (AI) machine learning (ML) systems is not readily available. This research uses grounded theory to develop a framework that leads to the most effective and informative white-box and black-box methods for the V&V of AI ML systems. Verification of the system ensures that the system adheres to the requirements and specifications developed and given by the major stakeholders, while validation confirms that the system properly performs with representative users in the intended environment and does not perform in an unexpected manner. Beginning with definitions, descriptions, and examples of ML processes and systems, the research results identify a clear and general process to effectively test these systems. The developed framework ensures the most productive and accurate testing results. Formerly, and occasionally still, the system definition and requirements exist in scattered documents that make it difficult to integrate, trace, and test through V&V. Modern system engineers along with system developers and stakeholders collaborate to produce a full system model using model-based systems engineering (MBSE). MBSE employs a Unified Modeling Language (UML) or System Modeling Language (SysML) representation of the system and its requirements that readily passes from each stakeholder for system information and additional input. The comprehensive and detailed MBSE model allows for direct traceability to the system requirements. xxiv To thoroughly test a ML system, one performs either white-box or black-box testing or both. Black-box testing is a testing method in which the internal model structure, design, and implementation of the system under test is unknown to the test engineer. Testers and analysts are simply looking at performance of the system given input and output. White-box testing is a testing method in which the internal model structure, design, and implementation of the system under test is known to the test engineer. When possible, test engineers and analysts perform both black-box and white-box testing. However, sometimes testers lack authorization to access the internal structure of the system. The researcher captures this decision in the ML framework. No two ML systems are exactly alike and therefore, the testing of each system must be custom to some degree. Even though there is customization, an effective process exists. This research includes some specialized methods, based on grounded theory, to use in the testing of the internal structure and performance. Through the study and organization of proven methods, this research develops an effective ML V&V framework. Systems engineers and analysts are able to simply apply the framework for various white-box and black-box V&V testing circumstances.
Burns, Swala B., "A Framework for the Verification and Validation of Artificial Intelligence Machine Learning Systems" (2023). Theses and Dissertations. 137.