Enhancing interpretability and explainability for fish farmers: decision tree approximation of DDPG for RAS control

Abstract

Recirculating Aquaculture Systems (RAS) have emerged as a sustainable approach to intensive fish production. In our previous work, we demonstrated the efficacy of a Deep Deterministic Policy Gradient (DDPG) reinforcement learning approach for optimizing feeding rates and maintaining water quality in RAS. However, the “black box” nature of deep reinforcement learning algorithms often impedes their adoption by fish farmers who lack specialized knowledge in artificial intelligence. This study addresses this gap by developing and evaluating methodologies to enhance the interpretability and explainability of DDPG-based control systems specifically for aquaculture practitioners. We employ decision tree approximation techniques to transform complex neural network policies into transparent rule-based systems while preserving performance. Our approach achieves 92.7% policy fidelity with only a 5.8% performance drop compared to the original DDPG controller. The resulting decision rules were validated with fish farmers, with 87% reporting improved understanding of system recommendations and 78% expressing increased trust in the technology. Feature importance analysis revealed dissolved oxygen, ammonia levels, and biomass as the primary decision factors. This work demonstrates how advanced artificial intelligence techniques can be made accessible to end-users through interpretable approximations, potentially accelerating the adoption of intelligent control systems in aquaculture and contributing to more sustainable food production practices.