ommx_pyscipopt_adapter

ommx_pyscipopt_adapter#

Submodules#

Exceptions#

OMMXPySCIPOptAdapterError

Common base class for all non-exit exceptions.

Classes#

`OMMXPySCIPOptAdapter`	Solve OMMX problems with SCIP through PySCIPOpt.
`SCIPDiagnosticsAnalyzer`	Pandas-like post-processor for PySCIPOpt diagnostics.
`SCIPProgressSnapshot`	SCIP solve progress observed during a solve.
`SCIPTerminationReport`	SCIP-side termination summary recorded after `model.optimize()`.

Package Contents#

exception OMMXPySCIPOptAdapterError#: Common base class for all non-exit exceptions.

class OMMXPySCIPOptAdapter(ommx_instance: Instance, *, initial_state: ToState | NamedState | None = None)#

Solve OMMX problems with SCIP through PySCIPOpt.

Install the ommx-pyscipopt-adapter package alongside a compatible OMMX SDK to use these examples. INPUT_CLASS defines the exact supported input. Nonconstant linear Indicator bodies and SOS1 constraints use SCIP’s native addConsIndicator and addConsSOS1 operations. A nonconstant equality indicator becomes two inequality indicators. A constant body is simplified: a satisfied body needs no row, while an unsatisfied body forces its indicator to zero. OneHot constraints are lowered to regular equalities by the recommended preparation policy.

solve() prepares an isolated copy, leaving the caller’s Instance unchanged. solve_without_preparation() and the constructor require the input to satisfy INPUT_CLASS already.

Build, solve, and inspect a 0-1 knapsack problem

Select items to maximize their total value while keeping their total weight at most the capacity. Each Binary variable is one when its item is selected. The names and subscripts preserve the connection to the input data:

>>> from ommx import Instance, Solution
>>> from ommx_pyscipopt_adapter import OMMXPySCIPOptAdapter
>>> values = [10, 13, 18, 31, 7, 15]
>>> weights = [11, 25, 20, 35, 10, 33]
>>> capacity = 47
>>> problem = Instance.maximize()
>>> items = [problem.new_binary("item", subscripts=[i]) for i in range(len(values))]
>>> problem.objective = sum(value * item for value, item in zip(values, items))
>>> weight_limit = problem.add_constraint(
...     sum(weight * item for weight, item in zip(weights, items)) <= capacity,
...     "Weight limit",
... )
>>> solution = OMMXPySCIPOptAdapter.solve(problem)
>>> assert solution.feasible
>>> assert solution.optimality == Solution.OPTIMAL
>>> selected = sorted(
...     i for (i,), value in solution.extract_decision_variables("item").items()
...     if value > 0.5
... )
>>> selected
[0, 3]
>>> assert abs(solution.objective - sum(values[i] for i in selected)) < 1e-6
>>> assert abs(solution.objective - 41.0) < 1e-6

ommx.Solution.objective is the returned candidate’s objective value; the value alone does not establish optimality. Here SCIP proves optimality for the problem being evaluated. The decision-variable table provides IDs, names, subscripts, and values for tabular analysis. The regular-constraint table reports the normalized left-hand side: total weight minus capacity.

>>> variables = solution.decision_variables_df()
>>> assert set(variables["name"]) == {"item"}
>>> assert sorted(variables["value"]) == [0.0, 0.0, 0.0, 0.0, 1.0, 1.0]
>>> selection = variables[["subscripts", "value"]].copy()
>>> selection["choice"] = selection["value"].map(
...     lambda value: "Include" if value > 0.5 else "Exclude"
... )
>>> assert selection["choice"].value_counts().to_dict() == {"Exclude": 4, "Include": 2}
>>> constraints = solution.constraints_df(kind=ConstraintKind.Regular)
>>> residual = constraints.loc[weight_limit.constraint_id, "value"]
>>> assert abs(residual - (sum(weights[i] for i in selected) - capacity)) < 1e-6
>>> assert abs(residual - (-1.0)) < 1e-6

The selected items weigh 46, so the inequality has one unit of slack. ommx.Solution documents result extraction and evaluation semantics. The input can also come from a modeler, ommx.Instance.load_mps(), or ommx.artifact.Artifact; model construction is independent of solve.

Native Indicator and SOS1 constraints

This model rewards enabling a feature and selecting options. When enabled is one, the Indicator requires option_a + option_b <= 1. The SOS1 allows at most one of option_b and option_c to be nonzero; it does not require either one to be selected. SCIP receives both structures directly:

>>> from ommx import Instance, Sos1Constraint, ConstraintKind
>>> from ommx_pyscipopt_adapter import OMMXPySCIPOptAdapter
>>> source = Instance.maximize()
>>> enabled = source.new_binary("enabled")
>>> option_a = source.new_binary("option_a")
>>> option_b = source.new_binary("option_b")
>>> option_c = source.new_binary("option_c")
>>> source.objective = 10 * enabled + 6 * option_a + 5 * option_b + 4 * option_c
>>> _ = source.add_indicator_constraint((option_a + option_b <= 1).with_indicator(enabled))
>>> _ = source.add_sos1_constraint(
...     Sos1Constraint(variables=[option_b, option_c], name="exclusive_options")
... )
>>> assert OMMXPySCIPOptAdapter.INPUT_CLASS.contains(source)
>>> solution = OMMXPySCIPOptAdapter.solve(source)
>>> assert solution.feasible
>>> assert abs(solution.objective - 20.0) < 1e-6
>>> assert source.active_special_constraint_kinds == {
...     ConstraintKind.Indicator, ConstraintKind.Sos1,
... }
>>> assert not source.removed_constraints

The selected values are enabled = option_a = option_c = 1 and option_b = 0. The ommx_highs_adapter.OMMXHighsAdapter Reference solves this same model through lowering, demonstrating the distinction between native support and preparation for a target input class.

OneHot preparation without changing the source

Exactly one of the three binary variables must be one. The easy solve path lowers the OneHot constraint on its copy before encoding SCIP input:

>>> from ommx import Instance, OneHotConstraint, ConstraintKind
>>> from ommx_pyscipopt_adapter import OMMXPySCIPOptAdapter
>>> choice_model = Instance.maximize()
>>> choices = [choice_model.new_binary("choice", subscripts=[i]) for i in range(3)]
>>> choice_model.objective = 5 * choices[0] + 10 * choices[1] + 3 * choices[2]
>>> _ = choice_model.add_one_hot_constraint(OneHotConstraint(variables=choices))
>>> assert not OMMXPySCIPOptAdapter.INPUT_CLASS.contains(choice_model)
>>> solution = OMMXPySCIPOptAdapter.solve(choice_model)
>>> assert solution.feasible
>>> assert abs(solution.objective - 10.0) < 1e-6
>>> assert choice_model.active_special_constraint_kinds == {ConstraintKind.OneHot}
>>> assert not choice_model.removed_constraints

SOS1 with continuous variables

Native SOS1 support also applies to continuous variables: at most one may be nonzero, and all may be zero.

>>> from ommx import Instance, Sos1Constraint
>>> from ommx_pyscipopt_adapter import OMMXPySCIPOptAdapter
>>> sparse_model = Instance.maximize()
>>> ys = [sparse_model.new_continuous("y", lower=0, upper=10, subscripts=[i]) for i in range(3)]
>>> sparse_model.objective = sum(ys)
>>> _ = sparse_model.add_sos1_constraint(Sos1Constraint(variables=ys))
>>> solution = OMMXPySCIPOptAdapter.solve(sparse_model)
>>> assert solution.feasible
>>> assert abs(solution.objective - 10.0) < 1e-6

Select a policy and use the strict solve path

Here the application permits only OneHot lowering and keeps SOS1 native. Constructing the policy does not change the model. prepare changes the Instance in place and stops when the target class is reached; an enabled operation need not run if the input already belongs to that class.

>>> from ommx import (
...     Instance, OneHotConstraint, PreparationPolicy, Sos1Constraint,
...     ConstraintKind, SpecialConstraintPreparation,
... )
>>> from ommx_pyscipopt_adapter import OMMXPySCIPOptAdapter
>>> model = Instance.maximize()
>>> primary = [model.new_binary("primary", subscripts=[i]) for i in range(2)]
>>> optional = [model.new_binary("optional", subscripts=[i]) for i in range(2)]
>>> model.objective = 10 * primary[0] + 8 * primary[1] + 4 * optional[0] + 3 * optional[1]
>>> _ = model.add_one_hot_constraint(OneHotConstraint(variables=primary))
>>> _ = model.add_sos1_constraint(Sos1Constraint(variables=optional))
>>> policy = PreparationPolicy(
...     special_constraints=SpecialConstraintPreparation.lower_special_constraints(
...         kinds={ConstraintKind.OneHot}
...     )
... )
>>> model.prepare(OMMXPySCIPOptAdapter.INPUT_CLASS, policy)
>>> assert OMMXPySCIPOptAdapter.check_applicability(model).is_member
>>> assert model.active_special_constraint_kinds == {ConstraintKind.Sos1}
>>> solution = OMMXPySCIPOptAdapter.solve_without_preparation(model)
>>> assert solution.feasible
>>> assert abs(solution.objective - 14.0) < 1e-6
>>> assert len(solution.decision_variables_df()) == 4

The strict call applies no additional preparation. Calling solve instead obtains a fresh recommended policy; it does not reuse the local policy variable. A recommendation is an editable starting point, not a guarantee that every model can reach INPUT_CLASS.

classmethod check_applicability(ommx_instance: Instance) → InstanceClassMembershipReport#: Check INPUT_CLASS membership without mutation.

decode(data: Model) → Solution#

Convert optimized pyscipopt.Model and ommx.Instance to ommx.Solution.

This method is intended to be used if the model has been acquired with solver_input for further adjustment of the solver parameters, and separately optimizing the model.

Note that alterations to the model may make the decoding process incompatible – decoding will only work if the model still describes effectively the same problem as the OMMX instance used to create the adapter.

Backend optimality is mapped through the instance’s output-objective semantics. It remains unspecified when active-formulation optimality does not transport to that objective.

Examples#

>>> from ommx_pyscipopt_adapter import OMMXPySCIPOptAdapter
>>> from ommx import Instance, DecisionVariable

>>> p = [10, 13, 18, 32, 7, 15]
>>> w = [11, 15, 20, 35, 10, 33]
>>> x = [DecisionVariable.binary(i) for i in range(6)]
>>> instance = Instance.from_components(
...     decision_variables=x,
...     objective=sum(p[i] * x[i] for i in range(6)),
...     constraints={0: sum(w[i] * x[i] for i in range(6)) <= 47},
...     sense=Sense.Maximize,
... )

>>> adapter = OMMXPySCIPOptAdapter(instance)
>>> model = adapter.solver_input
>>> # ... some modification of model's parameters
>>> model.optimize()

>>> solution = adapter.decode(model)
>>> solution.objective
42.0

decode_to_state(data: Model) → State#

Create an ommx.State from an optimized PySCIPOpt Model.

Examples#

The following example shows how to solve an unconstrained linear optimization problem with `x1` as the objective function.

>>> from ommx_pyscipopt_adapter import OMMXPySCIPOptAdapter
>>> from ommx import Instance, DecisionVariable

>>> x1 = DecisionVariable.integer(1, lower=0, upper=5)
>>> ommx_instance = Instance.from_components(
...     decision_variables=[x1],
...     objective=x1,
...     constraints={},
...     sense=Sense.Minimize,
... )
>>> adapter = OMMXPySCIPOptAdapter(ommx_instance)
>>> model = adapter.solver_input
>>> model.optimize()

>>> ommx_state = adapter.decode_to_state(model)
>>> ommx_state.entries
{1: 0.0}

classmethod recommended_preparation_policy() → PreparationPolicy#

Recommend lowering OneHot constraints before using PySCIPOpt.

PySCIPOpt accepts Indicator and SOS1 constraints directly, so this recommendation preserves those families and lowers only OneHot constraints. The easy API applies a fresh policy to an isolated copy; callers may instead edit and apply one before invoking solve_without_preparation().

classmethod require_applicable(ommx_instance: Instance) → InstanceClassMembershipReport#: Return the membership report or raise AdapterNotApplicableError.

classmethod solve(ommx_instance: Instance, *, initial_state: ToState | NamedState | None = None, diagnostics: DiagnosticsSink | None = None) → Solution#

Solve the given ommx.Instance using PySCIPopt, returning an ommx.Solution.

The input instance is not modified. An isolated copy is prepared with the recommended PySCIPOpt policy before preparation-free execution.

Parameters:

ommx_instance – The ommx.Instance to prepare and solve.
initial_state – Optional initial solution as a State, ID-keyed values, or values grouped by name and tuple subscripts.

Examples#

The OMMXPySCIPOptAdapter class documentation contains a complete knapsack example from model construction through named-variable extraction, tabular analysis, objective value, and constraint slack.

Start from a candidate

Evaluate the proposed assignment against the same instance you intend to solve. SCIP uses it as a starting point and can return a better solution. In this example the starting objective is 1 and the optimum is 5.

>>> from ommx import Instance
>>> from ommx_pyscipopt_adapter import OMMXPySCIPOptAdapter
>>> problem = Instance.maximize()
>>> x = problem.new_integer("x", lower=0, upper=5)
>>> y = problem.new_integer("y", lower=0, upper=5)
>>> problem.objective = x - y
>>> _ = problem.add_constraint(x + y <= 5)
>>> proposed = {x.id: 3, y.id: 2}
>>> initial = problem.evaluate(proposed)
>>> assert initial.feasible and initial.objective == 1
>>> result = OMMXPySCIPOptAdapter.solve(problem, initial_state=proposed)
>>> assert result.feasible and result.objective == 5

To reuse a plan across models whose IDs differ, supply names and tuple subscripts. Use an empty tuple for a scalar and (i,) for a single index. SCIP accepts the named mapping directly. To check it first, resolve the names against the target instance:

>>> named = {"x": {(): 3.0}, "y": {(): 2.0}}
>>> state = problem.state_from_named_values(named)
>>> assert problem.evaluate(state).feasible
>>> assert OMMXPySCIPOptAdapter.solve(problem, initial_state=named).objective == 5

You can also pass the resolved State or a previous solution’s state when the IDs match. ommx.Instance.state_from_named_values() describes unknown or ambiguous names. SCIP initializes omitted solver variables to zero when building the proposed solution. Evaluating a partial assignment with OMMX may require additional values. A starting point neither fixes the final assignment nor guarantees a shorter solve.

Infeasible Problem

>>> from ommx import Instance, DecisionVariable, Sense
>>> from ommx_pyscipopt_adapter import OMMXPySCIPOptAdapter

>>> x = DecisionVariable.integer(0, upper=3, lower=0)
>>> instance = Instance.from_components(
...     decision_variables=[x],
...     objective=x,
...     constraints={0: x >= 4},
...     sense=Sense.Maximize,
... )

>>> OMMXPySCIPOptAdapter.solve(instance)
Traceback (most recent call last):
    ...
ommx.InfeasibleDetected: Model was infeasible

Unbounded Problem

>>> from ommx import Instance, DecisionVariable, Sense
>>> from ommx_pyscipopt_adapter import OMMXPySCIPOptAdapter

>>> x = DecisionVariable.integer(0, lower=0)
>>> instance = Instance.from_components(
...     decision_variables=[x],
...     objective=x,
...     constraints={},
...     sense=Sense.Maximize,
... )

>>> OMMXPySCIPOptAdapter.solve(instance)
Traceback (most recent call last):
    ...
ommx.adapter.UnboundedDetected: Model was unbounded

classmethod solve_without_preparation(ommx_instance: Instance, *, initial_state: ToState | NamedState | None = None, diagnostics: DiagnosticsSink | None = None) → Solution#: Solve an exact PySCIPOpt Adapter input without preparing it.

INPUT_CLASS: ClassVar[InstanceClass]#: Required condition for an exact Adapter input.

property solver_input: Model#: The PySCIPOpt model generated from this OMMX instance

class SCIPDiagnosticsAnalyzer(diagnostics: Iterable[Any])#

Pandas-like post-processor for PySCIPOpt diagnostics.

The analyzer accepts either typed diagnostics collected by ommx.adapter.DiagnosticCollector or dictionaries loaded from ommx.experiment.Solve.diagnostics.

progress_history_records returns list[dict[str, object]] and does not require pandas. When diagnostics include a terminal SCIP report, the progress history includes a final TERMINATION row derived from that report. progress_history_df returns a pandas DataFrame indexed by solving_time_sec and imports pandas lazily. Time-series properties such as dual_bound return pandas Series with the same index. termination_result returns the terminal SCIP report as a dictionary.

property dual_bound: Any#: Return dual bounds indexed by solving time.

property event: Any#: Return progress event names indexed by solving time.

property gap: Any#: Return relative gaps indexed by solving time.

property incumbent_objective: Any#: Return incumbent objectives indexed by solving time.

property lp_iteration_count: Any#: Return LP iteration counts indexed by solving time.

property node_count: Any#: Return processed node counts indexed by solving time.

property primal_bound: Any#: Return primal bounds indexed by solving time.

property progress_history_df: Any#: Return progress history as a pandas DataFrame indexed by time.

property progress_history_records: list[dict[str, object]]#: Return SCIP progress history, with TERMINATION when present.

property progress_snapshots: tuple[SCIPProgressSnapshot, ...]#: Typed progress snapshots in the order they were recorded.

property solution_count: Any#: Return stored solution counts indexed by solving time.

property termination_result: dict[str, object] | None#: Return the terminal SCIP report as one dictionary, if present.

property total_node_count: Any#: Return total processed node counts indexed by solving time.

class SCIPProgressSnapshot#

SCIP solve progress observed during a solve.

The PySCIPOpt adapter records this snapshot for each tracked SCIP event and once more from the termination report after model.optimize() finishes. It currently listens for BESTSOLFOUND and DUALBOUNDIMPROVED. Event snapshots are the model state visible from that callback. SCIP may call a BESTSOLFOUND callback before every aggregate model statistic has been updated, so use the TERMINATION snapshot or SCIPTerminationReport for terminal values.

classmethod from_event(model: Model, event: Event) → SCIPProgressSnapshot#

classmethod from_termination_report(report: SCIPTerminationReport) → SCIPProgressSnapshot#

dual_bound: float#: SCIP dual bound reported at the snapshot.

event: str#

Progress marker.

Callback snapshots use SCIP event names such as "BESTSOLFOUND" and "DUALBOUNDIMPROVED". The terminal snapshot uses the synthetic "TERMINATION" marker.

gap: float#: SCIP relative gap reported at the snapshot.

incumbent_objective: float | None#

Objective value of SCIP’s current best solution.

This is None when PySCIPOpt cannot read an incumbent objective at that snapshot.

lp_iteration_count: int#: LP iterations at the snapshot.

node_count: int#: Processed branch-and-bound nodes at the snapshot.

primal_bound: float#: SCIP primal bound reported at the snapshot.

solution_count: int#: Number of solutions stored by SCIP at the snapshot.

solving_time_sec: float#: SCIP solving time when the snapshot was recorded.

total_node_count: int#: Total processed nodes including restarts at the snapshot.

class SCIPTerminationReport#

SCIP-side termination summary recorded after model.optimize().

The PySCIPOpt adapter records this report before decoding the optimized SCIP model back into an OMMX solution. It is therefore available even when decoding raises an adapter exception such as infeasible or unbounded detection.

classmethod from_model(model: Model) → SCIPTerminationReport#

applied_cut_count: int#: Number of cuts applied by SCIP.

best_solution_count: int#: Number of new incumbent solutions SCIP found.

cut_count: int#: Number of cuts available in SCIP’s cut pool.

dual_bound: float#: SCIP dual bound at termination.

gap: float#: SCIP relative gap reported by getGap().

lp_iteration_count: int#: Total LP iterations.

lp_solve_count: int#: Number of solved LPs.

max_depth: int#

Maximum branch-and-bound depth.

SCIP may report -1 when no branching occurred.

node_count: int#: Number of branch-and-bound nodes processed by SCIP.

objective_value: float | None#: Incumbent objective value, or None when SCIP has no solution.

presolving_time_sec: float#: SCIP presolving time in seconds.

primal_bound: float#: SCIP primal bound at termination.

primal_dual_integral: float#: SCIP primal-dual integral at termination.

pyscipopt_version: str | None#: PySCIPOpt package version, if available.

reading_time_sec: float#: SCIP reading time in seconds.

scip_version: str#: SCIP version used through PySCIPOpt.

solution_count: int#: Number of solutions stored by SCIP at termination.

solution_found_count: int#: Number of solutions SCIP found during the solve.

solving_time_sec: float#: SCIP solving time in seconds.

status: str#: SCIP termination status, such as "optimal", "infeasible", or "unbounded".

total_node_count: int#: Total processed nodes including restarts.