VERIFICATION AND VALIDATION OF AI SOFTWARE

NEW VERIFICATION AND VALIDATION (V&V) TECHNIQUES ARE NEEDED FOR USE WITH AI SOFTWARE. FORTUNATELY, AI SYSTEMS HAVE SEVERAL FEATURES THAT CAN BE USED IN NEW WAYS TO SUPPORT VERIFICATION AND VALIDATION; IN PARTICULAR, AUTOMATIC REASONING ABOUT DECLARATIVE REPRESENTATION CAN CHECK FOR CONSISTENCY AND OTHER PROPERTIES OF RULES, REASONING ABOUT UNCERTAINTY CAN BE USED TO ACHIEVE ROBUTNESS IN THE FACE OF ERRORS, REDUNDANCY AND SELF-CHECKING CAN ACHIEVE RUN-TIME VERIFICATION, SIMULATIONS CAN BE USED AS A TESTING METHODOLOGY, AND IMPROVED METHODS FOR MODULARITY, ABSTRACTION, AND USER INTERACTION CAN BE USED TO GIVE USERS A BETTER WORKING MODEL OF THE SYSTEM'S FUNCTIONALITY AND THUS ENABLE THE USERS TO GAIN CONFIDENCE IN THE SYSTEM AS THEY USE IT. THIS RESEARCH WILL EXPLORE THESE AND OTHER TECHNIQUES FOR VERIFYING AND VALIDATING AI SOFTWARE IN THE CONTEXT OF AN OVERALL SOFTWARE DEVELOPMENT MODEL THAT IS COMPATIBLE WITH AI SOFTWARE DEVELOPMENT METHODS. MULTIPLE V&V TECHNIQUES ARE NEEDED, AND THE CHOICE OF THE SET OF V&V TECHNIQUES USED FOR ANY SOFTWARE APPLICATION DEPENDS ON THE APPLICATION, ON THE AI FEATURES USED IN THE SOFTWARE, AND ON THE SOFTWARE DEVELOPMENT ENVIRONMENT. A GOAL OF THIS RESEARCH IS TO DEVELOP GUIDELINES FOR CHOOSING THE V&V TECHNIQUES THAT ARE EFFECTIVE FOR A GIVEN APPLICATION.