pyomo-cp

Constraint-programming backends for Pyomo: a discretization transform plus CP solver interfaces, starting with CP-SAT (OR-Tools).

Pyomo can build and solve disjunctive/logical models via pyomo.gdp, but it can only solve them by reformulating to MILP/MINLP (big-M, hull) and calling a MIP solver. pyomo-cp adds the missing path: take the same model and solve it with a constraint-programming solver, where disjunctions map to native reified constraints instead of being reformulated.

Status: alpha. Integer models, pyomo.gdp disjunctions, logical constraints / Boolean variables, and discretization of continuous variables (any step) work end-to-end via SolverFactory('cpsat'). Global constraints are not yet supported.

Scope

pyomo-cp is a backend framework, not a full CP modelling frontend. It translates what Pyomo can already express:

• integer and (once discretized) continuous variables,
• linear constraints and a linear objective,
• logical constraints (BooleanVar / LogicalConstraint),
• pyomo.gdp disjunctions.

Global constraints (alldifferent, no_overlap, element, cumulative, ...) are the heart of CP's modelling power, and Pyomo has no vocabulary for them. Adding one is roadmap, not part of the initial scope.

CP solvers are finite-domain, so continuous variables must be discretized first. That is an explicit step (TransformationFactory('cp.discretize')), never automatic, because the integrality assumption is a modelling decision that changes the problem.

Relationship to pyomo.contrib.cp

Pyomo ships an in-tree constraint-programming module, pyomo.contrib.cp. It is a different tool for a different job, and the two are complementary:

In one line: pyomo.contrib.cp is write an interval/scheduling model and solve it with CP Optimizer; this package is take a disjunctive/integer model you'd otherwise reformulate to MILP and solve it (optionally discretized) with CP-SAT. They overlap only on logical constraints; the paradigm (interval scheduling vs disjunctive/geometric) and the solver (commercial vs open) otherwise differ.

Install

pip install "pyomo-cp[cpsat]"   # includes OR-Tools for the CP-SAT backend

Usage

import pyomo.environ as pyo
from pyomo.gdp import Disjunct, Disjunction
import pyomo_cp  # registers cp.discretize and the cpsat solver

# Two boxes (lengths 2 and 3) that must not overlap on a line; minimize extent.
m = pyo.ConcreteModel()
m.x1 = pyo.Var(bounds=(0, 10))          # continuous positions
m.x2 = pyo.Var(bounds=(0, 10))
m.L = pyo.Var(bounds=(0, 10))
m.e1 = pyo.Constraint(expr=m.L >= m.x1 + 2)
m.e2 = pyo.Constraint(expr=m.L >= m.x2 + 3)
m.d1 = Disjunct(); m.d1.c = pyo.Constraint(expr=m.x1 + 2 <= m.x2)  # box1 left of box2
m.d2 = Disjunct(); m.d2.c = pyo.Constraint(expr=m.x2 + 3 <= m.x1)  # box2 left of box1
m.no = Disjunction(expr=[m.d1, m.d2])
m.obj = pyo.Objective(expr=m.L)

pyo.TransformationFactory("cp.discretize").apply_to(m)      # explicit; unit grid
res = pyo.SolverFactory("cpsat").solve(m, time_limit=10)
print(res.solver.termination_condition, pyo.value(m.obj))  # optimal 5.0

See examples/plant_layout.ipynb for a complete worked example: a 17-block plant-layout GDP model solved to proven optimality by CP-SAT via cp.discretize.

Solver options are passed as friendly aliases (time_limit, workers, seed, gap) or as raw CP-SAT parameter names via options={...}, e.g. solve(m, workers=8, options={"log_search_progress": True}).

Maintainer

Maintained by @devin-griff. Issues and pull requests welcome.

License

BSD 3-Clause License. See LICENSE.