An adaptive pairwise testing algorithm based on deep reinforcement learning

Pairwise testing is an important branch of combinatorial testing that focuses on finding a minimum test suite that satisfies pairwise coverage. However, most existing methods fail to achieve a good balance between exploration and exploitation capabilities when searching for the test suite, or may not fully utilize the information related to the already-generated test cases: This can lead to unsatisfactory performance in combination coverage. To address these limitations, we propose an adaptive pairwise testing framework based on deep reinforcement learning, APT-DRL. Using this, a deep reinforcement learning model for pairwise testing based on the Proximal Policy Optimization (PPO) method is developed. We design the pairwise coverage vector as the state space, and use neural networks to solve the search problem in this huge state space. To reduce the size of the Markov decision space, we also design a masking technique to avoid repeated generation of actions (test cases) that have already been used. We conducted experiments using APT-DRL and eight other baseline algorithms (representing three categories): The results show that APT-DRL, as a novel pairwise testing method, significantly outperforms four random-based pairwise testing methods (RT, ARTsum, FSCS-HD, and FSCS-SD); is comparable to, or surpasses, the two heuristic algorithms (IPOG and AETG); and has better test-suite-generation efficiency and superior effectiveness than the two swarm-intelligence-based algorithms (GSTG and DPSO).