Agent-based reinforcement learning framework for model-based testing

Abstract

Model-Based Testing (MBT) is an advanced approach for generating and executing software test cases by leveraging abstract models of the system under test (SUT). While MBT offers the advantage of test automation, it still faces challenges in task parallelization and the selection of appropriate test adequacy criteria. Agent-Based Software Testing (ABST) is a subfield of agent-oriented programming that aims at addressing complex software testing tasks using agents. ABST holds the potential to ameliorate these challenges by offering intelligent monitoring and resource management of testing activities. This study presents an agent-based reinforcement-learning framework for model-based testing, designed with four agent entities: a coordinator agent (CA), a monitor agent (MA), a test case generation and execution agent (TGEA), and a reinforcement learning agent (RLA). We implemented the framework using ModelJUnit as an MBT tool, the SPADE environment as a multiagent system, and Keras to integrate the deep reinforcement learning capabilities. Our findings highlight the framework’s effectiveness in achieving testing objectives. The results suggest a relation between the learning epochs and the complexity of the model, indicating balanced interaction cycles using the acquired knowledge.