Distributed Multi-Agent DBA With Reinforcement Learning for EPON System

The increasing demand for bandwidth hungry applications ranging from immersive content to Internet-of-Everything (IoE) services is placing extreme pressure on current network infrastructures. Moreover, the transition to 6G demands even higher performance, especially in terms of bandwidth, latency, and reliability. Passive Optical Networks (PONs), particularly Ethernet PON (EPON), are known to be the key enablers for high-capacity optical fronthaul. However, traditional Dynamic Bandwidth Allocation (DBA) mechanisms based on centralized scheduling fail to adapt under large-scale, bursty traffic loads. Recently, Machine Learning (ML) centralized DBA has improved performance but remains limited by high complexity, poor scalability and a single point of failure. To address these challenges, we propose a fully distributed DBA framework, leveraging Multi-Agent Reinforcement Learning (MARL). Each Optical Network Unit (ONU) hosts a lightweight agent trained and executed under the Decentralized Training and Decentralized Execution (DTDE) approach. Each agent autonomously adjusts its REPORT messages using queue- and age-aware scaling with Quality-of-Service (QoS) class weighting while operating within the standard Multi-Point-Control-Protocol (MPCP). A non-intrusive Software-Defined-Networking (SDN) layer supervises telemetry and policy distribution without cycle-by-cycle control. Simulation results demonstrate that the proposed architecture with MARL-based DBA significantly reduces mean packet delay, protects highest priority traffic under congestion, improves fairness, lowers the packet loss and increases the system throughput. The proposed distributed, edge intelligent architecture offers a scalable and resilient solution for the next generation based fronthaul networks.