DL-MTS: Dynamic Learning-Based Multipath Traffic Steering for Differentiated Service Customization

Traditional multipath scheduling algorithms exhibit insufficient flexibility in their path selection strategies, constraining their ability to accommodate the requirements of diverse services and effectively reconcile service requirements, path conditions, and priorities. This often leads to inefficient resource allocation, network congestion, and an inability to meet diverse service performance requirements. To address these limitations, this paper proposes Dynamic Learning-Based Multipath Traffic Steering (DL-MTS), employing an algorithm based on the Contextual Multi-Armed Bandit (CMAB). By continuously monitoring path conditions and leveraging a rewardpunishment mechanism, the proposed approach dynamically adjusts scheduling strategies to mitigate congestion and maximize available bandwidth utilization. Moreover, it incorporates service priority considerations, ensuring that high-priority services receive preferential bandwidth allocation across different service requirements. Experimental results demonstrate that, compared to traditional static scheduling methods, the DL-MTS offers superior adaptability and resource allocation efficiency, better meets the bandwidth requirements of services, and significantly reduces delay jitter, and improving overall transmission performance.