CRL-KEA: A Deep Reinforcement Learning Assisted Evolutionary Algorithm for Multipath Routing Optimization Problem

The rapid growth of Internet traffic necessitates the adoption of effective routing algorithms to ensure faster and more stable network transmission. Single-path routing is insufficient for current demands, and the performance of multipath routing needs enhancement to address the constraints and requirements of practical scenarios. In this paper, we propose a comprehensive reinforcement learning assisted knowledge-based evolutionary algorithm (CRL-KEA) to solve multipath routing problems with the constraints of practical applications. In this method, deep reinforcement learning with a differentiated encoder and decoder (DRL-DC) is utilized to assist in constructing the initial population. By integrating the current network load state, DRL-DC achieves efficient subpaths construction. Moreover, various operators with specific problem knowledge are adopted to guide the solution updates and repair infeasible solutions. In this way, our method enables the rapid provision of high-quality multipath routing schemes for all network flows. Through experiments conducted under various network topologies, we demonstrate that CRL-KEA has significant advantages in both quality and speed.