Android GUI Test Generation with SARSA

Abstract

Android applications are often challenging to test because of large event spaces with an exponential number of event sequences. Several studies employ reinforcement learning to generate test suites in an effort to optimize code coverage and fault-finding effectiveness under limited testing budgets. In this paper, we generate test cases using the SARSA rein-forcement learning algorithm for seven Android applications, each with a two-hour testing window. The SARSA generated test suites achieve 9.87% to 24.79% better line coverage, 6.9% to 20.09% better branch coverage, 7.88% to 28.48% better method coverage and 3.74% to 35.02% better class coverage than the test suites generated at random by the Monkey tool.