Fetching contributors…
Cannot retrieve contributors at this time
45 lines (35 sloc) 1.81 KB

# Introduction

Divide-and-conquer and dynamic programming algorithms have a rich history in computer science for problems with large numbers of variables. Many hard problems that can benefit from quantum computers are too large to map directly to a QPU. To solve a problem with more variables than the available number of qubits, we break the problem into subproblems, solve the subproblems, and then reconstruct an answer to the original problem from the subproblem solutions.

qbsolv is one such decomposing solver. It provides two interfaces:

• :ref:`usage` (CLI)

The tabu algorithm is executed on the problem which is divided into subproblems of several dozen variables each.

• :ref:`python`

The Python interface provides a :class:`~dwave_qbsolv.QBSolv` class wrapper for the qbsolv C code. A dimod sampler can be substituted for the default tabu algorithm.

For a description of the algorithm and implementation, see :download:`Partitioning Optimization Problems for Hybrid Classical/Quantum Execution <../qbsolv_techReport.pdf>`.

For a description of the tabu search algorithm, see Tabu search.

## Example

This example sends 30-variable sub-problems of a 500-variable QUBO to the dwave-neal sampler to be incorporated into the tabu results run in the main loop of qbsolv.

```>>> from dwave_qbsolv import QBSolv
>>> import neal
>>> import itertools
>>> import random
...
>>> qubo_size = 500
>>> subqubo_size = 30
>>> Q = {t: random.uniform(-1, 1) for t in itertools.product(range(qubo_size), repeat=2)}
>>> sampler = neal.SimulatedAnnealingSampler()
>>> response = QBSolv().sample_qubo(Q, solver=sampler, solver_limit=subqubo_size)
>>> print("energies=" + str(list(response.data_vectors['energy'])))   # doctest: +SKIP
energies=[-2800.794817495185]```
You can’t perform that action at this time.