Progress on class integration test order generation approaches: A systematic literature review

Class integration test order generation approach based on Sarsa algorithm

“…When n actions have been selected, the state space S is presented as a tree diagram, which stores a maximum of n n+1 −1 (n−1) states (Czibula et al 2018), and the action space stores n actions, which form a possible action order, which is a candidate class integration test order. The reward function of the method is designed primarily around the stubbing complexity (Equation 7) in the reinforcement learning-based approach proposed by Czibula et al (2018) and Ding et al (2022), so that the agent will be biased to select the class with the lowest stubbing complexity to be constructed, and achieve the purpose of generating a final CITO with the lowest possible stubbing cost. However, the graph-based approach uses the principle of breaking as many cycles as feasible by removing the fewest number of edges, and the stubbing cost is related to the cost of breaking cycles.…”

“…In this paper, the overall stubbing complexity is used as the evaluation indicator to assess the effectiveness of the proposed class integration test order generation approach based on Sarsa algorithm (CITO Sarsa), and the CITO with the lowest stubbing complexity is selected using the stubbing complexity measure formula in Section 2.1.2. The results in Table 7 are acquired by running 30 times in eight systems and choosing the average value, and comparing the approach in this work with the graph theory approach (GT) of Tai and Daniels (1997), the genetic algorithm (GA) presented by Briand et al (2004), the particle swarm algorithm (PSO) of Zhang et al (2018a), the random interaction algorithm (RIA) of Wang et al (2010), and the reinforcement learning-based approach (CITO RL) proposed by Ding et al (2022), other comparative approach parameter settings are consistent with each literature. Table 7 compares the overall stubbing complexity OCplx data produced by running the experimental systems 30 times, with the lowest values obtained by each system using different approaches highlighted in bold.…”

“…Research was conducted in terms of both reward function design and reward strategy to improve the bug detection capabilities of continuous integration test orders. Ding et al (2022) proposed a reinforcement learning-based approach to solving the CITO problem, which included using the Q-learning algorithm, calculating the weights of method complexity and attribute complexity using the entropy weight method, and designing an action selection strategy in reinforcement learning, with the final experimental results being effective in terms of overall stubbing complexity.…”