Reinforcement Learning (RL) Based Admission Control in Advance Bandwidth Reservation

The infrastructure provider (InP) guarantees bandwidth and request blocking probabilities (RBPs) as the service level agreement (SLA) metrics for an advance bandwidth reservation (ABR) service. Although the InP can guarantee the total RBPs, users who require a virtual circuit (VC) setup within a short time (a short booking window (BkWd)) tend to experience higher RBPs than the guarantee. This paper proposes a reinforcement learning (RL) based admission control for an ABR service to maximize an InP's profit and decrease the RBPs for short BkWd requests. The proposed RL overbooks resources with different overbooking ratios for requests with different BkWds. Booking-Window Pricing, which offers lower prices for requests with a long BkWd, is utilized. Our proposal achieves higher revenues and lower RBPs than conventional heuristic approaches.