End-to-End RPA-Like Testing Using Reinforcement Learning

Even though test automation has an increased presence in industry nowadays, there is still room for improvement, especially in the area of end-to-end testing. Most testing methods in the literature focus on techniques that do not test these applications as a typical end user would, i.e., starting from the user interface (UI) level. Our work, done in collaboration with UiPath company, a leader in Robotic Process Automation (RPA), proposes deep reinforcement learning methods that can test applications from end to end at the UI level. In the current implementation of our prototype, abstractions and separation of concerns are considered so that methods can be reused between applications and algorithms can be used with minimal user effort. The testing process that results after training the agents is similar to that of a human tester going through the functions of the application. Empirical evaluation of these agents shows that, on the one hand, they can almost perfectly mimic the behavior of human testers and, on the other hand, they can exceed the human performance level.