Modular AI for supply chain resilience: Blending RL, forecasting, causal inference, and evolutionary coordination for inventory optimization

Retail supply chains today face persistent uncertainty, with demand fluctuations, fragile supply networks, and interlinked product relationships making traditional inventory control increasingly inadequate. In response, this study proposes a modular framework of eight hybrid reinforcement learning (RL) models designed to improve decision making under dynamic and uncertain conditions. Each model integrates deep RL with a complementary technique, including causal inference, demand forecasting, graph-based representations, and genetic algorithms (GA), to address specific operational challenges. Some variants incorporate counterfactual reasoning to evaluate policy robustness under alternative scenarios, while others combine reinforcement learning with heuristic and evolutionary coordination mechanisms to enhance decision stability. Forecast-driven RL further supports anticipatory inventory control by incorporating predicted demand signals into the decision process. The proposed framework is evaluated on a single-SKU, single-store subset derived from the Walmart M5 dataset, providing a controlled experimental setting for consistent comparison across all hybrid models. Experimental results indicate that hybrid approaches, particularly counterfactual-enhanced RL and GA-coordinated variants, outperform standard RL baselines in terms of service level, cost efficiency, and decision stability. The models demonstrate improved adaptability to demand variability, promotions, and supply disruptions within the evaluated setting. While the current study focuses on a controlled subset, the proposed framework is inherently modular and designed to be extensible to multi-SKU and multi-store supply chain environments . These findings highlight the potential of hybrid AI-driven approaches for enhancing inventory decision-making under uncertainty.