Research On Multi-access Heterogeneous Network Traffic Splitting Algorithm Based On Machine Learning

With the rapid development of modern network technology,the increasing number of devices and the demand for various applications have brought huge mobile network traffic.A single network cannot meet the bandwidth and latency requirements of devices.Therefore,transmitting data through multi-access networks has become an effective way to increase bandwidth and reduce latency.The current heterogeneous network diversion technology is mainly based on MPTCP(Multipath TCP),which has the ability to simultaneously use multiple communication channels to transmit data packets and effectively aggregate the transmission bandwidth of heterogeneous networks.However,the path performance differences between complex and time-varying heterogeneous networks can lead to a decrease in MPTCP transmission performance.Therefore,how to improve the performance of MPTCP in heterogeneous network environments has become a key challenge.This thesis conducts in-depth research on complex and time-varying multi-access heterogeneous network diversion algorithms,and the main technical innovations are as follows:(1)An MPTCP subflow control algorithm(MSCA)based on SVM(Support Vector Machine)prediction model is proposed to address the issues of greedy MPTCP schedulers and throughput degradation caused by large path differences.This method is based on the SDN architecture and continuously monitors the network status using an SDN controller.Based on the monitored path parameters,the impact factors of the network are predicted using an SVM model.The system dynamically configures subflows for each user based on the value of the impact factor,selects the optimal subflow to transmit data during the transmission process,avoids the impact of poor subflow on transmission performance,and thus improves average throughput.The results show that compared with traditional schemes,this method can cope with complex network environments and respond quickly to changes in the network,with an average throughput increase of 10%.(2)An MPTCP scheduling optimization algorithm based on multi-expert learning forward delay prediction is proposed to address the issue of receiver buffer blocking caused by network asymmetry and differences in network performance(latency,jitter)in multi-access heterogeneous network environments.This algorithm predicts forward delay through a established multi-expert learning model.Based on the predicted forward delay,data packets are scheduled to various paths through delay compensation.The simulation results show that the MPTCP scheduling optimization algorithm based on this method can more accurately calculate the forward delay compared to other methods.Compared to FDPS and MPTCP,the algorithm reduces the number of unordered packets by 30% and60%,respectively,and improves throughput by 33% and 42%.