Choosing a solver¶

How to solve on a solver other than HiGHS, pass it options, or write the model to a file for a solver specsolve does not run. HiGHS ships with specsolve and is the default. Nothing in the spec names a solver, so the call chooses one.

See which solvers this build has¶

Name a solver that does not exist, and the refusal lists the ones that do. This block runs when the site is built, so the list is this commit's:

import specsolve as sps

try:
    sps.solve('examples/dispatch.yaml', {}, solver_name='cplex')
except sps.errors.SpecsolveError as exc:
    print(exc)

unknown solver 'cplex' — this build solves with gurobi, highs, xpress.

The name is checked before any data is attached.

Install the solver¶

Every solver except HiGHS comes with the extra of the same name:

pip install "specsolve[gurobi]"

A name the build knows, but whose package is not installed, raises ModuleNotFoundError with the pip install line to run.

Name it in the call¶

solver_name= goes on solve, Model.solve and solve_over:

result = sps.solve('dispatch.yaml', sources, solver_name='gurobi')

Pass solver options¶

solver_options= goes to the solver verbatim, in its own vocabulary. A time limit is time_limit on HiGHS and TimeLimit on Gurobi:

result = sps.solve('dispatch.yaml', sources, solver_name='gurobi', solver_options={'TimeLimit': 60})

Solve an LP by interior point without crossover¶

An LP can be solved by interior point and left there, without the crossover that moves the answer to a vertex:

The answer is optimal to the solver's tolerance and carries duals. It ends on no basis, so it has no variable_basis or constraint_basis, and an LP cannot start from it.

On HiGHS, measure before you turn it on. On the benchmark models at rung m, it was the fastest of the three ways only on transport, by 14%. Simplex was the fastest on dispatch and storage. On nodal, interior point with crossover was the fastest, and without crossover it took 28 times as long. Every objective was within 7e-9, relative, of the simplex one (#1879). Time each way on your own model before you keep one.

A solver that cannot take the model¶

Solvers take different constructs: HiGHS has no SOS sets, for example. solve refuses a built model the solver cannot take, before the load, and the refusal names the solvers that do take the construct. To ask without solving, build and check:

import specsolve as sps

model = sps.build('dispatch.yaml', sources)
model.check('gurobi')

The full table is what each sink takes.

The answer is read off the model the build produced, not off the file. A file may declare a quadratic cost that the data prices at zero everywhere, or an integer variable that a where: leaves without a column; that model is an LP, and every solver takes it.

Write the model to a file¶

For a solver or tool that specsolve does not run, write the model. The suffix picks the format, and this build writes .lp and .mps:

sps.write('dispatch.yaml', sources, 'dispatch.lp')

The two formats carry different constructs: MPS has no quadratic objective or indicator, for example. The table above covers both.