Variety in Autonomous Playtesting: A Case Study on Reinforcement Learning for 'NHL26'

Reinforcement learning (RL) is already successfully used for playtesting games without human intervention, allowing game studios to reduce the significant time and cost spent for this important task in game development. We present a case study on testing the goalie AI in the hockey game NHL26, which is using a traditional rule-based algorithm, by training a forward player with RL to find exploits in the goalie's behavior. We show that out of the box RL algorithms only provide limited value for this kind of testing scenario, because they converge to a single exploit strategy. With some simple steps we can enhance the algorithm to instead provide a set of qualitative and diverse solutions that provide additional value to game designers, since they allow finding multiple fix-worthy exploits in a single experiment iteration. For our first deployment of our approach, within a single experiment we were able to find six exploit strategies that were qualitatively similar to those that game testers found in hour long manual testing sessions.