一句话：提出首个多智能体强化学习 Web GUI 测试系统 MARG，用集中/分布式经验共享协调多智能体，大幅提升探索覆盖与缺陷检出。
基于强化学习（RL）的 Web GUI 测试已被广泛用于自动、智能地探索被测网站，但单个 RL 智能体往往难以充分探索结构复杂、内容动态的大规模网站状态空间。作者观察到多智能体的潜在优势，探索将多智能体强化学习（MARL）用于自动 Web GUI 测试，以提升测试效率与覆盖率。其核心难点在于如何在智能体间共享信息以避免冗余动作、实现有效协作。为此，论文提出首个基于 MARL 的 Web GUI 测试系统 MARG，协调多个测试智能体协同探索网站，并设计两种数据共享方案：以共享 Q-table 实现高效通信的集中式方案，以及通过数据交换降低 Q-table 维护开销的分布式方案。作者在 9 个流行的真实网站上评估 MARG：配置 5 个智能体时，其探索状态数较两种代表性方法平均提升 4.34 与 3.89 倍，检出的失败数相应提升 4.03 与 3.76 倍；相比独立运行相同数量的智能体，MARG 多探索 36.42% 的唯一 Web 状态，表明 MARL 能有效提升 Web GUI 测试的效率与表现。
Reinforcement learning (RL)-based web GUI testing techniques have attracted significant attention in both academia and industry due to their ability to facilitate automatic and intelligent exploration of websites under test. Yet, the existing approaches that leverage a single RL agent often struggle to comprehensively explore the vast state space of large-scale websites with complex structures and dynamic content. Observing this phenomenon and recognizing the benefit of multiple agents, we explore the use of Multi-Agent RL (MARL) algorithms for automatic web GUI testing, aiming to improve test efficiency and coverage. However, how to share information among different agents to avoid redundant actions and achieve effective cooperation is a non-trivial problem. To address the challenge, we propose the first MARL-based web GUI testing system, MARG, which coordinates multiple testing agents to efficiently explore a website under test. To share testing experience among different agents, we have designed two data sharing schemes: one centralized scheme with a shared Q-table to facilitate efficient communication, and another distributed scheme with data exchange to decrease the overhead of maintaining Q-tables. We have evaluated MARG on nine popular real-world websites. When configuring with five agents, MARG achieves an average increase of 4.34 and 3.89 times in the number of explored states, as well as a corresponding increase of 4.03 and 3.76 times in the number of detected failures, respectively, when compared to two state-of-the-art approaches. Additionally, compared to independently running the same number of agents, MARG can explore 36.42% more unique web states. These results demonstrate the usefulness of MARL in enhancing the efficiency and performance of web GUI testing tasks.
DOI 原文 ·
@inproceedings{paperbot8416,
title = {Can Cooperative Multi-Agent Reinforcement Learning Boost Automatic Web Testing? An Exploratory Study},
author = {Yujia Fan and Sinan Wang and Zebang Fei and Yao Qin and Huaxuan Li and Yepang Liu},
booktitle = {ASE},
year = {2024},
doi = {10.1145/3691620.3694983}
}