Improving Android GUI Automated Testing via Knowledge-Guided Reinforcement Learning and LLM-Generated Inputs
Wentao Liu, Ya Pan, Hanli Bai · 2025
Recent studies have shown that, despite the increasing application of model-based and machine learning-driven approaches in Android GUI testing, these methods face significant limitations. Model-based techniques, while systematic, are costly to construct and maintain, and deep learning-based methods rely heavily on large labeled datasets, lacking generalization and interpretability. Reinforcement learning offers a promising alternative by balancing exploration and exploitation without extensive labeled datasets but often struggles with inefficient early-stage learning in complex state spaces. Additionally, all these methods face challenges in generating effective text inputs, a critical yet underexplored aspect of GUI testing. To address these issues, this paper proposes GuideBot, a model-enhanced reinforcement learning framework that integrates human prior knowledge to accelerate learning and reduce initial exploration time. GuideBot also incorporates a dynamic adaptive strategy, a fuzzy reward mechanism, and a context-aware, LLM-based text generation module to handle text input challenges effectively.