Slice-on-the-Fly: AI-based Network Slicing in O-RAN for Dynamic Traffic Demands

Open Radio Access Network (O-RAN) is expected to support a diverse range of cutting-edge, real-time, and heterogeneous dynamic traffic demands. In response, we formulate a network-slicing resource allocation optimization problem to enhance the Quality of Service (QoS) in O-RAN. The proposed problem is formulated as a Nonlinear Programming (NLP) problem, designed to efficiently meet User Equipment (UE) traffic demands in a dynamic environment while maintaining QoS and optimizing network radio resources. Given the combinatorial nature of this problem, finding an optimal solution is challenging and often impractical in a dynamic traffic environment. To address this, we propose decomposing the prime problem into two sub-problems: a Long-Term Resource Allocation (LTRA) problem focusing on resource allocation decisions, and a Short-Term Resource Scheduling (STRS) problem that manages the scheduling of demanded resources. We propose the Long Short-Term Memory Actor-Critic-based (LS-RLSlice) algorithm, a novel approach that modifies LSTM and Deep Deterministic Policy Gradient (DDPG) algorithms for efficiently solving LTRA and STRS. We perform simulations to show that our proposed algorithm outperforms state-of-the-art schemes and significantly reduces the utilization of network resources. Finally, our proposed solution is validated using the real-world POWDER testbed supporting the O-RAN stack and provides configuration setups for Non-Real-Time and Near-Real-Time RAN Intelligent Controllers (Non-RT RIC and Near-RT RIC).