Traditional hierarchical optimization methods in next-generation communication networks struggle with resource allocation in dynamic environments, leading to reduced throughput, increased latency, and low energy efficiency. To address this, we propose a cross-layer collaborative optimization method using Deep Q-Network (DQN). First, we model cross-layer resource scheduling as a Markov Decision Process (MDP), enabling joint optimization across physical, link, and network layers through DQN agents. An experience replay mechanism and target network enhance training stability and sample efficiency. Experiments in heterogeneous networks demonstrate that our method achieves 48.1% higher throughput, 59.4% lower latency, and 16% improved bandwidth utilization under 80% load compared to traditional greedy algorithms. These results validate the method’s adaptability and robustness in dynamic environments, offering a novel solution for intelligent resource management in future networks.