AI‑Powered Compatibility Testing for Mobile Games: Platform Design, Scene Traversal, and Anomaly Detection

Following a previous introduction of the "Intelligent Compatibility Testing Platform," the authors discuss the need for specialized algorithms to handle the diverse and complex scenarios encountered during mobile game compatibility testing.

Three major challenges are identified: the vast variety of Android devices, intricate game play flows with many UI states, and the difficulty of building a universal model to recognize numerous types of UI anomalies.

The proposed solution is a three‑part framework: a cloud‑based real‑device farm for unified device management, a scene‑traversal module that extracts UI trees via the Poco SDK, filters interactive controls, and generates intelligent scripts, and an image‑based anomaly detection module that classifies UI defects using a CNN.

In the scene‑traversal module, UI components are identified from the layout tree, filtered by type and hierarchy, and then selected for interaction using a reinforcement‑learning (UCB) algorithm that favors less‑visited controls. Two traversal modes are offered: a fast mode covering basic screens and a deep mode exploring all sub‑scenes for regression testing.

The anomaly detection module leverages computer‑vision techniques: a CNN classifier trained on normal and synthetic abnormal screenshots, with data augmentation achieved through bug injection (camera flag changes, disabling cameras, adding post‑processing effects) and image corruption. Special handling is provided for cut‑out screen designs (notches, punch‑holes) by overlaying mask data and checking UI overlap.

Experimental results show successful detection of issues such as screen tearing, missing textures, and black borders. Future work includes addressing text‑related UI problems, extending support to MMORPGs, reducing SDK intrusion by using pure image‑based UI detection, and providing easier integration and debugging tools.