Counterfactual multi-agent reinforcement learning for causal defect analysis in manufacturing environments

Effective defect management is critical for maintaining product quality in today’s manufacturing environments. However, traditional analysis often relies on correlational models that fail to capture the underlying complexity of defect causation. To address this challenge, we introduce a novel framework, Counterfactual Multi-Agent Reinforcement Learning with Soft-Masked Vector to DAG (CF-MARL-SMV2D). The proposed framework features a collaborative multi-agent policy that explores the causal Directed Acyclic Graph (DAG) space. The framework integrates a learnable domain-informed soft mask to encode expert knowledge and a counterfactual-based reward function to ensure robust causal identification. The framework was validated using a case study from a leading computer server manufacturer. Empirical results demonstrate that the proposed approach outperforms state-of-the-art causal discovery algorithms, achieving Structural Hamming Distance of 3.90 and low False Discovery Rate. Further, the resulting causal feature set enables superior downstream root cause identification with a predictive accuracy of 0.96.