Intelligent Network Slicing in 5G: A Multi-Agent Deep Reinforcement Learning Framework for Dynamic Resource Orchestration

Abstract This paper presents a comprehensive framework for AI-driven dynamic network slice orchestration in 5G networks. We propose a Deep Reinforcement Learning-based Network Slice Orchestrator (DRL-NSO) employing a multi-agent system to optimize resource distribution across enhanced Mobile Broadband (eMBB), Ultra-Reliable Low-Latency Communication (URLLC), and massive Machine-Type Communication (mMTC) network slices. The framework integrates centralized training with decentralized execution (CTDE), enabling slice-aware optimization while maintaining inter-slice coordination. Our theoretical analysis demonstrates polynomial-time computational complexity O (|S|·|A|·d·h·w) suitable for real-time operation. Economic feasibility assessment indicates potential operational cost reductions of 11. 8-34. 2 million annually for large operators, with payback periods of 12-24 months and 5-year NPV of 22. 5-85. 3 million. Index Terms—5G networks, artificial intelligence, deep reinforcement learning, multi-agent systems, network slicing, resource optimization