Automated simulation testing for complex software environments using multi-agent reinforcement learning

The growing complexity of software systems poses major challenges for automated testing in continuous integration and continuous delivery pipelines.This paper proposes multi-agent-based software automated testing, a multi-agent reinforcement learning framework that models testing tasks as a decentralised partially observable Markov decision process.Using the Q-network mixing algorithm, the system enables coordinated decisions across testing agents.Evaluation on a TravisTorrent-based simulation environment shows multi-agent-based software automated testing achieves a 95.2% defect detection rate -showing a 4.7% improvement over the best multi-agent baseline -while reducing test execution time to 70% of conventional rule-based scheduling.Compared with multi-agent deep deterministic policy gradient, it demonstrates a large effect size (Cohen's d = 0.89) in defect detection.These results demonstrate the framework's effectiveness in improving testing quality and efficiency, offering a viable solution for intelligent test automation in complex software environments.