An integrating RAG-LLM and deep Q-network framework for intelligent fish control systems

Abstract

The fish farming industry is advancing by adopting technologies designed to enhance efficiency, productivity, and sustainability. This study investigates integrating a Retrieval-Augmented Generation Large Language Model (RAG-LLM) with a Deep Q-Network (DQN) in autonomous aquaculture. It compares their performance to traditional expert-led methods and other AI-based systems. The developed autonomous system employs ensemble learning of RAG-LLM and DQN, incorporating IoT devices to thoroughly monitor feeding schedules, disease management, growth, and water quality parameters. This integration allows the system to generate optimal policies through majority voting, leveraging pre-trained LLM knowledge to improve initialization conditions and accelerate learning convergence. The hybrid approach of RAG-LLM and DQN demonstrates superior growth rates and rapid stabilization of automation policies. This highlights its potential to enable non-experts to manage fish farms and efficiently scale production for global food sustainability.