| 摘要: |
| 针对现有智能路由技术无法适用于动态拓扑的不足，提出了一种面向动态拓扑的深度强化学习智能路由技术，通过使用图神经网络近似PPO(Proximal Policy Optimization)强化学习算法中的策略函数与值函数、策略函数输出所有链路的权值、基于链路权值计算最小成本路径的方法，实现了路由智能体对不同网络拓扑的泛化。仿真结果表明，所提方法可适应动态拓扑的变化并具有比传统的最短路由算法更高的网络吞吐量。 |
| 关键词: 智能通信网络 智能路由 深度强化学习 图神经网络 动态拓扑 |
| DOI： |
|
| 基金项目: |
|
| Deep reinforcement learning routing for dynamic topology networks |
| WU Yuansheng |
| (Southwest China Institute of Electronic Technology,Chengdu 610036,China) |
| Abstract: |
| For the shortcoming that most state-of-the-art deep reinforcement learning(DRL)-based routing solutions cannot be used in networks with dynamic topology,a DRL-based routing solution for dynamic topology networks is proposed.Graph networks are used to approximate policy function and value function in proximal policy optimization(PPO) reinforcement learning algorithm,the link weights are output by the graph nets policy function and the least weight path is computed by the traditional constrained shortest path routing algorithm based on the link weights,thus achieving the generalization of the routing agent to different network topologies.Simulation results indicate that the proposed routing solution can adapt to dynamic network topology and outperforms the shortest path routing in network throughputs. |
| Key words: intelligent communication network intelligent routing deep reinforcement learning graph neural network dynamic topology |
