Study on Computer Network Routing Optimization Algorithm Based on Deep Reinforcement Learning

Traditional routing protocols, such as RIP and OSPF, encounter a series of issues—including slow convergence and high average latency—when faced with the exponential growth of network traffic and varying service demands. To address these challenges, routing algorithms leverage the powerful capabilities of deep reinforcement learning, enabling autonomous learning and optimization of network routing, thereby better adapting to complex network environments and fluctuating traffic demands. For the optimization of continuous state spaces, a routing optimization algorithm named DDPGOR (Deep Deterministic Policy Gradient On Routing) has been designed based on the aforementioned control mechanism. This algorithm employs Long Short-Term Memory (LSTM) networks as recurrent neural networks (RNNs) and incorporates an attention mechanism to capture the relationships among various network traffic streams. In addition, to enhance the correlation among network traffic, we optimized the LSTM network by utilizing bidirectional LSTM networks instead of standard LSTM networks. Bidirectional LSTMs can capture longer-term information and identify correlations in network traffic from both forward and backward directions. Simulation results indicate that both DDPGOR and TD3OR significantly improve convergence and stability, markedly reduce end-to-end latency, and enhance throughput and link utilization compared to traditional routing algorithms. Furthermore, as an optimized version of the DDPGOR algorithm, the TD3OR algorithm outperforms DDPGOR in all performance metrics.