| Although commercial computer games usually undergo intensive testing before release, many bugs and sweet spots still exist and make games less attractive than expected. In this thesis, a Semi-Automated Gameplay Analysis (SAGA-ML) system is developed to summarize game behaviors as human readable rules, which can be presented to game designers to check if those behaviors are as intended. Unexpected game behaviors can be found this way. Machine learning and selective sampling techniques are incorporated into automated software testing. Machine learning is used to create a summary of the gameplay log that is comprehensible by humans. Selective sampling is used to sample instance space intelligently to build a good model. Four existing selective sampling algorithms (Uncertainty Sampling, Bagging, Boosting and BootStrapLV), and a new rule-based selective sampling method, are implemented and compared. SAGA-ML has been tested on Electronic Arts' FIFA99 soccer game and shown to be a practical game behavior testing solution. |
