D3PG-Light: A Lightweight and Stable Resource Scheduling Framework for UAV-Integrated Sensing, Communication, and...

Area of Science:

• Wireless Communication Networks
• Artificial Intelligence
• Robotics

Background:

• Unmanned Aerial Vehicles (UAVs) are increasingly vital for Integrated Sensing, Communication, and Computation (ISCC) in next-generation wireless networks.
• Static resource allocation is inefficient for UAV-driven ISCC systems due to dynamic environments and resource constraints.
• Minimizing response time is critical for time-sensitive sensing data in mission-critical applications.

Purpose of the Study:

• To dynamically allocate communication bandwidth, sensing resources, and computing power in a UAV-driven ISCC system.
• To reduce system latency while maintaining sensing quality and energy efficiency.
• To propose a novel deep reinforcement learning framework, D3PG-Light, for real-time resource scheduling under UAV hardware constraints.

Main Methods:

• Developed D3PG-Light, a stability-enhanced refinement of deep reinforcement learning, incorporating adaptive gradient stabilization, Long Short-Term Memory (LSTM), and feature fusion.
• Tailored the framework for real-time resource scheduling considering UAV hardware limitations.
• Validated the approach using simulations based on real air-ground channel measurements.

Main Results:

• D3PG-Light demonstrated faster convergence and more stable learning compared to DDPG, TD3, and the original D3PG.
• Achieved a significant reduction in 95th-percentile latency, from over 100 ms to approximately 24 ms.
• Required fewer than 50,000 model parameters, indicating efficiency.

Conclusions:

• D3PG-Light is effective for latency-sensitive UAV-ISCC applications.
• The proposed method enhances training stability and performance in dynamic wireless environments.
• The framework offers a practical solution for real-time resource management in resource-constrained UAV systems.