SAGA-ML: An Active Learning System for Semiautomated Gameplay Analysis

Book One

Volume

Proceedings of the AAAI Conference on Artificial Intelligence, 20

Intelligent Systems Demonstrations

As software systems grow, software testing has become increasingly onerous. Consequently, statistical software testing and machine learning techniques have become increasingly attractive. Following these approaches, we present SAGA-ML (Semi-Automated Gameplay Analysis by Machine Learning), an active learning system for blackbox software testing, and SoccerViz, an analysis vizualization tool for Electronic Arts’ FIFA Soccer game. SAGA-ML was developed in the context of commercial video games, complex virtual worlds with state spaces too large for exhaustive testing. Beyond program correctness, developers must evaluate the gameplay of a game (e.g. its difficulty). SAGA-ML, which is not game-specific, samples the blackbox software system to learn a model of the system’s behaviour. This model is used to i) select new points for sampling, and ii) summarize the game’s behaviour for the developer. The demonstration shows how models learned by SAGA-ML (generated offline) can be visualized and explored by the developer via the game-specific SoccerViz tool.