Deep Reinforcement Learning-Based Resource Allocation for Satellite Internet of Things with Diverse QoS Guarantee.

Area of Science:

• * Satellite Internet of Things (S-IoT)
• * Wireless Communication Resource Management
• * Artificial Intelligence in Networking

Background:

• * Diverse Quality of Service (QoS) demands from large-scale terminals pose significant challenges for S-IoT resource allocation.
• * Existing methods struggle to simultaneously optimize channel allocation and power control for uplink S-IoT scenarios.
• * Efficient resource management is critical for the viability and performance of S-IoT networks.

Purpose of the Study:

• * To develop a deep reinforcement learning (DRL)-based online algorithm for joint channel allocation and power control in S-IoT uplink.
• * To enhance resource efficiency and ensure QoS requirements are met through intelligent decision-making.
• * To propose a practical deployment strategy using transfer learning for onboard training efficiency in space environments.

Main Methods:

• * An intelligent agent utilizing DRL to simultaneously determine transmission channel and power based on contextual information.
• * A weighted normalized reward function balancing success rate, power efficiency, and QoS adherence.
• * A transfer learning mechanism to optimize onboard training and reduce computational load.

Main Results:

• * The proposed DRL algorithm effectively balances success rate and power efficiency while guaranteeing QoS requirements.
• * Demonstrated significant improvements in power efficiency: 60.91% over GA and 144.44% over DRL_RA.
• * Achieved power efficiency close to DRL-EERA, with only a 4.55% difference.
• * Transfer learning enabled deployment with minimal onboard training (100 steps).

Conclusions:

• * The DRL-based approach provides an effective solution for resource allocation challenges in S-IoT uplink.
• * The method offers a practical and efficient way to manage S-IoT resources, crucial for normal operation.
• * Transfer learning significantly reduces the computational burden for deploying AI in space-based S-IoT systems.