Glitch Hunter: Autonomous Game Testing Using Reinforcement Learning
IOSR Journal of Computer Engineering · 2026-06 · 28 (6) · 42-49
DOI: 10.9790/0661-2806014249 ↗Kurz gefasst
Fachartikel in der Zeitschrift IOSR Journal of Computer Engineering, 2026-06. Verfasst von Aryan Kumar und 3 weiteren Autor:innen.
Abstract (Originalsprache)
Recently, deep reinforcement learning has produced agents that play video games from raw screen images, while the discovery of collision and physics defects in games continues to depend on repeated manual play-testing. This paper presents Glitch Hunter, an autonomous testing system that couples such an agent with thirteen rule-based checks evaluated on every engine frame of a two-dimensional platform game. Trained with Proximal Policy Optimization on visual input only, the agent first learns to finish the level and is then rewarded for entering unvisited parts of the physically reachable space, while a divergence bound towards a healthy reference policy safeguards the finishing skill. The collision checks compare engine behaviour against the level as rendered and the remaining checks encode general engine rules, so no prior knowledge of fault locations is required. Once a check is violated, testing halts on the offending frame, the evidence is archived, an isolated replay confirms the fault, and Markdown and PDF reports indicate the most likely responsible source lines. Experiments on World 1-1 of a publicly available Super Mario Bros. clone, both unmodified and with six deliberately injected faults, show that the final agent completes 56.6% of 500 seeded episodes, against 46.8% after the first stage, and covers 84.25% of the reachable space. Every injected fault is detected, reproduced and reported, and no false alarm arises on the unmodified game across the measured runs
Zitieren
APA
Kumar, A., G, S. G. S., Singh, T., & H, P. H. P. (2026). Glitch Hunter: Autonomous Game Testing Using Reinforcement Learning. IOSR Journal of Computer Engineering, 28(6), 42-49. https://doi.org/10.9790/0661-2806014249
Vancouver
Kumar A, G SGS, Singh T, H PHP. Glitch Hunter: Autonomous Game Testing Using Reinforcement Learning. IOSR Journal of Computer Engineering. 2026;28(6):42-49. doi:10.9790/0661-2806014249
BibTeX
@article{Kumar2026,
author = {Kumar, Aryan and G, Suman G Suman and Singh, Tanmay and H, Pavana H Pavana},
title = {Glitch Hunter: Autonomous Game Testing Using Reinforcement Learning},
journal = {IOSR Journal of Computer Engineering},
year = {2026},
volume = {28},
number = {6},
pages = {42--49},
doi = {10.9790/0661-2806014249}
}
EUVIMED