A Testability Strategy Optimization Method Under Multi-Valued Dependency Condition Based on Deep Reinforcement Learning

Full abstract

Read the full abstract

The multi-valued dependency matrix (MVD matrix) is an important testability modeling approach, which can deliver more comprehensive testability information than the traditional dependency matrix (D-matrix). However, existing testability strategy optimization algorithms perform poorly in handling the MVD matrix, and the high-dimensional MVD matrix further aggravates these limitations as system complexity increases. To address these problems, a novel testability strategy optimization method under multi-valued dependency conditions based on deep reinforcement learning (DRL) is proposed. Firstly, the sets of elements and two reward functions to minimize test sequence length and test cost are established from the MVD matrix. Subsequently, the algorithm for selecting test points based on Deep Q-Network (DQN) is proposed. The DQN parameters are updated to fit the Q-value of test points. Thirdly, Double DQN (DDQN) and the prioritized experience replay (PER) mechanism are introduced to address the overestimation problem and sample redundancy problem, respectively, in high-dimensional matrix environments. The experimental results show that the testability strategy generated by this method can isolate all faults with fewer steps or at a lower cost. In a high-dimensional matrix environment, it can reduce test costs compared with the other heuristic algorithms while maintaining a good level of stability.