QHDOPT (QHD-based OPTimizer) is a software package for nonconvex optimization.
QHDOPT implements a quantum optimization algorithm named Quantum Hamiltonian Descent (QHD) on available quantum computers (such as the D-Wave systems). QHD is a quantum-upgraded version of gradient descent (GD). Unlike the classical GD, QHD demonstrates a significant advantage in solving nonconvex optimization problems.
QHDOPT aims to eliminate the technical barrier of using QHD for the broader operations research (OR) community. We do not assume users to have prior knowledge of quantum computing, while we allow expert users to specify advanced solver parameters for customized experience. Our target users include:
- Professionals pursuing an off-the-shelf nonconvex optimization solver to tackle problems in operations research (e.g., power systems, supply chains, manufacturing, health care, etc.),
- Researchers who hope to advance the theory and algorithms of optimization via quantum technologies,
- Experts in quantum computation who want to experiment with hyperparameters and/or encodings in QHD to achieve even better practical performance.
QHDOPT has a built-in compiler powered by SimuQ, a framework for programming and compiling quantum Hamiltonian systems.
QHDOPT automatically post-processes the results returned by the quantum machines. The post-processing includes decoding the raw measurement results and improving their precision (i.e., fine-tuning) via a classical local solver. Users may disable the fine-tuning if needed.
The dependency of simuq
now requires you to install a specific version of SimuQ manually. Please install by running the following
git clone https://github.com/PicksPeng/SimuQ.git
cd SimuQ/
git checkout qhd-work
pip install ".[dwave, ionq, qutip]"
To install QHDOPT, you can clone this repo and install by
git clone https://github.com/jiaqileng/QHDOPT.git
cd QHDOPT/
pip install ".[all]"
Finally, run the following in your conda environment:
conda install -c conda-forge cyipopt
Two example notebooks for a jump start are examples/QP-example.ipynb
and examples/PO-example.ipynb
. The following illustrates the basic building blocks of QHDOPT and their functionalities briefly.
Import QHDOPT by running
from qhdopt import QHD
You can create a problem instance by directly constructing the function via SymPy.
from sympy import symbols, exp
x, y = symbols("x y")
f = y**1.5 - exp(4*x) * (y-0.75)
model = QHD.SymPy(f, [x, y], bounds=(0,1))
Then you need to setup the solver and the backend device (D-Wave in this example).
model.dwave_setup(resolution=8, api_key="API_key")
Here resolution
represents the resolution of the QHD algorithm, and api_key
represents the API key of the D-Wave account obtained at D-Wave Leap.
Now you can solve the target problem.
minimum = model.optimize()
The minimal value of minimum
. To print more details in the process, you can run model.optimize(verbose=1)
.
Jiaqi Leng [email protected]
Yuxiang Peng [email protected]
Samuel Kushnir, Jiaqi Leng, Yuxiang Peng, Lei Fan
If you use QHDOPT in your work, please cite our paper
@article{kushnir2024qhdopt,
title = {QHDOPT: A Software for Nonlinear Optimization with Quantum Hamiltonian Decent},
author = {Kushnir, Sam and Leng, Jiaqi and Peng, Yuxiang and Fan, Lei and Wu, Xiaodi},
journal = {arXiv preprint arXiv:xxxx.xxxxx},
year = {2024}
}