Research on Deep Reinforcement Learning-Based Resource Allocation Strategies for Wireless Networks

The evolution to 5G-Advanced and 6G networks demands intelligent resource allocation solutions capable of trading off dynamic channel conditions, multi-service QoS trade-offs, and edge deployment constraints. This paper presents a lightweight deep reinforcement learning (DRL) framework with hierarchical spatiotemporal attention mechanisms and edge-centric co-design principles. A three-stage curriculum learning strategy progressively exposes the policy network to difficult scenarios—from static channel conditions to 60 m/s high-mobility cases—at 2.1× speed over conventional training protocols. The HAN proposed herein leverages temporal LSTM layers to track Doppler-induced channel variations and spatial CNN transformers to identify interference-critical user pairs, achieving 40% latency reduction for feature extraction and 22% improvement for policy stability.