Deep Reinforcement Learning-Based 3C Optimization and UAV Trajectory Planning for Semantic-Aware Edge Intelligence Networks

Abstract

Unmanned aerial vehicle (UAV)-assisted edge intelligence networks have emerged as a promising paradigm for supporting semantic-aware tasks, which rely on the joint orchestration of communication, computation, and caching (3C) resources. This paper investigates a multi-user collaborative urban sensing scenario and aims to minimize the worst-case task completion latency among all users. We formulate a joint optimization problem that integrates multi-dimensional resource allocation, semantic offloading decisions, and UAV trajectory planning. Due to the non-convexity and high dimensionality of the problem, a deep reinforcement learning (DRL)-based approach is developed to obtain efficient solutions. Compared with exhaustive search (ES), the proposed method significantly reduces computational complexity while incurring only marginal performance degradation. Simulation results demonstrate that the proposed joint 3C optimization framework effectively reduces task completion latency compared with benchmark schemes without caching or offloading.