Seite - 134 - in The Future of Software Quality Assurance

134 G.Numan In the movieRobocopofficer Murphyhad a “primedirective” programmedinto his system: if he would try to arrest a managing director of his home company, his system would shut down. AI systems could have prime directives too, or unacceptableresults, likeoffensivelanguage,pornsitesordrivingintoapedestrian. We call these “contracts”: possible unwanted results that should be flagged in the test resultsasblocking issuesorat least begivena highweight. Required contracts have to be part of the test set. Possible negative side effects ofexistingcontracts shouldbepartof the test set too. 4.8 Test Automation AItestingneedssubstantialautomation.Theamountoftestcasesrequest itandtests need to be run repetitively with every new version. When the AI system is trained constantly, testing is necessary, as in the case of search engines where there are feedbackloopsfromrealdata.ButevenwhentheAIsystemisnottrainedconstantly andversionsof thesystemarestable,achangingcontextdemandsconstant training. Evenwhen thesystemdoesnotchange, theworldwill. Test automationconsists of a test frameworkwhere the test cases will be run on the AI system and the output from the AI system will be processed. Below a basic setupof sucha test frameworkis shown. 4.9 Overall EvaluationandInput for Optimising The product of testing is not just a list of bugs to be fixed. Bugs cannot be fixed directly without severe regression, as stated above. The AI-system has to be evaluatedas a whole since with the many test cases and regression,no versionwill be perfect. Programmers want to know which version to take, if a new version is better than a previousone. Therefore the test results should be amalgamated into a total result:aquantitatedscore.Forprogrammerstogetguidanceintowhat to tweak (training data, labelling, parametrisation) they need areas that need improvement. This is as close that we can get to bug fixing. We need metrics, weighing and contracts to achieve a meaningful overall score and clues for optimisation. Low scoring test cases should be analysed as to their causes: is it over-fitting, under- fittingoranyof theother riskareas? 4.10 ExampleofAITest Framework (Fig. 2) Fromleft up tobottomand thenrightup: 1. Identifyingusergroups 2. Creatingpersonaperusergroup