Online swarm exploration in virtual 3D environments : Possible approaches for expanding simulation-based game testing

Abstract: The use of AI-driven and simulation-based testing of games is becoming a more and more practiced method in the game development industry as present- day virtual 3D environments become more and more semantically complex. Test coverage (here defined as a metric of how many game states are reached in a set of simulation-based tests) in games can be generated by creating interaction between the virtual characters and their virtual environments. This work aims to explore simulation-based testing of virtual 3D environments by using swarm intelligence: a field within artificial intelligence where algorithms are inspired by swarm behaviours observed in the animal kingdom. Four different agent swarm types (wander, repulsive wander, reactive Lévy walk and repel- lent pheromone ants) are evaluated with the strive of finding robustness and scalability over different types of environments with the goal of spatial exploration. The experiments are carried out in four different virtual environments constructed in the game engine Unity3D. Three environments have simple de- signs (sparse obstacles, maze and multi-floored) and one environment have a more complex design (a virtual forest). The methods are evaluated on: their strengths and weaknesses in terms of spatial coverage over time in the different environment types and for different agent quantities (10, 50, 100), their dispersion/edge-covering capabilities and how well they presumably can be applied in the targeted application (game testing systems). The study finds that the most scalable and robust method of the studied methods seems to be the wander behaviour, while the most complex method: repellent pheromone ants, proves very efficient in some environments but less in others and can be more cumbersome to implement and computationally heavy to run. Over- all, the work can conclude that swarm intelligence could be of interest from a game testing perspective as it could help to increase test coverage of virtual 3D environments in general simulation-based testing.