Define your own constraints

In this chapter, you will learn how to define your own constraints propagation mechanism and how to make them efficient and incremental.

Constraint’s semantic

We will start by specifying the semantic of the constraint. It is defined by:

  • A constant \(M\).
  • Two variables \(X\) and \(Y\) with \(X\) the result of the operation \(Y \bmod M\).

let’s see how to implement this constraint it the Artelys Kalis library framework:

Implementation of the constraint

The KUserConstraint is the base class of all the user defined constraints in Artelys Kalis. When you want to define your own constraint, you have to create a new class that will inherit from the KUserConstraint class. The KUserConstraint class contains six kind of interfaces:

class KUserConstraint: public KConstraint
{
    public:
        KIntVarArray _vars;

        /// Constructor for unary constraints
        KUserConstraint(KIntVar &v1);

        /// Constructor for binary constraints
        KUserConstraint(KIntVar &v1,KIntVar &v2);

        /// Constructor for n-ary constraints
        KUserConstraint(KIntVarArray &vars);

        /// Copy constructor
        KUserConstraint(const KUserConstraint & toCopy);

        // destructor
        virtual ~KUserConstraint();

        /// virtual method called when the domain of some or several variables has changed
        virtual void propagate(void) = 0;

        /// virtual method called upon initialization of the constraint
        virtual void awake(void) = 0;

        /// virtual method called when the lower bound of var has been raised
        virtual void awakeOnInf(KIntVar & var);

        /** virtual method called when the upper bound of var has been lowered
            @param var the variable with modified domain
        */
        virtual void awakeOnSup(KIntVar & var);

        /** virtual method called when the variable var has been instantiated
            @param var the variable with modified domain
        */
        virtual void awakeOnInst(KIntVar & var);

        /** virtual method called when the value removedValue has been removed from the domain of var
            @param var the variable with modified domain
            @param removedValue the value that has been removed from the domain of var
        */
        virtual void awakeOnRem(KIntVar & var,int removedValue);

        /** virtual method called when the domain of variable var has changed
            @param var the variable with modified domain
        */
        virtual void awakeOnVar(KIntVar & var);

        /** virtual method for use within boolean connectors
            @return CTRUE whenever the constraint is definitively satisfied
            @return CFALSE whenever the constraint is definitively violated
            @return CUNKNOWN otherwhise
        */
        virtual int askIfEntailed(void);

        virtual void print(void)=0;
};
  • Several constructors that need to be called by the subclasses constructors;
  • An awake and a propagate method that need to be implemented;
  • Severals awake() methods that may be implemented for an incremental propagation scheme (awakeOnInf(), awakeOnSup(), awakeOnInst(), awakeOnRem(), awakOnVar())
  • A print() method that needs to be implemented;
  • An askIfEntailed() method;
  • A getCopy(KProblem*) method.

The purpose of these methods will be explained later. Let’s first start by creating a class named ModuloConstraintXYM that inherit from the KUserConstraint interface. We will implement one constructor, the awake and propagate methods, the awakeOnInst() method, the print and the askIfEntailed() methods:

//*********************************************************************************
// Definition of the constraint X == Y mod M.
//*********************************************************************************
class ModuloConstraintXYM: public KUserConstraint
{
    private:
        int _M;
        KIntVar * _VX;
        KIntVar * _VY;

    public:
        ModuloConstraintXYM(KIntVar &X, KIntVar &Y, const int modulo);
        virtual void awake(void);
        virtual void propagate(void);
        virtual void awakeOnInst(KIntVar & var);
        virtual void print();
        virtual int askIfEntailed(void);
        virtual KConstraint* getCopyPtr(const KProblem& problem) const;
};
# **************************************************************************************
#    Definition of the constraint X == Y mod M.
# **************************************************************************************
class ModuloConstraintXYC (KUserConstraint):

    def __init__(self, X, Y, modulo):
        # Constructor of the constraint
        KUserConstraint.__init__(self, X, Y)
        self._M = modulo
        self._VX = X
        self._VY = Y

    def awake(self):
        # Initialisation and first propagation of the constraint
        # Trivial deduction: 0 <= Y mod C < C
        self._VX.setInf(0)
        self._VX.setSup(self._M - 1)
        # To avoid issues with the '%' function: 0 <= Y
        self._VY.setInf(0)
        # Now do some AC3 like propagation (arc consistency)
        self.propagate()

    def propagate(self):
        # Propagating domain updates since last propagation.
        # Now do some AC3 like propagation
        # Propagating from X to Y
        for  vy in range(int(self._VY.getInf()), int(self._VY.getSup())+1):
            support = False
            for vx in range(int(self._VX.getInf()), int(self._VX.getSup())+1):
                if vx == vy % self._M:
                    support = True
                    break
            if not support:
                self._VY.remVal(vy)
        # Propagating from Y to X
        for vx in range(int(self._VX.getInf()), int(self._VX.getSup())+1):
            support = False
            for vy in range(int(self._VY.getInf()), int(self._VY.getSup())+1):
                if vx == vy % self._M:
                    support = True
                    break
            if not support:
                self._VX.remVal(vx)

    def awakeOnInst(self, var):
        if var.isEqualTo(self._VX):
            # X has been instantiated , removing all values in domain of Y inconsistent with the constraint
            for v in range(int(self._VY.getInf()), int(self._VY.getSup())+1):
                if int(self._VX.getValue()) != v % self._M:
                    self._VY.remVal(v)
        elif var.isEqualTo(self._VY):
            # Y has been instantiated , removing all values in domain of X inconsistent with the constraint
            for v in range(int(self._VX.getInf()), int(self._VX.getSup())+1):
                if v != int(self._VY.getValue()) % self._M:
                    self._VX.remVal(v)

    def askIfEntailed(self):
        #Determine whether the constraint is satisfied or not
        if self._VX.getIsInstantiated() and self._VY.getIsInstantiated():
            # Here X and Y have been instantiated so we now if the constraint is satisfied or violated
            if self._VX.getValue() == self._VY.getValue() % self._M:
                # the constraint hold since X == Y mod C
                return KUserConstraint.CTRUE
            else:
                # the constraint does not hold since X != Y mod C
                return KUserConstraint.CFALSE
        else:
            # X or Y is not yet instantiated so we don't know yet if the constraint holds
            return KUserConstraint.CUNKNOWN

    def display(self):
        # For pretty printing of the constraint, sys.stdout is use here to avoid newline insertion
        sys.stdout.write("%s == %s mod %i"%(self._VX.getName(), self._VY.getName(), self._M))

    def getCopyPtr(self, *args):
        # Obtain a copy of the constraint
        return ModuloConstraintXYC(self._VX, self._VY, self._M)

    def getInstanceCopyPtr(self, *args):
        # Obtain an instance copy of the constraint
        problem = args[0]
        return ModuloConstraintXYC(problem.getInstanceOf(self._VX), problem.getInstanceOf(self._VY), self._M)
public static class ModuloConstraintXYC extends KUserConstraint {
    private KIntVar x, y;
    private int mod;

    //*********************************************************************************
    // Definition of the constraint X == Y mod M.
    //*********************************************************************************
    public ModuloConstraintXYC(KIntVar x, KIntVar y, int mod) {
        super(x, y);
        this.x = x;
        this.y = y;
        this.mod = mod;
    }

    public void awake() {
        x.setInf(0);
        x.setSup(mod - 1);
        y.setInf(0);
        propagate();
    }

    public void propagate() {
        int vx, vy;
        for (vy = (int) (y.getInf()); vy <= y.getSup(); vy++) {
            boolean support = false;
            for (vx = (int) (x.getInf()); vx <= x.getSup(); vx++) {
                if (vx == vy % mod) {
                    support = true;
                    break;
                }
            }
            if (!support) {
                y.remVal(vy);
            }
        }
        for (vx = (int) (x.getInf()); vx <= x.getSup(); vx++) {
            boolean support = false;
            for (vy = (int) (y.getInf()); vy <= y.getSup(); vy++) {
                if (vx == vy % mod) {
                    support = true;
                    break;
                }
            }
            if (!support) {
                x.remVal(vx);
            }
        }
    }

    public void awakeOnInst(KIntVar var) {
        if (var.isEqualTo(x)) {
            for (int v = (int) (y.getInf()); v <= y.getSup(); v++) {
                if ((int) (x.getValue()) != v % mod) {
                    y.remVal(v);
                }
            }
        } else if (var.isEqualTo(y)) {
            for (int v = (int) (y.getInf()); v <= y.getSup(); v++) {
                if (v != (int) (y.getValue()) % mod) {
                    x.remVal(v);
                }
            }
        }
    }

    public int askIfEntailed() {
        if (x.getIsInstantiated() && y.getIsInstantiated()) {
            if (x.getValue() == y.getValue() % mod) {
                // the constraint hold since X == Y mod C
                return askRet.CTRUE;
            } else {
                // the constraint does not hold since X != Y mod C
                return askRet.CFALSE;
            }
        } else {
            return askRet.CUNKNOWN;
        }

    }

    public void print() {
        System.out.print(String.format("%s == %s mod %d", x.getName(),
                y.getName(), mod));
    }

    public KConstraint getCopyPtr() {
        return new ModuloConstraintXYC(x, y, mod);
    }

    public KConstraint getInstanceCopyPtr(KProblem problem) {
        return new ModuloConstraintXYC(problem.getInstanceOf(x),
                problem.getInstanceOf(y), mod);
    }

}

Construction of the constraint

To create the constraint, we need to define a constructor for the class that will initialize the variables and the data linked to the constraint. Note that we need to explicitly make a call to the constructor for binary constraints of the super-class KUserConstraint.

ModuloConstraintXYC::ModuloConstraintXYC(KIntVar &X, KIntVar &Y, const int modulo): KUserConstraint(X,Y)
{
    // note that you must call the superclass constructor
    _M = modulo;
    _VX = &X;
    _VY = &Y;
}
class ModuloConstraintXYC (KUserConstraint):
    # Constructor of the constraint
    def __init__(self, X, Y, modulo):
        KUserConstraint.__init__(self, X, Y)  # note that you must call the superclass constructor
        self._M = modulo
        self._VX = X
        self._VY = Y
public static class ModuloConstraintXYC extends KUserConstraint {
    public ModuloConstraintXYC(KIntVar x, KIntVar y, int mod) {
        super(x, y);
        this.x = x;
        this.y = y;
        this.mod = mod;
    }
}

Now that we have implemented a constructor for the constraint, we will look in detail the purpose and the implementation of the awake and propagate methods.

Constraint’s events

Two kind of events can be thrown to a constraint: events linked to the constraint itself and events linked to a specific variable of the constraint. We will describe here the first kind of event.

Events linked to the constraint can be of two kinds:

  • Awake events which purpose is to initialize the constraint and do an initial propagation.
  • Propagate events that occur when some unknown changes have been applied to the domain of the variables involved in the constraint.

The awake method

Upon initialization of the constraint, it is possible to deduce that the \(X\) variable representing the result of the operation \(Y \bmod C\) is bounded by 0 and M-1:

void ModuloConstraintXYC::awake()
{
    // Trivial deduction: 0 <= Y mod M < M
    _VX->setInf(0);
    _VX->setSup(_M-1);

    // Now do some AC3 like propagation (arc consistency)
    propagate();
}
def awake(self):
    # Initialisation and first propagation of the constraint
    # Trivial deduction: 0 <= Y mod C < C
    self._VX.setInf(0)
    self._VX.setSup(self._M - 1)
    # To avoid issues with the '%' function: 0 <= Y
    self._VY.setInf(0)
    # Now do some AC3 like propagation (arc consistency)
    self.propagate()
public void awake() {
    x.setInf(0);
    x.setSup(mod - 1);
    y.setInf(0);
    propagate();
}

Note that this method can throw a contradiction since it updates the domains of the variables that may be empty after the deductions.

propagate method

Now we have to implement the propagate method that will be triggered by any modifications of the domains of the variables (This execution can be forced by calling the constAwake method of KUserConstraint). In order to find impossible values we will use an AC3 like algorithm that try to find a supporting value (in the domain of the other variable) for each values in the domain of one variable. For example if \(X\) can take value 1,2 and 3 and \(Y\) can take value in 1,2,4 and the \(M\) constant equal 4, the value ‘3’ of \(X\) has no supporting value in the domain of \(Y\) since neither 1 mod 4, 2 mod 4 and 4 mod 4 equal 3. Reciprocally, the value 4 in the domain of \(Y\) has no supporting value in the domain of \(X\) since 0 do not belong to the domain of \(X\). Hence we can deduce that \(X \neq 3\) and \(Y \neq 4\). The following algorithm implement an AC3 like propagation for the constraint \(X = Y \bmod M\):

void ModuloConstraintXYC::propagate()
{
    int vx, vy;

    // Now do some AC3 like propagation
    // Propagating from X to Y
    for ( vy = _VY->getInf();vy<=_VY->getSup();vy ++)
    {
        bool support = false;
        for ( vx = _VX->getInf();vx<=_VX->getSup();vx ++)
        {
            if ( vx == vy % _M)
            {
                support = true;
                break;
            }
        }

        if ( !support)
        {
            _VY->remVal(vy);
        }
    }

    // Propagating from Y to X
    for ( vx = _VX->getInf();vx<=_VX->getSup();vx ++)
    {
        bool support = false;
        for ( vy = _VY->getInf();vy<=_VY->getSup();vy ++)
        {
            if ( vx == vy % _M)
            {
                support = true;
                break;
            }
        }

        if ( !support)
        {
            _VX->remVal(vx);
        }
    }
}
def propagate(self):
    # Propagating domain updates since last propagation.
    # Now do some AC3 like propagation
    # Propagating from X to Y
    for  vy in range(int(self._VY.getInf()), int(self._VY.getSup())+1):
        support = False
        for vx in range(int(self._VX.getInf()), int(self._VX.getSup())+1):
            if vx == vy % self._M:
                support = True
                break
        if not support:
            self._VY.remVal(vy)
    # Propagating from Y to X
    for vx in range(int(self._VX.getInf()), int(self._VX.getSup())+1):
        support = False
        for vy in range(int(self._VY.getInf()), int(self._VY.getSup())+1):
            if vx == vy % self._M:
                support = True
                break
        if not support:
            self._VX.remVal(vx)
public void propagate() {
    int vx, vy;
    for (vy = (int) (y.getInf()); vy <= y.getSup(); vy++) {
        boolean support = false;
        for (vx = (int) (x.getInf()); vx <= x.getSup(); vx++) {
            if (vx == vy % mod) {
                support = true;
                break;
            }
        }
        if (!support) {
            y.remVal(vy);
        }
    }
    for (vx = (int) (x.getInf()); vx <= x.getSup(); vx++) {
        boolean support = false;
        for (vy = (int) (y.getInf()); vy <= y.getSup(); vy++) {
            if (vx == vy % mod) {
                support = true;
                break;
            }
        }
        if (!support) {
            x.remVal(vx);
        }
    }
}

Events on variables

The second kind of events that can be triggered in a constraint is events linked to the variables of the constraint (following decisions taken by the search algorithm or by the filtering of other constraints of the problem). The differents methods related to these events are:

  • awakeOnInf that is triggered upon raising of the lower bound of one variable;
  • awakeOnSup that is triggered upon lowering of the upper bound of one variable;
  • awakeOnRem that is triggered upon deletion of a value from the domain of one variable;
  • awakeOnInst that is triggered upon instantiation of one variable;
  • awakeOnVar that is triggered upon unknown modification of the domain of one variable.

When these methods are not overloaded, the default behavior of the KUserConstraint is to call the propagate method. Overloading one of the awakeOnXXX() methods allows the user to implement a more incremental propagation scheme by giving some information about what has changed since the last call. Note that you can also use the constAwake() method inside one of the awakeOnXXX() methods (in the propagate() method, the call would cause an infinite loop) to force the constraint engine to call the propagate method of this constraint later (Usually when you want to delay the propagation or do dot want to react incrementally to one specific event). For our constraint we will only overload the awakeOnInst() method. (The other events will trigger the propagate() method).

awakeOnInst method

This method is called when a specific variable is instantiated to a specific value. The variable is passed in parameter of the method. The instantiation value can be found by calling the getValue() method of the variable. Since one of the variable is instantiated, the loop on the values of the domain of this variable is useless. Hence we can define a lighter propagation algorithm looking only for one supporting value for the instantiation value of the instantiated variable thus avoiding the loop on the values and the filtering on the other variable:

void ModuloConstraintXYC::awakeOnInst(KIntVar & var)
{
    int v;

    if ( var.isEqualTo(*_VX) )
    {
        // X has been instantiated , removing all values in domain of Y inconsistent with the constraint
        for ( v = _VY->getInf();v<=_VY->getSup();v ++)
        {
            if (_VX->getValue() != v % _M)
            {
                _VY->remVal(v);
            }
        }
    }
    else if ( var.isEqualTo(*_VY) )
    {
        // Y has been instantiated , removing all values in domain of X inconsistent with the constraint
        for ( v = _VX->getInf();v<=_VX->getSup();v ++)
        {
            if ( v != _VY->getValue() % _M)
            {
                _VX->remVal(v);
            }
        }
    }
}
def awakeOnInst(self, var):
    if var.isEqualTo(self._VX):
        # X has been instantiated , removing all values in domain of Y inconsistent with the constraint
        for v in range(int(self._VY.getInf()), int(self._VY.getSup())+1):
            if int(self._VX.getValue()) != v % self._M:
                self._VY.remVal(v)
    elif var.isEqualTo(self._VY):
        # Y has been instantiated , removing all values in domain of X inconsistent with the constraint
        for v in range(int(self._VX.getInf()), int(self._VX.getSup())+1):
            if v != int(self._VY.getValue()) % self._M:
                self._VX.remVal(v)
public void awakeOnInst(KIntVar var) {
    if (var.isEqualTo(x)) {
        for (int v = (int) (y.getInf()); v <= y.getSup(); v++) {
            if ((int) (x.getValue()) != v % mod) {
                y.remVal(v);
            }
        }
    } else if (var.isEqualTo(y)) {
        for (int v = (int) (y.getInf()); v <= y.getSup(); v++) {
            if (v != (int) (y.getValue()) % mod) {
                x.remVal(v);
            }
        }
    }
}

Additionnal methods

The print() method is used by the constraint engine to print the name of the constraint to the standard output. It is mainly used by the KProblem::print() method to print the problem.

void ModuloConstraintXYC::print()
{
    printf("%s == %s mod %i",_VX->getName(),_VY->getName(),_M);
}
def display(self):
    # For pretty printing of the constraint, sys.stdout is use here to avoid newline insertion
    sys.stdout.write("%s == %s mod %i"%(self._VX.getName(), self._VY.getName(), self._M))
public void print() {
    System.out.print(String.format("%s == %s mod %d", x.getName(),
            y.getName(), mod));
}

The askIfentailed() method is used by composite constraints such as KImplies, KGuard, KEquiv, etc. The return value of the method correspond to one of the three following values depending on the status of the constraint:

  • KUserConstraint::CTRUE when the constraint is definitively satisfied
  • KUserConstraint::CFALSE when the constraint is definitively violated
  • KUserConstraint::CUNKNOWN when the status of the constraint is unknown

This listing show an implementation of this method for the \(X=Y mod M\) constraint :

int ModuloConstraintXYC::askIfEntailed()
{
    if (_VX->getIsInstantiated() && _VY->getIsInstantiated() )
    {
        // Here X and Y have been instantiated so we now if the constraint is satisfied or violated
        if ( _VX->getValue() == _VY->getValue() % _M )
        {
            // the constraint hold since X == Y mod M
            return CTRUE;
        }
        else
        {
            // the constraint does not hold since X != Y mod M
            return CFALSE;
        }
    }
    else
    {
        // X or Y is not yet instantiated so we do not know yet if the constraint holds
        return CUNKNOWN;
    }
}
def askIfEntailed(self):
    #Determine whether the constraint is satisfied or not
    if self._VX.getIsInstantiated() and self._VY.getIsInstantiated():
        # Here X and Y have been instantiated so we now if the constraint is satisfied or violated
        if self._VX.getValue() == self._VY.getValue() % self._M:
            # the constraint hold since X == Y mod C
            return KUserConstraint.CTRUE
        else:
            # the constraint does not hold since X != Y mod C
            return KUserConstraint.CFALSE
    else:
        # X or Y is not yet instantiated so we don't know yet if the constraint holds
        return KUserConstraint.CUNKNOWN
public int askIfEntailed() {
    if (x.getIsInstantiated() && y.getIsInstantiated()) {
        if (x.getValue() == y.getValue() % mod) {
            // the constraint hold since X == Y mod C
            return askRet.CTRUE;
        } else {
            // the constraint does not hold since X != Y mod C
            return askRet.CFALSE;
        }
    } else {
        return askRet.CUNKNOWN;
    }
}

The getCopyPtr(KProblem*) method is used by Artelys Kalis to obtain an instance copy of the constraint (see Section 9.4) for details about implementation of this method).

KConstraint* ModuloConstraintXYC::getCopyPtr(const KProblem& problem) const
{
    return new ModuloConstraintXYC(*problem.getCopyPtr(_VX), *problem.getCopyPtr(_VY), _M);
}
def getCopyPtr(self, *args):
    # Obtain a copy of the constraint
    return ModuloConstraintXYC(self._VX, self._VY, self._M)
public KConstraint getCopyPtr() {
    return new ModuloConstraintXYC(x, y, mod);
}