A controlled benchmark of differentiable optimization and reinforcement learning for traffic engineering in transmission backhaul networks

Abstract

Simulation framework and reference implementations for the paper "A controlled benchmark of differentiable optimization and reinforcement learning for traffic engineering in transmission backhaul networks". The archive contains the implementations of all nine routing methods evaluated in the paper — a multi-commodity flow linear program (LP) giving the lower bound on maximum link utilization, Dijkstra shortest-path routing, Equal-Cost Multi-Path (ECMP), offline-tuned OSPF, a multi-agent deep reinforcement learning agent (MADRL), centralized and graph Proximal Policy Optimization agents (PPO, GPPO), and two direct differentiable optimization models (MLP-TE, GNN-TE). It covers every implementation-dependent stage of the study so that each can be reproduced or inspected independently: traffic-matrix preprocessing and anomaly filtering, train/test selection, link-capacity calibration, candidate-path construction, OSPF weight optimization, model training for all five learned methods, link-failure sampling, and the evaluation and timing harness. All random seeds used in the paper are recorded in the configuration files and a README documents the run order. Experiments use three real backbone topologies (Abilene, GEANT, Germany50) with real traffic matrices from SNDlib. Access to the files is restricted during peer review and will be opened on publication.