An AI-Enhanced Multipath TCP Scheduler for Open Radio Access Networks

Abstract

Multipath transmission technology has recently emerged as a crucial solution to address bandwidth resource constraints and uneven load distribution across access points caused by the surge in data-intensive applications. A well-designed multipath scheduler can improve the quality of service and balance the power consumption in evolving Open Radio Access Networks (O-RANs). However, wireless channel instability and RAN heterogeneity challenge the scheduler’s bandwidth aggregation capability. This paper introduces a Neural Aggregation Bandwidth Optimization (NABO) scheduler for O-RAN, combining bandwidth prediction with scheduling policy optimization. NABO employs an innovative approach by first constructing a Transformer-optimized Throughput (ToT) prediction model based on historical path characteristics. To train the model, we design a system to simulate various network conditions and collect datasets. This model is then integrated into a dual-network collaborative learning framework that combines ToT predictions with heterogeneity levels to guide the scheduler’s optimization process. The ToT model achieves a throughput prediction error of less than 2%. In numerous heterogeneous simulation scenarios and real-world wireless environments, NABO significantly outperforms state-of-the-art multipath transmission methods, with bandwidth aggregation improvements of approximately 51% and 30% over existing benchmarks, respectively. These findings demonstrate NABO’s superior efficacy and potential in enhancing the performance and energy efficiency of O-RANs.