| 摘要: |
| 针对低轨卫星网络拓扑高动态性与传统固定模型适应性不足的问题,提出了一种融合深度强化学习(Deep Reinforcement Learning,DRL)与增量进化策略的动态在线调整方法。该方法通过构建集中-分布式增量学习框架,在拓扑变化时仅对受影响的局部子图进行参数更新,同时引入记忆回放机制和动态切片算法,有效平衡历史知识保留与新数据学习。实验结果表明,在星间链路切换率达每秒2次的高动态场景下,该方法的识别准确率较传统DRL提升22%,调整时延降低45%,网络吞吐量稳定性提升30%。 |
| 关键词: 低轨卫星 网络拓扑 动态调整 增量学习 深度强化学习 进化策略 |
| DOI:10.20079/j.issn.1001-893x.250515002 |
|
| 基金项目: |
|
| Dynamic Online Adjustment and Recognition Method Based on Incremental Learning for LEO Satellites |
| SHEN Jie,ZHANG Jiaxin |
| (1.Suzhou Guoxing High-Tch Co.,Ltd.,Suzhou 215200,China;2.School of Information and Communication Technology,Beijing University of Posts and Telecommunications,Beijing 100876,China) |
| Abstract: |
| For the problems of the high dynamic nature of the low Earth orbit(LEO) satellite network topology and the insufficient adaptability of traditional fixed models,the authors propose a dynamic online adjustment method that integrates deep reinforcement learning(DRL) and incremental evolutionary strategy.This method constructs a centralized-distributed incremental learning framework,updates the parameters of the affected local subgraph when the topology changes,and introduces a memory replay mechanism and a dynamic slicing algorithm to effectively balance the retention of historical knowledge and the learning of new data.The experimental results show that in a high dynamic scenario where the inte-satellite link switching rate reaches 2 times per second,compared with that of traditional DRL,the recognition accuracy of this method is increased by 22 |
| Key words: LEO satellites network topology dynamic adjustment incremental learning deep reinforcement learning evolutionary strategy |