MAAT: Applied Computational Intelligence for Autonomous Software Testing via Deep Learning, Reinforcement Learning and Knowledge-Based Reasoning

Abstract

Software testing consumes 30–40% of total development effort in modern continuous integration pipelines, yet existing automation tools treat it as a static, hand-coded problem rather than an applied intelligence challenge. We present MAAT (Multi-Agent Autonomous Testing), a computational intelligence framework that orchestrates five s pecialized A I a gents through a synergistic combination of three CI paradigms: (i) a deep learning Foundation LLM Engine for natural-language-to-test synthesis; (ii) a reinforcement learning Orchestrator modelled as a Markov Decision Process (MDP) with empirically validated reward parameters and convergence analysis; and (iii) knowledge-based reasoning via a typed property graph (CodeEntity, TestEntity, DefectEntity) enabling semantic coordination across agents. A three-stage LLM validation pipeline with confidence-scored retry achieves 97.9% test acceptance, directly addressing LLM reliability concerns. Evaluation on five o pen-source projects (Apache Commons, Spring Boot, TensorFlow Lite, React Testing Library, Kubernetes; 145K–950K LOC) demonstrates 94.3% statement coverage (95% CI: 93.7–94.9), 91.7% defect detection and 83% generation speed-up over manual methods with a 3.2% false positive rate; all gains are significant under Holm-corrected Wilcoxon tests (p<0.001, Cliff’s δ ≥ 0.71). An industrial case study on a production fintech microservices platform validates real-world viability: MAAT quarantined 23 previously misclassified flaky tests, caught a critical race condition and raised coverage from 79.2% to 91.4% within a 22-minute CI budget.