Lightweight deep deterministic policy gradient for edge computing in recirculating aquaculture systems: real-time...

Area of Science:

• Artificial Intelligence
• Aquaculture Technology
• Edge Computing

Background:

• Real-time aquaculture management faces challenges with advanced algorithms in resource-constrained recirculating aquaculture systems (RAS).
• Previous work established the effectiveness of Deep Deterministic Policy Gradient (DDPG) controllers in commercial RAS.
• The need for optimized, lightweight models for edge deployment in RAS is critical.

Purpose of the Study:

• To introduce a lightweight DDPG architecture (Edge-DDPG) for efficient edge computing deployment in RAS.
• To reduce computational complexity and memory footprint for real-time aquaculture management.
• To enable widespread adoption of intelligent control in small to medium-scale aquaculture operations.

Main Methods:

• Developed a lightweight DDPG architecture with compact neural networks and reduced layer dimensions.
• Implemented memory-efficient replay buffers and CPU-optimized operations for ARM-based edge devices.
• Validated the Edge-DDPG framework on Raspberry Pi 4B for real-time performance and resource consumption.

Main Results:

• Edge-DDPG achieved an 85% reduction in computational complexity, retaining 92% of the original model's accuracy.
• Experimental validation showed average inference times of 15.2 ms on Raspberry Pi 4B, meeting real-time control requirements.
• The system demonstrated 94.3% feeding accuracy and 96.1% water quality stability with low memory usage (47 MB).

Conclusions:

• The Edge-DDPG framework offers a computationally efficient solution for real-time aquaculture management on edge devices.
• Significant cost reductions ($56,900 to $8,400) are achievable, facilitating the adoption of intelligent systems.
• This research enables advanced AI-driven control for improved efficiency and sustainability in aquaculture.