Guiding Random Graphical and Natural User Interface Testing Through Domain Knowledge

Users have access to a diverse set of interfaces that can be used to interact with software. Tools exist for automatically generating test data for an application, but the data required by each user interface is complex. Generating realistic data similar to that of a user is difficult. The environment which an application is running inside may also limit the data available, or updates to an operating system can break support for tools that generate test data. Consequently, applications exist for which there are no automated methods of generating test data similar to that which a user would provide through real usage of a user interface. With no automated method of generating data, the cost of testing increases and there is an increased chance of bugs being released into production code. In this thesis, we investigate techniques which aim to mimic users, observing how stored user interactions can be split to generate data targeted at specific states of an application, or to generate different subareas of the data structure provided by a user interface. To reduce the cost of gathering and labelling graphical user interface data, we look at generating randomised screen shots of applications, which can be automatically labelled and used in the training stage of a machine learning model. These trained models could guide a randomised approach at generating tests, achieving a significantly higher branch coverage than an unguided random approach. However, for natural user interfaces, which allow interaction through body tracking, we could not learn such a model through generated data. We find that models derived from real user data can generate tests with a significantly higher branch coverage than a purely random tester for both natural and graphical user interfaces. Our approaches use no feedback from an application during test generation. Consequently, the models are “generating data in the dark”. Despite this, these models can still generate tests with a higher coverage than random testing, but there may be a benefit to inferring the current state of an application and using this to guide data generation.