Automated Game Testing with ICARUS: Intelligent Completion of Adventure Riddles via Unsupervised Solving
Abstract
With ICARUS, we introduce a framework for autonomous video game playing, testing, and bug reporting. We report on the design rationale, the practical implementation, and its use in game development industry projects. The underlying solving mechanic is based on discrete reinforcement learning in a dualistic fashion, encompassing volatile short-term memory as well as persistent long-term memory that spans across distinct game iterations. In combination with heuristics that reduce the search space and the possibility to employ pre-defined situation-dependent action choices, the system manages to traverse complete playthrough iterations in roughly the same amount of time that a professional game tester requires for a speedrun. The ICARUS project was developed at Daedalic Entertainment. The software can be used to generically run all adventure games built with the popular Visionaire Engine and is currently used for evaluating daily builds, for large-scale hardware compatibility and performance tests, as well as for semi-supervised quality assurance playthroughs. The supplementary video depicts real-time solving with active control and observation via a web control panel.