Skip to content

Welcome to XLRON

Code style: black codecov


Live render of the per-link spectrum allocations made by a Graph Transformer agent trained with RL on DeepRMSA-NSFNET.

XLRON ("ex-el-er-on") is a JAX-based simulation framework for resource allocation in optical networks. It combines a fast simulation engine that runs entirely on accelerator hardware, an integrated PPO trainer, classical heuristics, capacity bound estimators, an end-to-end differentiable physical layer, and a browser GUI — all in one library.

If you are deciding whether XLRON is the right tool for your work, the short version is:

  • It is the most comprehensive of the open-source optical-network simulation libraries — see Comparisons & Speed.
  • It is fast: 222–1,494× faster end-to-end RL training than other libraries; up to \(6 \times 10^6\) steps/s on a single A100. See Comparisons & Speed.
  • It includes an accurate ISRS GN physical layer model with distributed Raman amplification that agrees with state-of-the-art C+L-band experiments to within 0.5 dB. See Physical Layer.
  • It is the first fully differentiable optical-network simulator: gradients flow through the entire pipeline, enabling gradient-based pump optimization and direct RSA optimization. See Differentiable Simulation.
  • It ships all 119 real-world topologies from TopologyBench out of the box, including very large ones like USA100 and TataInd. See Topologies.
  • It is the framework behind the first Graph Transformer trained with RL to consistently match or beat the strongest heuristics for dynamic RMSA. See Graph Transformer for RMSA.
  • It exposes everything through both a CLI and a browser GUI designed to lower the barrier to entry. See GUI.
  • It is fully tested and comprehensively documented — a unit-test suite covers the environments, training loop, heuristics, bounds, and GN model, and every flag, environment, and execution mode is documented on this site.
  • The flat-flag CLI is designed to be driven by LLM coding agents as well as humans: every option is a single hyphenated flag, every parameter is documented, and the same interface is used by the GUI under the hood. See GUI.

XLRON GUI

Get started

Papers and reproduction guides

If you are here because you read one of our papers and want to reproduce a figure, the per-paper reproduction guides give the exact commands and scripts:

  • XLRON: Accelerated Reinforcement Learning Environments for Optical Networks — OFC 2024
  • Reinforcement Learning with Graph Attention for Routing and Wavelength Assignment with Lightpath Reuse — ONDM 2025 (arXiv:2502.14741) — Reproduce
  • Reinforcement Learning for Dynamic Resource Allocation in Optical Networks: Hype or Hope? — JOCN 17(9), D1 (2025), DOI 10.1364/JOCN.559990, arXiv:2406.01919 — Reproduce
  • Comparison of Dynamic Elastic Optical Network Capacity Bound Estimation Methods — submitted to ECOC 2026 — Reproduce
  • XLRON: A Framework for Hardware-Accelerated and Differentiable Simulation of Optical Networks — in preparation — Reproduce
  • Graph Transformers and Stabilized Reinforcement Learning for Large-Scale Dynamic Routing, Modulation and Spectrum Allocation in Elastic Optical Networks — in preparation — Reproduce

A consolidated list of papers with BibTeX entries lives on the Papers page.

  • TopologyBench — 119 real-world optical network topologies, all bundled with XLRON.
  • Gymnax — gym-style API in JAX; XLRON's environment interface follows the Gymnax pattern.
  • PureJaxRL — XLRON's PPO implementation derives from PureJaxRL.
  • Stoix — multi-device sharding inspiration.

Contact and acknowledgements

Maintained by Michael Doherty (michael.doherty.21@ucl.ac.uk) at UCL. Supported by EPSRC grant EP/S022139/1 (CDT in Connected Electronic and Photonic Systems) and EPSRC Programme Grant TRANSNET (EP/R035342/1).