Scholarlink

<p>Recirculating aquaculture systems (RAS) represent a sustainable approach to intensive fish production, but their operational complexity poses significant challenges for fish farmers. Addressing these challenges requires managing the complexity of multi-objective optimization with shifting priorities, ensuring interpretability of sophisticated AI models for practical users, and adapting control strategies to the dynamic requirements of fish development stages. This study introduces an adaptive multi-objective reinforcement learning framework that dynamically adjusts priorities across different growth stages, addressing a critical gap in current RAS management approaches. The proposed system implements a hierarchical Deep Deterministic Policy Gradient architecture with growth stage-specific policies, complemented by an interpretable visualization framework that makes complex AI decisions accessible to aquaculture practitioners. Experimental validation in a commercial-scale RAS facility with tilapia demonstrated substantial improvements in both production efficiency and system usability. Key findings include an 18.7% improvement in feed conversion ratio compared to standard implementations, while maintaining exceptional water quality stability (98.3% of time within optimal ranges). Particularly noteworthy was the system’s performance during growth phase transitions—historically challenging periods in RAS management—where adjustment time was reduced by 42.5% while maintaining stable environmental conditions. User studies revealed that the interpretable visualization components significantly enhanced farmers’ trust and understanding, with 89% of participants expressing confidence in the system compared to 65% for previous approaches. These results demonstrate how adaptive reinforcement learning with appropriate interpretability tools can bridge the gap between advanced AI techniques and practical aquaculture management, potentially accelerating adoption of intelligent control systems in commercial aquaculture.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Adaptive multi-objective reinforcement learning with interpretable visualization for integrated RAS management across growth cycles

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年11月7日

查看全文
  • Ashwaq M. Alnemari,
  • Wael M. Elmessery,
  • Said Elshahat Abdallah,
  • Abdallah Elshawadfy Elwakeel,
  • Dina Saif,
  • Mohamed Fawzi Abdalshefie Abuhussein,
  • Péter Szűcs,
  • Mohamed Hamdy Eid,
  • Aml Abubakr Tantawy,
  • Fawaz Alzahrani,
  • Atef Fathy Ahmed,
  • Engy S. E. Abdallah,
  • Rahmah N. AlQthanin,
  • Mohamed Ragab

摘要

Recirculating aquaculture systems (RAS) represent a sustainable approach to intensive fish production, but their operational complexity poses significant challenges for fish farmers. Addressing these challenges requires managing the complexity of multi-objective optimization with shifting priorities, ensuring interpretability of sophisticated AI models for practical users, and adapting control strategies to the dynamic requirements of fish development stages. This study introduces an adaptive multi-objective reinforcement learning framework that dynamically adjusts priorities across different growth stages, addressing a critical gap in current RAS management approaches. The proposed system implements a hierarchical Deep Deterministic Policy Gradient architecture with growth stage-specific policies, complemented by an interpretable visualization framework that makes complex AI decisions accessible to aquaculture practitioners. Experimental validation in a commercial-scale RAS facility with tilapia demonstrated substantial improvements in both production efficiency and system usability. Key findings include an 18.7% improvement in feed conversion ratio compared to standard implementations, while maintaining exceptional water quality stability (98.3% of time within optimal ranges). Particularly noteworthy was the system’s performance during growth phase transitions—historically challenging periods in RAS management—where adjustment time was reduced by 42.5% while maintaining stable environmental conditions. User studies revealed that the interpretable visualization components significantly enhanced farmers’ trust and understanding, with 89% of participants expressing confidence in the system compared to 65% for previous approaches. These results demonstrate how adaptive reinforcement learning with appropriate interpretability tools can bridge the gap between advanced AI techniques and practical aquaculture management, potentially accelerating adoption of intelligent control systems in commercial aquaculture.