Smart Aquaculture Feeding Strategy Optimization Based on Explainable Deep Reinforcement Learning

To address the “black box” nature and low user trust in deep reinforcement learning (DRL)-based smart feeding systems, this study proposes X-DDQN (Explainable Dueling Double Deep Q-Network), an interpretable aquaculture feeding optimization framework. X-DDQN integrates multi-source real-time sensor data into a high-dimensional state space and employs a multi-objective reward function—balancing feed conversion rate, fish growth, welfare, and water quality—to train a DDQN agent toward near-optimal feeding policies. The key innovation lies in incorporating SHAP (SHapley Additive exPlanations) for local decision attribution and CART (Classification and Regression Trees) for global rule extraction, transforming the trained model into an interpretable “IF-THEN” rule set and visual contribution maps. This enhances transparency and allows expert verification. Experimental results show that X-DDQN maintains decision-making performance comparable to standard DDQN while significantly improving interpretability and user trust. The system achieves a Feed Conversion Ratio (FCR) of 1.28 and a Water Quality Compliance (WQC) rate of 94.7 %, outperforming rule-based and fixed strategies. The extracted decision tree attains 92.5 % fidelity to the original DDQN policy. This work provides a trustworthy AI solution for smart aquaculture that harmonizes performance with explainability, supporting human-machine collaborative decision-making in agricultural intelligence.