Reward-Adaptive Iterative Discovery: A Case Study on Automated Game Testing for NHL26

Game testing is a resource-intensive process, often requiring significant human effort to identify behavioral exploits in complex AI systems. While Reinforcement Learning (RL) has been used to automate this, standard RL agents typically converge to a single "optimal" strategy. This forces a sequential, human-in-the-loop workflow where developers must fix one exploit before the agent can be retrained to find another. The authors aim to streamline this by developing an automated method that discovers a diverse set of high-performing exploits without requiring human intervention between iterations.

Reward-Adaptive Iterative Discovery (RAID) addresses the lack of diversity in standard RL by training a population of agents sequentially. After each agent converges, its strategy—defined by the shot type and the final puck position—is recorded. For all subsequent training iterations, the reward function is modified to mask any strategies similar to those already found. In the context of NHL 26, the authors define "similarity" as any shot of the same type occurring within a 2-meter radius of a previously discovered strategy. This simple, intuitive constraint allows developers to easily adjust the level of diversity without requiring deep expertise in RL.

The authors evaluated RAID on a development version of EA SPORTS NHL 26. While a naive baseline (standard RL) consistently converged to the same two scoring patterns regardless of the random seed, RAID successfully discovered a wide variety of distinct, high-quality scoring strategies. In a single experiment, the algorithm identified six unique exploits that were qualitatively similar to those discovered by human playtesters in hours of manual testing. The results demonstrate that RAID can effectively automate the discovery of diverse exploits, significantly reducing the manual burden on game development teams.