Boosting input data sequences generation for testing EFSM-specified systems using deep reinforcement learning

• Published in: Information and software technology Vol. 155; p. 107114
• Main Authors: Shu, Ting, Wu, Cuiping, Ding, Zuohua
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.03.2023
• Subjects: Deep reinforcement learning; EFSM; Feasible test sequences; Input data; Input data; Deep reinforcement learning; EFSM; Feasible test sequences
• ISSN: 0950-5849; 1873-6025
• DOI: 10.1016/j.infsof.2022.107114

Input data sequence (IDS) is an important component of test sequences for testing from the Extended Finite State Machine (EFSM) model. During test generation, frequent IDS derivation is time-consuming. Therefore, it has become one of the key factors that restrict the efficiency of test sequences generation. To address this issue, this paper introduces deep reinforcement learning (DRL) to propose a novel approach named IDSG-DRL to accelerate IDS derivation. Our method first formalizes the problem of generating IDS for EFSM-based testing as a Markov decision problem. It then incorporates the DRL algorithm to learn experience from previous input data generation and train a decision-making model to significantly enhance the efficiency of subsequent data derivation for the newly generated test sequences. To improve the convergence of DRL algorithm and ensure the success rate of data generation, a state representation based on variable deviation and action formulation using adaptive exploration are elaborately designed. Finally, a DRL-based algorithm for efficiently yielding IDS is presented for any subject test sequence. We evaluate the proposed approach against the random method, GA-based method as well as a particle swarm optimization (PSO) based method. Experimental statistics show that IDSG-DRL significantly outperforms the baselines in terms of iteration steps, runtime cost, and the success rate of input data derivation. Specifically, compared to random, GA-based and PSO-based methods, IDSG-DRL can reduce the average number of iteration steps by up to 87.09%, 78.57%, and 56.35%, respectively. Regarding the average runtime, our approach is about 3.52 and 1.58 times faster than the GA-based and PSO-based methods. Additionally, given a larger input range, we observed that the performance of IDSG-DRL is more stable and its advantages are more significant. The experimental results suggest that our method is very promising to speed up IDS generation for EFSM-based testing.