NeurOptimiser

NeurOptimiser is a neuromorphic optimisation framework in which metaheuristic search emerges from asynchronous spiking dynamics. It defines optimisation as a decentralised process executed by interconnected Neuromorphic Heuristic Units (NHUs), each embedding a spiking neuron model and a spike-triggered heuristic rule.

This framework enables fully event-driven, low-power optimisation by integrating spiking computation with local heuristic adaptation. It supports multiple neuron models, perturbation operators, and network topologies.

✨ Key Features

• Modular and extensible architecture using Intel’s Lava.
• Supports linear and Izhikevich neuron dynamics.
• Implements random, fixed, directional, and Differential Evolution operators as spike-triggered perturbations.
• Includes asynchronous neighbourhood management, tensor contraction layers, and greedy selectors.
• Compatible with BBOB (COCO) suite.
• Designed for scalability, reusability, and future deployment on Loihi-class neuromorphic hardware.

📖 Documentation

For detailed documentation, examples, and API reference, please visit the Neuroptimiser Documentation.

📦 Installation

Requirements

• Python 3.10 (tested and recommended version)
• Lava-NC environment configured

Install with pip

# After cloning the repository and navigating to the project directory
pip install -e .
# or just install from PyPI
pip install neuroptimiser

Install with uv

# After cloning the repository and navigating to the project directory
uv pip install -e .
# or just install from PyPI
uv pip install neuroptimiser

You can also use the provided Makefile for additional installation options and commands.

Known Issues

"Too many open files" error: On some systems, you may encounter this error during execution. To fix it:

Unix/Linux/macOS:

ulimit -n 65536

Windows (PowerShell):

# No direct equivalent - typically not needed on Windows
# If issues persist, check system file handle limits via registry

Windows (Command Prompt):

REM Windows typically has higher default limits
REM If needed, adjust via Registry Editor or contact system administrator

🚀 Example Usage

from neuroptimiser import NeurOptimiser
import numpy as np

problem_function    = lambda x: np.linalg.norm(x)
problem_bounds      = np.array([[-5.0, 5.0], [-5.0, 5.0]])

optimiser = NeurOptimiser()

optimiser.solve(
    obj_func=problem_function,
    search_space=problem_bounds,
    debug_mode=True,
    num_iterations=1000,
)

For more examples, please, visit Neuroptimiser Usage

📊 Benchmarking

Neuroptimiser has been validated over the BBOB suite, showing:

• Competitive convergence versus Random Search
• Consistent results across function types and dimensions
• Linear runtime scaling with number of units and problem size

🔬 Citation

@misc{neuroptimiser2025,
  author={Cruz-Duarte, Jorge M. and Talbi, El-Ghazali},
  title        = {Neuroptimiser: A neuromorphic optimisation framework},
  year         = {2025},
  url          = {https://github.com/neuroptimiser/neuroptimiser},
  note         = {Version 1.0.X, accessed on 20XX-XX-XX}
}

🔗 Resources

• 📘 Documentation
• 📜 Paper
• 🧠 Intel Lava-NC
• 🧪 COCO Platform

🛠️ License

BSD-3-Clause License — see LICENSE

🧑‍💻 Authors

• Jorge M. Cruz-Duarte — University of Lille
• El-Ghazali Talbi — University of Lille