Specialised theory solver for neural networks using SoPlex. More...

#include <NNSoplexTheorySolver.h>

Inheritance diagram for dlinear::NNSoplexTheorySolver:

Classes
struct	PlConstraint

Public Member Functions
void	AddLiteral (const Variable &formula_var, const Formula &formula) override
	Add a Literal to the theory solver.

Explanations	EnableLiteral (const Literal &lit) override
	Activate the literal that had previously been added to the theory solver.

SatResult	CheckSatCore (mpq_class *actual_precision, std::set< LiteralSet > &explanations) override
	Check the satisfiability of the theory.

void	Reset () override
	Reset the linear problem.

Public Member Functions inherited from dlinear::SoplexTheorySolver
void	AddVariable (const Variable &var) override
	Add a variable (column) to the theory solver.

void	AddLiterals () override
	Add a vector of literals to the theory solver.

void	Consolidate (const Box &box) override
	Consolidate the solver.

void	Reset () override
	Reset the linear problem.

Public Member Functions inherited from dlinear::TheorySolver
	TheorySolver (const PredicateAbstractor &predicate_abstractor, const std::string &class_name="TheorySolver")
	Construct a new Theory Solver object.

virtual Explanations	PreprocessFixedLiterals (const LiteralSet &fixed_literals)
	Add the fixed literals to the theory solver.

Explanations	EnableLiterals (std::span< const Literal > theory_literals)
	Activate the literals that have previously been added to the theory solver.

const Box &	model () const
	Get read-only access to the model that satisfies all the constraints of the TheorySolver.

const Config &	config () const
	Get read-only access to the configuration of the TheorySolver.

const PredicateAbstractor &	predicate_abstractor () const
	Get read-only access to the predicate abstractor of the TheorySolver.

const std::map< Variable::Id, int > &	var_to_theory_col () const
	Get read-only access to the map of variables to the theory columns of the TheorySolver.

const std::vector< Variable > &	theory_col_to_var () const
	Get read-only access to the map of theory columns to the variables of the TheorySolver.

const std::map< Variable::Id, int > &	lit_to_theory_row () const
	Get read-only access to the map of literals to the theory rows of the TheorySolver.

const std::vector< Literal > &	theory_row_to_lit () const
	Get read-only access to the map of theory rows to the literals of the TheorySolver.

const BoundVectorMap &	theory_bounds () const
	Get read-only access to the bounds of each theory variable of the TheorySolver.

const BoundVectorMap &	fixed_theory_bounds () const
	Get read-only access to the fixed bounds of each theory variable of the TheorySolver.

const BoundPreprocessor &	preprocessor () const
	Get read-only access to the bound preprocessor of the TheorySolver.

BoundPreprocessor &	m_preprocessor ()
	Get read-write access to the bound preprocessor of the TheorySolver.

const BoundPreprocessor &	fixed_preprocessor () const
	Get read-only access to the fixed bound preprocessor of the TheorySolver.

BoundPreprocessor &	m_fixed_preprocessor ()
	Get read-write access to the fixed bound preprocessor of the TheorySolver.

std::size_t	theory_row_count () const
	Get read-only access to the number of rows of the TheorySolver.

std::size_t	theory_col_count () const
	Get read-only access to the number of columns of the TheorySolver.

std::set< Literal >	enabled_literals () const
	Get read-only access to the vector of enabled literals of the TheorySolver.

SatResult	CheckSat (const Box &box, mpq_class *actual_precision, std::set< LiteralSet > &explanations)
	Check the satisfiability of the theory.

const IterationStats &	stats () const
	Get the statistics of the theory solver.

Private Member Functions
void	EnableSpxRow (int spx_row, bool truth) override
	Enable the `spx_row` row for the LP solver.

void	Consolidate (const Box &box) override
	Consolidate the solver.

void	SoiToObjFunction ()
	Convert the soi_ expression to an objective function inside soplex.

bool	InvertGreatestViolation ()
	Invert the greatest violation found in the objective function.

Additional Inherited Members
Public Types inherited from dlinear::TheorySolver
using	Bound = std::tuple<const Variable &, LpColBound, const mpq_class &>
	Bound on the variable.

using	Violation = BoundIterator
	Bound iterator over some violated bounds.

using	Explanations = std::set<LiteralSet>
	Set of explanations of the conflict.

Protected Types inherited from dlinear::TheorySolver
enum class	BoundViolationType { NO_BOUND_VIOLATION , LOWER_BOUND_VIOLATION , UPPER_BOUND_VIOLATION }
	Enum used to describe how the bounds on a variable participate in the infeasibility result of an LP problem. More...

Protected Member Functions inherited from dlinear::SoplexTheorySolver
void	UpdateModelBounds () override
	Update each variable in the model with the bounds passed to the theory solver.

void	UpdateModelSolution () override
	Update the model with the solution obtained from the LP solver.

void	UpdateExplanation (LiteralSet &explanation) override
	Use the result from the theory solver to update the explanation with the conflict that has been detected.

virtual void	EnableSpxVarBound ()
	Set the bounds of the variables in the LP solver.

void	EnableSpxRow (int spx_row)
	Enable the `spx_row` row for the LP solver.

void	DisableSpxRows ()
	Disable all disabled spx rows from the LP solver.

soplex::DSVectorRational	ParseRowCoeff (const Formula &formula)
	Parse a `formula` and return the vector of coefficients to apply to the decisional variables.

void	SetSPXVarCoeff (soplex::DSVectorRational &coeffs, const Variable &var, const mpq_class &value) const
	Set the coefficients to apply to `var` on a specific row.

void	GetSpxInfeasibilityRay (soplex::VectorRational &farkas_ray)
	Get the infeasibility ray of the LP problem.

void	GetSpxInfeasibilityRay (soplex::VectorRational &farkas_ray, std::vector< BoundViolationType > &bounds_ray)
	Get the infeasibility rays of the LP problem.

std::vector< std::pair< int, soplex::Rational > >	GetActiveRows ()
	Get all active rows in the current solution.

std::vector< std::pair< int, soplex::Rational > >	GetActiveRows (const std::vector< int > &spx_rows)
	Get active rows in the current solution among the rows in `spx_rows`.

std::optional< soplex::Rational >	IsRowActive (int spx_row)
	Check if the `spx_row` is active.

bool	IsRowActive (int spx_row, const soplex::Rational &value)
	Check if the `spx_row` is active with value `value`.

Protected Member Functions inherited from dlinear::TheorySolver
void	UpdateExplanations (Explanations &explanations)
	Use the result from the theory solver to update the explanations with the conflict that has been detected.

Protected Attributes inherited from dlinear::SoplexTheorySolver
soplex::SoPlex	spx_
	SoPlex exact LP solver.

soplex::LPColSetRational	spx_cols_
	Columns of the LP problem.

soplex::LPRowSetRational	spx_rows_
	Rows of the LP problem.

std::vector< mpq_class >	spx_rhs_
	Right-hand side of the rows.

std::vector< LpRowSense >	spx_sense_
	Sense of the rows.

Protected Attributes inherited from dlinear::TheorySolver
const Config &	config_
	Configuration of the theory solver.

bool	is_consolidated_
	Whether the solver has been consolidated.

const PredicateAbstractor &	predicate_abstractor_
	Predicate abstractor used to create the theory solver.

std::map< Variable::Id, int >	var_to_theory_col_
	Variable ⇔ theory column.

std::vector< Variable >	theory_col_to_var_
	Theory column ⇔ Variable.

std::map< Variable::Id, int >	lit_to_theory_row_
	Literal ⇔ theory row.

std::vector< Literal >	theory_row_to_lit_
	Theory row ⇔ Literal The row is the constraint used by the theory solver.

std::vector< bool >	theory_rows_state_
	Whether each theory row is enabled or not.

BoundPreprocessor	fixed_preprocessor_
	Preprocessor for the bounds. Only computed on fixed literal theories.

BoundPreprocessor	preprocessor_
	Preprocessor for the bounds.

Box	model_
	Model produced by the theory solver.

IterationStats	stats_
	Statistics of the theory solver.

Static Protected Attributes inherited from dlinear::SoplexTheorySolver
static mpq_class	infinity_ {soplex::infinity}
	Positive infinity.

static mpq_class	ninfinity_ {-soplex::infinity}
	Negative infinity.

Detailed Description

Specialised theory solver for neural networks using SoPlex.

The solver will use special care when dealing with piecewise linear constraints. Instead of encoding them as a set of linear constraints, they will create a new objective function representing the sum of infeasibilities to minimise. If the problem is feasible, but the objective value is not 0, the solver will try to invert the greatest violation to attempt making the problem feasible.

Definition at line 29 of file NNSoplexTheorySolver.h.

Member Function Documentation

◆ AddLiteral()

void dlinear::NNSoplexTheorySolver::AddLiteral	(	const Variable &	formula_var,
		const Formula &	formula )

overridevirtual

Add a Literal to the theory solver.

A Literal is formed by a variable that corresponds to a theory formula inside the PredicateAbstractor and the formula itself

Parameters

formula_var	boolean variable that corresponds to the theory formula
formula	symbolic formula that represents the theory formula

Implements dlinear::TheorySolver.

Definition at line 35 of file NNSoplexTheorySolver.cpp.

◆ CheckSatCore()

SatResult dlinear::NNSoplexTheorySolver::CheckSatCore	(	mpq_class *	actual_precision,
		std::set< LiteralSet > &	explanations )

overridevirtual

Check the satisfiability of the theory.

Called by CheckSat. Run the internal LP solver to check whether the underlying linear programming problem is feasible. If it is, SAT will be returned, along with the actual precision required to obtain that result. Otherwise, UNSAT will be returned, along with an explanation of the conflict. In that case, the precision will remain the same as the one passed as input.

Parameters

[in,out]	actual_precision	desired precision. It will be updated with the actual precision if SAT is returned
[out]	explanations	set of sets of literals that explain the conflict if UNSAT is returned

Returns: SAT if the problem is feasible, along with the actual precision required to obtain that result and the model; UNSAT if the problem is infeasible, along with an explanation of the conflict

Implements dlinear::TheorySolver.

Definition at line 114 of file NNSoplexTheorySolver.cpp.

◆ Consolidate()

void dlinear::NNSoplexTheorySolver::Consolidate ( const Box & box )

overrideprivatevirtual

Consolidate the solver.

This method must be called after all the literals have been added to the solver and before calling any other method. Once the solver has been consolidated, no more literals can be added to it. A previously added literal can be enabled using the EnableLiteral method and disabled with Reset.

Note: A solver can be consolidated only once. If you need to change the variables or constraints, you must create a new theory solver.

Parameters

box	box with the bounds for the variables

Reimplemented from dlinear::TheorySolver.

Definition at line 285 of file NNSoplexTheorySolver.cpp.

◆ EnableLiteral()

TheorySolver::Explanations dlinear::NNSoplexTheorySolver::EnableLiteral ( const Literal & lit )

overridevirtual

Activate the literal that had previously been added to the theory solver.

If the literal is a row inside the TheorySolver, then the the corresponding constraint is activated by giving it the proper values.

If the literal is a simple bound on the variable (column), some additional checks are made. Theory solvers struggle to handle problems with inverted bounds. It is convenient to have a method that checks them beforehand. If at least a bound is inverted, the problem is UNSAT, since it is impossible to satisfy the constraint. An explanation is produced and returned to the SAT solver.

Parameters

lit	literal to activate

Returns: an explanation with the literals that correspond to the conflicting bounds; an empty optional if no conflicts are detected at this step

Implements dlinear::TheorySolver.

Definition at line 67 of file NNSoplexTheorySolver.cpp.

◆ EnableSpxRow()

void dlinear::NNSoplexTheorySolver::EnableSpxRow	(	int	spx_row,
		bool	truth )

overrideprivatevirtual

Enable the spx_row row for the LP solver.

Precondition: The row's coefficients must have been set correctly before calling this function

Parameters

spx_row	index of the row to enable
truth	truth value of the row

Implements dlinear::SoplexTheorySolver.

Definition at line 291 of file NNSoplexTheorySolver.cpp.

◆ Reset()

void dlinear::NNSoplexTheorySolver::Reset ( )

overridevirtual

Reset the linear problem.

All constraints' state will be set to disabled and the bounds for each variable will be cleared.

Note: The variables and constraints themselves will not be modified. If you need to change them, you must create a new theory solver.

Reimplemented from dlinear::TheorySolver.

Definition at line 279 of file NNSoplexTheorySolver.cpp.

The documentation for this class was generated from the following files:

dlinear/solver/NNSoplexTheorySolver.h
dlinear/solver/NNSoplexTheorySolver.cpp

Classes

Public Member Functions

Private Member Functions

Additional Inherited Members

Detailed Description

Member Function Documentation

◆ AddLiteral()

◆ CheckSatCore()

◆ Consolidate()

◆ EnableLiteral()

◆ EnableSpxRow()

◆ Reset()