Coherent Continous-Variable Machine Simulators

This software package includes solvers for continuous optimization problems. The solvers are simulators of coherent continuous-variable machines (CCVM), which are novel coherent network computing architectures based on NTT Research’s coherent Ising machines (CIM). In CCVMs, the physical properties of optical pulses (e.g., mean-field quadrature amplitudes and phase) represent the continuous variables of a given optimization problem. Various features of CCVMs can be used along with programming techniques to implement the constraints imposed by such an optimization problem. Included are methods for solving box-constrained quadratic programming (BoxQP) problems using CCVM simulators by mapping the variables of the problems into the random variables of CCVMs.

Table of Contents

0. Requirements
1. Quickstart
2. Usage
3. Documentation
   • BoxQP Problem Definition
   • ccvmplotlib
4. Contributing
5. References
6. License

Requirements

• Python (supported version: 3.10)

Supported operating systems

• macOS (Monterey, Big Sur)
• Ubuntu (20.04)

Quickstart

1. Once you have cloned the repository, install the package using pip.

   pip install ccvm-simulators

2. Go into examples/ and run the demo scripts.

   • ccvm_boxqp_dl.py: Solve a BoxQP problem using a CCVM simulator, w/o plotting
   • ccvm_boxqp_plot.py: Solve a BoxQP problem using a CCVM simulator, w/ time-to-solution (TTS) plotting
   • For more demo scripts see examples/README.md

3. View the generated plot.

   • The ccvm_boxqp_plot.py script solves a single problem instance, and will create an image of the resulting TTS plot in a plots/ directory. The resulting output image, DL-CCVM_TTS_cpu_plot.png, will look something like this:

Extending the Example

4. Plotting larger datasets
   • The ccvm_boxqp_plot.py script is a toy example that plots the TTS for only a single problem instance.
   • It can be extended to solve multiple problems over a range of problem sizes.
   • When solution metadata is saved for multiple problems, the graph becomes more informative, for example:

5. Other types of plots
   • ccvmplotlib can also plot the success probability data, for example:

Usage

Solving a BoxQP Problem

1. Import Modules

from ccvm_simulators.problem_classes.boxqp import ProblemInstance
from ccvm_simulators.solvers import DLSolver

2. Define a Solver

solver = DLSolver(device="cpu", batch_size=100)  # or "cuda"
solver.parameter_key = {
    20: {"pump": 2.0, "dt": 0.005, "iterations": 15000, "noise_ratio": 10},
}

3. Load a Problem Instance

boxqp_instance = ProblemInstance(
    instance_type="test",
    file_path="./examples/benchmarking_instances/single_test_instance/test020-100-10.in",
    device=solver.device,
)

4. Scale the Coefficients

The BoxQP problem matrix Q and vector V are normalized to obtain a uniform performance across different problem sizes and densities. The ideal range depends on the solver. For best results, Q should be passed to the solver's get_scaling_factor() method to determine the best scaling value for the problem–solver combination.

boxqp_instance.scale_coefs(solver.get_scaling_factor(boxqp_instance.q_matrix))

5. Solve the Problem Instance

solution = solver(
    instance=boxqp_instance,
    post_processor=None,
)

print(f"The best known solution to this problem is {solution.optimal_value}")
# The best known solution to this problem is 799.560976

print(f"The best objective value found by the solver was {solution.best_objective_value}")
# The best objective value found by the solver was 798.1630859375

print(f"The solving process took effectively {solution.solve_time} seconds to solve a single instance")
# The solving process took 0.008949262142181396 seconds

Documentation

The package documentation can be found here.

Additional links:

• Problem definition: BoxQP problem class
• Plotting library: ccvmplotlib

Contributing

We appreciate your pull requests and welcome opportunities to discuss new ideas. Check out our contribution guide and feel free to improve the ccvm_simulators package. For major changes, please open an issue first to discuss any suggestions for changes you might have.

Thank you for considering making a contribution to our project! We appreciate your help and support.

References

This repository contains architectures and simulators presented in the paper "Non-convex Quadratic Programming Using Coherent Optical Networks" by Farhad Khosravi, Ugur Yildiz, Martin Perreault, Artur Scherer, and Pooya Ronagh.

License

APGLv3