GDPopt LBB sends linear MIP relaxed-node subproblem to minlp_solver

[bernalde]

Summary

gdpopt.lbb can create a linear MIP relaxed-node subproblem and then dispatch it through minlp_solver instead of mip_solver. When the MINLP role is configured as GAMS/DICOPT and the MIP role is configured as GAMS/Gurobi, this fails before solve because the GAMS writer rejects DICOPT for a MIP model.

This appears specific to the LBB relaxed-node path. In the same environment, the LOA/GLOA/RIC paths route linear discrete or fixed subproblems through the MIP solver role.

Related but distinct: #3941 tracks the LBB time-limit finalization path calling stale _get_final_results_object.

Steps to reproduce the issue

The following small linear GDP reproduces the dispatch problem. It requires GAMS with DICOPT, IPOPTH, and Gurobi solver names available; the important point is that DICOPT is assigned to the MINLP role while Gurobi is assigned to the MIP role.

$ python lbb_mip_dispatch_repro.py

# lbb_mip_dispatch_repro.py
import pyomo.environ as pyo
from pyomo.gdp import Disjunct, Disjunction

m = pyo.ConcreteModel()
m.x = pyo.Var(bounds=(0, 10))
m.y = pyo.Var(bounds=(0, 10))

m.d1 = Disjunct()
m.d2 = Disjunct()
m.d1.c = pyo.Constraint(expr=m.x + m.y <= 2)
m.d2.c = pyo.Constraint(expr=m.x - m.y >= 1)
m.disj = Disjunction(expr=[m.d1, m.d2])
m.obj = pyo.Objective(expr=m.x + 2 * m.y, sense=pyo.maximize)

pyo.SolverFactory("gdpopt.lbb").solve(
    m,
    tee=True,
    time_limit=10,
    mip_solver="gams",
    mip_solver_args={
        "solver": "gurobi",
        "add_options": ["option reslim=10;"],
        "tee": True,
    },
    nlp_solver="gams",
    nlp_solver_args={
        "solver": "ipopth",
        "add_options": ["option reslim=10;"],
        "tee": True,
    },
    minlp_solver="gams",
    minlp_solver_args={
        "solver": "dicopt",
        "add_options": [
            "option reslim=10;",
            "option nlp=ipopth;",
            "option mip=gurobi;",
        ],
        "tee": True,
    },
    local_minlp_solver="gams",
    local_minlp_solver_args={
        "solver": "dicopt",
        "add_options": [
            "option reslim=10;",
            "option nlp=ipopth;",
            "option mip=gurobi;",
        ],
        "tee": True,
    },
)

Error Message

Starting GDPopt version 22.5.13 using LBB algorithm
...
Original model has 2 constraints (0 nonlinear) and 1 disjunctions, with 4 variables, of which 2 are binary, 0 are integer, and 2 are continuous.
Nodes: 0 LB -inf Unbranched 1
Nodes: 1 LB -inf Unbranched 1
Exploring node 0 with LB -inf UB inf and 1 inactive disjunctions.

Solved in 0 iterations and 0.30524 seconds
Optimal objective value -inf
Relative optimality gap nan%
09/06/22: The GDPopt LBB algorithm currently has known issues. Please use the results with caution and report any bugs!
No feasible solutions found.
Traceback (most recent call last):
  File "<stdin>", line 15, in <module>
  File "/home/bernalde/repos/gdplib/.pixi/envs/default/lib/python3.12/site-packages/pyomo/contrib/gdpopt/branch_and_bound.py", line 90, in solve
    return super().solve(model, **kwds)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/bernalde/repos/gdplib/.pixi/envs/default/lib/python3.12/site-packages/pyomo/contrib/gdpopt/algorithm_base_class.py", line 128, in solve
    self._solve_gdp(model, config)
  File "/home/bernalde/repos/gdplib/.pixi/envs/default/lib/python3.12/site-packages/pyomo/contrib/gdpopt/branch_and_bound.py", line 255, in _solve_gdp
    new_node_data = self._evaluate_node(node_data, node_model, config)
                    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/bernalde/repos/gdplib/.pixi/envs/default/lib/python3.12/site-packages/pyomo/contrib/gdpopt/branch_and_bound.py", line 410, in _evaluate_node
    new_lb, new_ub = self._solve_rnGDP_subproblem(node_model, config)
                     ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/bernalde/repos/gdplib/.pixi/envs/default/lib/python3.12/site-packages/pyomo/contrib/gdpopt/branch_and_bound.py", line 441, in _solve_rnGDP_subproblem
    result = SolverFactory(config.minlp_solver).solve(
             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/bernalde/repos/gdplib/.pixi/envs/default/lib/python3.12/site-packages/pyomo/solvers/plugins/solvers/GAMS.py", line 882, in solve
    _, smap_id = model.write(
                 ^^^^^^^^^^^^
  File "/home/bernalde/repos/gdplib/.pixi/envs/default/lib/python3.12/site-packages/pyomo/core/base/block.py", line 1995, in write
    filename, smap = problem_writer(self, filename, solver_capability, io_options)
                     ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/bernalde/repos/gdplib/.pixi/envs/default/lib/python3.12/site-packages/pyomo/repn/plugins/gams_writer.py", line 542, in __call__
    self._write_model(
  File "/home/bernalde/repos/gdplib/.pixi/envs/default/lib/python3.12/site-packages/pyomo/repn/plugins/gams_writer.py", line 783, in _write_model
    raise ValueError(
ValueError: GAMS writer passed solver (dicopt) unsuitable for model type (mip)

Information on your system

Pyomo version: 6.10.0
Python version: 3.12.13
Operating system: Linux-6.6.114.1-microsoft-standard-WSL2-x86_64-with-glibc2.39
How Pyomo was installed (PyPI, conda, source): Pixi-managed environment
Solver (if applicable): GAMS 51.2.1 with DICOPT, IPOPTH, and Gurobi roles

Additional information

The direct cause appears to be pyomo/contrib/gdpopt/branch_and_bound.py: _solve_rnGDP_subproblem() creates a Big-M transformed subproblem and calls SolverFactory(config.minlp_solver).solve(...) unconditionally. For a linear GDP, that transformed relaxed-node subproblem is a MIP, so the configured MIP solver role is the natural fit.

Expected behavior: LBB should either dispatch a linear relaxed-node subproblem through mip_solver, or fail earlier with a clearer configuration error/documentation that the minlp_solver role may receive MIP models in LBB.

Downstream GDPlib context:

• GDPlib model issue: Refactor med_term_purchasing to fix benchmark issues SECQUOIA/gdplib#69
• GDPlib PR with local benchmark-case routing workaround: Fix med-term benchmark solver routing SECQUOIA/gdplib#137

In GDPlib's med_term_purchasing benchmark case, the direct gdp.bigm and gdp.hull reformulations are linear MIPs. A bounded LBB probe with gams_minlp_solver=dicopt reproduced the same GAMS writer passed solver (dicopt) unsuitable for model type (mip) error. Re-running with the MINLP role set to Gurobi got past this solver-type failure, then reached the separate LBB time-limit finalization issue tracked in #3941.

For comparison, short GDPlib probes for gdpopt.loa, gdpopt.gloa, and gdpopt.ric on the same model completed without this routing failure; their logs showed GAMS generating MIP models and executing Gurobi through the MIP role.