基于优化决策树的时延敏感流智能感知调度

Delay-sensitive traffic intellisense scheduling based on optimal decision tree

目前流量调度策略无法做到智能按需化,尤其对于网络突发故障造成的拥塞以及高价值业务的护航场景,无法按需保障时延敏感的业务体验。通过分析研究不同网络业务流量时延敏感性属性需求,探索挖掘不同网络业务流量的行为特征与其时延敏感性需求之间的内在关联关系。然后利用AI技术对这种内在的关联关系进行学习,构建其映射关系,实现了时延敏感流智能感知调度。同时,考虑AI模型的可解释性及可部署性实际问题,采用强化学习剪枝优化可解释性决策树模型,提高模型的鲁棒性同时使模型更轻量化,易于设备部署实现。通过真实网络流量实验,强化学习优化后的决策树模型在单次推理情形下感知正确率提高1.75%,推理速度提升约30%;同时,实验也证明了使用局部微观统计特征多次推理有助于提高模型感知正确率。在所有实验中,强化学习优化的决策树模型规模缩小了60.0%~87.2%,并且Saras比Q-learning具有更好的优化表现。

Currently,network traffic scheduling strategy cannot be intelligent and on-demand,especially in the congestion caused by sudden network failures and escort scenarios of high-value services.They cannot guarantee latency-sensitive service experience on demand.The delay-sensitive attribute requirements of different network traffic were analyzed and studied,and the internal correlation between the behavior characteristics of varying network traffic and its delay sensitivity requirements was explored.Then,AI technology was used to learn this inherent relationship and construct its mapping relationship,realizing a traffic scheduling technical solution based on the intelligent awareness of delay-sensitive traffic.At the same time,considering the practical issues of interpretability and deploy-ability of AI models,reinforcement learning(RL)technology was used to prune and optimize the interpretable decision tree model,which improved the robustness of the model and made model lighter and easier to implement in equipment deployment.Through experiments by the collected real network traffic,the decision tree model optimized by reinforcement learning could improve the awareness accuracy by 1.75%in a single inference case,and the inference performance was improved by about 30%.The experiment also proved that using micro-statistical features for multiple inferences could help improve the model accuracy;in all experiments,the scale of the decision tree model optimized by RL was reduced by about 60.0%~87.2%,and the Saras had better optimization performance than Q-learning.