I have a non-convex QUBO problem that I'd like to solve by warm-starting QAOA with a solution obtained from a continuous relaxation solution obtained by a classical algorithm. The specifics of the problem is shown below in the code.

I have 2 questions:

  • In the code below is CPLEX able to solve the original QUBO. However, when I use CPLEX as input to the WarmStartQAOA optimization in QisKit, it tells me it cannot solve because the problem is non-convex?
  • For non-convex problems, there must be an easy reformulation that QisKit and WarmStartQAOA can do on its own since most problems are non-convex. Can someone help me find that functionality in QisKit?
import random

def invert_counts(counts):
    return {k[::-1]:v for k, v in counts.items()}

qp = QuadraticProgram()
p = 1
m = 4
n = 4
for i in range(m):
    for j in range(n):

for i in range(m):
    for p in range(n):
        for i_2 in range(m):
            for p_2 in range(n):
                qp.quadratic_dict[("x_{}_{}".format(i,p), "x_{}_{}".format(i_2, p_2))] = 0
for i in range(m):
    for p in range(n):
        x_i_p = "x_{}_{}".format(i,p)
        curr_entry = qp.quadratic_dict[(x_i_p, x_i_p)]
        qp.quadratic_dict[(x_i_p, x_i_p)]= curr_entry - 40

def relax_problem(problem) -> QuadraticProgram:
    """Change all variables to continuous."""
    relaxed_problem = copy.deepcopy(problem)
    for variable in relaxed_problem.variables:
        variable.vartype = VarType.CONTINUOUS
    return relaxed_problem

sol = CplexOptimizer().solve(qp)

qaoa_mes = QAOA(sampler=Sampler(), optimizer=COBYLA())
ws_qaoa = WarmStartQAOAOptimizer(
    pre_solver=CplexOptimizer(), relax_for_pre_solver=True, qaoa=qaoa_mes, epsilon=0.0)

ws_result = ws_qaoa.solve(qp)
  • $\begingroup$ Did you figure out what was the issue? I get a similar problem. $\endgroup$
    – Stéphane
    Jun 29, 2023 at 16:10


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Browse other questions tagged or ask your own question.