Dynamic Resource Allocation for Satellite Edge Computing: An Adaptive Reinforcement Learning-based Approach

Abstract

Satellite edge computing has gained traction in geohazard monitoring, agriculture monitoring, and traffic surveillance. However, expanding applications have introduced load imbalance as a challenge. The geolocation-dependency of inference tasks leads to task influx in disaster-prone regions, potentially overloading satellites and impacting system performance. Traditional proximity-based scheduling fails to reduce task waiting latency, causing computational hotspots and overloaded satellites, consuming bandwidth and reducing efficiency. To address this, we propose a micro-satellite cloud architecture enabling each satellite to form a collaborative computing system with neighbors. Resources are divided into private and shared sections, and reinforcement learning is used for adaptive resource allocation. Simulation experiments show reduced waiting latency and failure ratio, improving overall performance.