A Progress-Based Algorithm for Interpretable Reinforcement Learning in Regression Testing

In video games, the validation of design specifications throughout the development process poses a major challenge as the project grows in complexity and scale and purely manual testing becomes very costly. This paper proposes a new approach to design validation regression testing based on a reinforcement learning technique guided by tasks expressed in a formal logic specification language (TLTL) and the progress made in completing these tasks. This requires no prior knowledge of machine learning to train testing bots, is naturally interpretable and debuggable, and produces dense reward functions without the need for reward shaping. We investigate the validity of our strategy by comparing it to an imitation baseline in experiments organized around three use cases of typical scenarios in commercial video games on a 3D stealth testing environment created in Unity. For each scenario, we analyze the agents' reactivity to modifications in common assets to accommodate design needs in other sections of the game, and their ability to report unexpected gameplay variations. Our experiments demonstrate the practicality of our approach for training bots to conduct automated regression testing in complex video game settings.