DeepRLGUIMAT: Deep Reinforcement Learning-based GUI Mobile Application Testing Approach

• Ano de defesa: 2022
• Autor(a) principal: Ribeiro, Eliane Figueiredo Collins
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Biblioteca Digitais de Teses e Dissertações da USP
• Palavras-chave em Português: Aprendizado de máquina por reforço; Automação de testes; Automated testing; Mobile application Testing; Reinforcement learning; Teste de aplicações móveis
• Link de acesso: https://www.teses.usp.br/teses/disponiveis/55/55134/tde-03062022-162453/

The constant advances in mobile computing technologies and the market demand for new products and applications that serve an increasingly broad audience represent an opportunity and a need to reflect on how to guarantee the quality of these mobile applications, as well as software testing important and fundamental for these applications. Research in this area has been increasingly needed because it serves as evidence of system quality. However, mobile applications have some characteristics and limitations such as the amount of memory to use, battery life, amount of input data, mobile device screen size and different operating systems. The testing tools available are still limited in the testing strategies and criteria they support. Considering the growth in the use of these applications and the challenges that test automation faces, such as many possible combinations of operations, transitions, functionality coverage, interface element changes and fault reproduction, research to overcome these difficulties is encouraged. Thus, studies of Artificial Intelligence techniques, such as Reinforcement Learning, emerge as an opportunity to improve this area for the generation of test cases through the exploration of applications. Objective: This work proposes the DeepRLGUIMAT approach, which uses the technique of deep reinforcement learning, Deep Q-Network algorithm to generate test cases through the exploration of the mobile application using trial and error, producing a variety of input test data according to the Systematic Functional Testing method and following the probability distribution to satisfy the purpose of covering application functionality. Method: An investigation was carried out in the technical literature through a systematic mapping to know the main studies in this area. The approach was designed and developed based on the information acquired, built a tool for concept proof to execute the elaborated strategy. Finally, empirical experiments were carried out in 30 mobile applications. A comparison was made with similar approaches in the literature (tools with Monkey random approach, Model-based approach with Droidbot Machine Learning, and approaches that use DroidbotX reinforcement machine learning and Q-testing) in terms of code coverage metrics, found flaws, and feature coverage. Results: It was observed that the proposed approach achieved greater value for code coverage in instruction, branches, lines of code and methods compared to state-of-the-art tools. The tool achieved the same result as state-of-the-art tools in terms of failures and crashes encountered. The proposed approach exercised more functional operations than the compared tools in functionality coverage. Conclusion: The proposed approach showed promising results, being more effective for navigation and executing operations in applications, generating useful and effective tests that exercise data entry variations that facilitate the coverage of mobile application functionalities and with greater possibilities of identifying the presence of errors/failures.