Automation of User Interface Testing by Reinforcement Learning-Based Monkey Agents.

The complexity of Graphical User Interfaces (GUIs) in consumer applications such as home appliances has significantly risen in recent years. For example, the number of different views in the GUIs has increased from simple selection views to complex sub-menu structures. Alongside the development, both testing complexity and cost have risen drastically. A way of handling this increase is test automation by the use of machine learning algorithms. This work focuses on reinforcement learning-based autonomous grey-box monkey testing for consumer GUIs. As a monkey tester, a Deep Q-Network is interacting with the device under test. Experiments are performed on an oven GUI as well as on a desktop training environment. A known feature representation for monkey testing is compared to three alternative representations, as well as a random agent. A careful selection of the feature representation can improve the exploration performance. Empirical results for the autonomous exploration of GUIs show the usefulness of reinforcement learning-based monkey testing over pure random testing on consumer GUIs. This can lead to an efficiency advantage in practice, as pure random testing is often the status quo in many well-known GUI test-frameworks such as Squish.