Evaluating Reinforcement learning for Error Detection in Model-Based Testing

Software testing ensures systems meet requirements and identifies defects while minimizing manual effort, reducing costs, and enhancing automation. Model-Based Testing (MBT) leverages abstract models for automated test case generation and execution. However, traditional exploration techniques often struggle to balance exploration and exploitation, limiting coverage and error detection. Reinforcement Learning (RL) offers a promising alternative, adapting dynamically for test case generation, execution, and prioritization. This study evaluates state-of-the-art RL algorithms—A2C, PPO, and Recur-rentPPO—utilizing an Agent-Based MBT framework. Using four Software Under Test (SUT) models with four error sizes, we assess the effectiveness of RL in error detection. Results suggest RL could outperform traditional techniques in both efficacy and efficiency, excelling at detecting complex errors and exploring extended execution paths. These findings highlight the potential of RL to advance MBT and broader software testing tasks.