Abstract
Generating tests for games is challenging due to the high degree ofrandomisation inherent to games and hard-to-reach program states that requiresophisticated gameplay. The test generator NEATEST tackles these challenges bycombining search-based software testing principles with neuroevolution tooptimise neural networks that serve as test cases. However, since NEATEST isdesigned as a single-objective algorithm, it may require a long time to coverfairly simple program states or may even get stuck trying to reach unreachableprogram states. In order to resolve these shortcomings of NEATEST, this workaims to transform the algorithm into a many-objective search algorithm thattargets several program states simultaneously. To this end, we combine theneuroevolution algorithm NEATEST with the two established search-based softwaretesting algorithms, MIO and MOSA. Moreover, we adapt the existingmany-objective neuroevolution algorithm NEWS/D to serve as a test generator.Our experiments on a dataset of 20 SCRATCH programs show that extending NEATESTto target several objectives simultaneously increases the average branchcoverage from 75.88% to 81.33% while reducing the required search time by93.28%.