Multi agent federated reinforcement learning for Distributed MPTCP Agents

The integration of Multi access Edge Computing (MEC) with low Earth orbit (LEO) satellite constellations is a promising paradigm for global, low latency connectivity. However, the dynamic topology and heterogeneous link qualities of satellite networks pose significant challenges for efficient multipath transport protocol (MPTCP) scheduling. Traditional schedulers, often based on heuristics or designed for fixed size inputs, struggle to adapt to the variable number of available paths. We propose a novel Multi Agent Federated Reinforcement Learning (MAFRL) framework that leverages Set Transformers for permutation invariant encoding of variable path sets. Each agent learns a local policy using Proximal Policy Optimization (PPO), augmented with a soft fairness constraint to ensure equitable performance. A federated learning scheme, using FedProx aggregation, enables collaborative training across distributed agents without sharing raw data, preserving privacy and improving robustness to non IID data. Extensive emulation experiments show our approach outperforms heuristic and learning based baselines in aggregate throughput, latency, and fairness, particularly under path variability. This work demonstrates the viability of set based learning and federated optimization for intelligent resource management in next generation satellite terrestrial networks.