Dynamic confidence-based constraint adjustment in distributional constrained policy optimization: enhancing supply chain management through adaptive reinforcement learning

In this study, we introduce the dynamic confidence-based constraint adjustment (DCCA) approach, an innovative enhancement to the distributional constrained policy optimization (DCPO) algorithm, tailored to optimize decision-making process in intricate supply chain management environments. DCCA continuously tunes the reshaping parameter in response to real-time confidence estimations in satisfying operational constraints, enabling more adaptive and risk-aware policy updates. Through a comprehensive evaluation involving a multi-echelon, multi-period supply chain case study, DCCA demonstrates superior performance in balancing return maximization with stringent constraint adherence, outperforming traditional baseline algorithms such as Vanilla TRPO, Saute TRPO, CPO, and DCPO. Our results, highlighted by reduced variability in performance metrics and improved average returns, underscore DCCA’s effectiveness in navigating the intricate trade-offs between risk and reward in dynamic supply chain scenarios. This study not only validates DCCA’s theoretical underpinnings but also establishes its practical applicability, offering a promising avenue for advancing supply chain optimization methodologies.