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

Jul 16, 2026

Testing consumes significant budget and manpower in the gaming industry, especially when AI behaviors must be re-validated after iterative changes. This work studies a development version of EA SPORTS NHL 26, focusing on testing a goalie AI via human playtesters exploiting behavioral weaknesses. To reduce costly re-testing, Reward-Adaptive Iterative Discovery (RAID) uses iterative reinforcement learning to train a population of agents and extend beyond overfitting by producing multiple diverse, high-quality exploit strategies without human intervention.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages7

Reading Time18 min

Introduction

Related Work

Automated Playtesting

Behavior Diversity

Reward-Adaptive Iterative Discovery: A Case Study on Automated Game Testing for NHL26 Florian Fuchs* , Jessy Gosselin-Grant, Boris Skuin, Michele Petteni, Alessandro Sestini, Joakim Bergdahl, Amir Baghi, Linus Gissln* Electronic Arts (EA) {ffuchs,[lgisslen](lgisslen}@ea.com)[}](lgisslen}@ea.com)[@ea.com](lgisslen}@ea.com) * Corresponding authors arXiv :2607 .07498v 1 [ cs .LG] 8 Jul 2026 Abstract—Testing is a major effort for the gaming industry, requiring a significant part of development budget and people power. We present a case study on a development version of the ice hockey game EA SPORTS NHL 26 , for which human playtesters test the goalie AI for behavioral exploits. To reduce the effort of re-testing the goalie AI after every game or behavior modification in the development phase, we propose Reward-Adaptive Iterative Discovery (RAID), a novel approach to automatically find exploits using an iterative Reinforcement Learning (RL) approach that trains a population of goal scoring agents. While previous approaches can already successfully find exploits, RL algorithms tend to overfit to a single solution. We introduce a simple extension on top of existing RL algorithms, such that they find multiple diverse high-quality solutions. For our first deployment of this approach, within a single experiment we were able to find six hockey scoring exploit strategies that were qualitatively similar to those that playtesters had found in hours-long manual testing sessions. Index Terms—Automated playtesting, Reinforcement Learning, Diversity SUPPLEMENTARY VIDEO This paper is accompanied by videos of the agents trained with our algorithm, RAID: [go.ea.com/RAID](go.ea.com/RAID) I. INTRODUCTION With game worlds becoming vaster and game systems becoming more complex, the effort to test games grows likewise. To reduce repetitive aspects of game testing, prior work has investigated automating parts of it through the use of autonomous agents playing a game and reporting bugs [1, 2, 3] . While Reinforcement Learning (RL) agents can find individual exploits within a game [4], they tend to converge to one single “best” solution of solving a task. In this paper, we use EA SPORTS NHL 26 (short NHL) as our test-case. In particular, we aim to test the goalie AI of the game, training a forward agent to find high chance scoring strategies, which could represent potential exploits. Our results indicate that standard RL algorithms tend to collapse to a small set of high-reward behaviors, rather than exploring a larger, more diverse set. Thus, exploit discovery becomes sequential: developers must fix one issue before retraining the agent to find others. Since this slows down the exploit finding process, we aim to develop an algorithm that can find multiple potential exploits without human intervention. 1 2 3 Fig. 1. Top: Each image shows the last frame before the goal of a different scoring strategy learned by an agent trained with RAID. Our agent’s forward player wears a blue jersey, the goalie a white jersey. The strategies are learned in an iterative fashion, with a reward function enforcing each new strategy’s shot position to be at least 2 meters away from all previous strategies using the same shot type. The numbers indicate which iteration learned this specific strategy and the colors indicate the shot types used to score the goal. Bottom: The shot locations and types of the strategies found by the first 4 iterations of RAID. The small dots represent the shot location and type of 100 goals taken after convergence of each iteration. The large circles represent a 2 meter radius around the average position of those goals inside which no reward is given for all succeeding iterations of RAID if scoring with the same shot type. The shots by the agent of iteration 3 are for example outside of the 2 meter radius of iteration 1, since the two agents use the same shot type. Prior work on RL for game testing has proposed methods

FAQs

What problem does RAID address in automated game testing?

It reduces the repeated effort required to re-test the goalie AI after each game or behavior modification during development.

How does RAID improve over standard reinforcement learning approaches?

It avoids RL overfitting to a single high-reward behavior by finding multiple diverse, high-quality exploit strategies.

What role does diversity enforcement play in RAID’s results?

RAID uses a reward function that discourages scoring strategies similar to previously discovered ones, maintaining diversity across sequential training iterations.

Jul 16, 2026

Upload by Seraphina