Next: Graph Constraints, Previous: Arithmetic-Logical Constraints, Up: Available Constraints [Contents][Index]
element(?X,+List,?Y)
where X and Y are domain variables and List is a list of domain variables. True if the X:th element of List is Y. Operationally, the domains of X and Y are constrained so that for every element in the domain of X, there is a compatible element in the domain of Y, and vice versa.
Maintains domain consistency in X and bounds consistency in
List and Y. Corresponds to nth1/3
in
library(lists)
and to element
and member
in MiniZinc.
relation(?X,+MapList,?Y) since release 4.0.5,deprecated
where X and Y are domain variables and MapList is a
list of integer-ConstantRange
pairs, where the
integer keys occur uniquely (see Syntax of Indexicals). True if
MapList contains a pair X-R
and Y is in
the range denoted by R. Maintains domain consistency.
An arbitrary binary constraint can be defined with relation/3
.
relation/3
is implemented by straightforward transformation to
the following, more general constraint, with which arbitrary relations
can be defined compactly:
table(+Tuples,+Extension)
table(+Tuples,+Extension,+Options)
Defines an n-ary constraint by extension. Extension should be a list of lists of integers, each of length n. Tuples should be a list of lists of domain variables, each also of length n. The constraint holds if every Tuple in Tuples occurs in the Extension. The constraint will maintain domain consistency.
Corresponds to table
in MiniZinc.
For convenience, Extension may contain ConstantRange (see Syntax of Indexicals) expressions in addition to integers.
Options is a list of zero or more of the following:
order(Order) since release 4.1
Determines the variable order used internally. The following values are valid:
leftmost
The order is the one given in the arguments (the default).
id3
Each tuple, and the columns of the extension, is permuted according to the heuristic that more discriminating columns should precede less discriminating ones.
method(Method) since release 4.1
Controls the choice of internal data structure and algorithm. The following values are valid:
default since release 4.4
SICStus choice handles propagation as it sees fit (the default).
noaux
SICStus handles propagation with the case/[3,4]
constraint, see below,
converting the extension to a DAG.
aux
SICStus handles propagation with the case/[3,4]
constraint, see
below. Before converting the extension to a DAG, an auxiliary, first
variable is introduced, denoting extension row number.
table/[2,3]
can be implemented in terms of the following, more
general constraint, which allows a compact representation of arbitrary relations:
case(+Template, +Tuples, +Dag)
case(+Template, +Tuples, +Dag, +Options)
This constraint encodes an n-ary constraint, defined by extension and/or linear inequalities. It uses a DAG-shaped data structure where nodes corresponds to variables and every arc is labeled by an admissible interval for the node above it and optionally by linear inequalities. The variable order is fixed: every path from the root node to a leaf node should visit every variable exactly once, in the order in which they occur lexically in Template. The constraint is true for a single ground tuple if there is a path from the root node to a leaf node such that (a) each tuple element is contained in the corresponding Min..Max interval on the path, and (b) any encountered linear inequalities along the path are true.
The case
constraint is true for a set of ground tuples if it is
true for each tuple of the set. The details are given below.
Template is a nonground Prolog term, each variable of which should occur exactly once. Its variables are merely place-holders; they should not occur outside the constraint nor inside Tuples.
Tuples is a list of terms of the same shape as Template. They should not share any variables with Template.
Dag is a list of nodes of the form
node(ID,X,Children)
, where X is a
template variable, and ID should be an integer, uniquely
identifying each node. The first node in the list is the root
node.
Nodes are either internal nodes or leaf nodes. For an
internal node, Children is a list of terms
(Min..Max)-ID2
or
(Min..Max)-SideConstraints-ID2
, where
ID2 is the ID of a child node, Min is an integer or the atom
inf
(minus infinity), and Max is an integer or the atom
sup
(plus infinity). For a leaf node, Children is a list
of terms (Min..Max)
or
(Min..Max)-SideConstraints
.
SideConstraints is a list of side constraints of the form
scalar_product(Coeffs, Xs, #=<, Bound)
, where
Coeffs is a list of length k of integers, Xs is a list
of length k of template variables, and Bound is an integer.
Variables in Tuples for which their template variable counterparts are constrained by side constraints, must have bounded domains. In the absence of side constraint, the constraint maintains domain consistency.
The use of side constraint is deprecated, because the level of consistency is no better than the alternative, namely to use separate, reified arithmetic constraints and possibly auxiliary variables.
Options is a list of zero or more of the following:
scalar_product(Coeffs, Xs, #=<, Bound) since release 4.2,deprecated
A side constraint located at the root of the DAG.
For example, recall that element(X,L,Y)
wakes
up when the domain of X or the lower or upper bound of Y has
changed, performs full pruning of X, but only prunes the bounds of
Y. The following two constraints:
element(X, [1,1,1,1,2,2,2,2], Y), element(X, [10,10,20,20,10,10,30,30], Z)
can be replaced by the following single constraint, which is equivalent declaratively, but which maintains domain consistency:
elts(X, Y, Z) :- case(f(A,B,C), [f(X,Y,Z)], [node(0, A,[(1..2)-1,(3..4)-2,(5..6)-3,(7..8)-4]), node(1, B,[(1..1)-5]), node(2, B,[(1..1)-6]), node(3, B,[(2..2)-5]), node(4, B,[(2..2)-7]), node(5, C,[(10..10)]), node(6, C,[(20..20)]), node(7, C,[(30..30)])]).
The DAG of the previous example has the following shape:
elts/3
.
A couple of sample queries:
| ?- elts(X, Y, Z). X in 1..8, Y in 1..2, Z in {10}\/{20}\/{30} | ?- elts(X, Y, Z), Z #>= 15. X in(3..4)\/(7..8), Y in 1..2, Z in {20}\/{30} | ?- elts(X, Y, Z), Y = 1. Y = 1, X in 1..4, Z in {10}\/{20}
As an example with side constraints, consider assigning tasks to machines with given unavailibility periods. In this case, one can use a calendar constraint [CHIP 03, Beldiceanu, Carlsson & Rampon 05] to link the real origins of the tasks (taking the unavailibility periods into account) with virtual origins of the tasks (not taking the unavailibility periods into account). One can then state machine resource constraints using the virtual origins, and temporal constraints between the tasks using the real origins.
Assume, for example, three machines with unavailibility periods given by the following table:
Machine 1
is not available during (real) time periods 1-2
and 6-6
, machine 2
is not available during (real) time
periods 3-4
and 7-7
, and machine 3
is always
available.
The following can then be used to express a calendar constraint for a
given task scheduled on machine M in 1..3
, with virtual
origin V in 1..8
, and real origin R in 1..8
:
calendar(M, V, R) :- M in 1..3, V in 1..8, R in 1..8, smt((M#=1 #/\ V in 1..3 #/\ R#=V+2) #\/ (M#=1 #/\ V in 4..5 #/\ R#=V+3) #\/ (M#=2 #/\ V in 1..2 #/\ R#=V) #\/ (M#=2 #/\ V in 3..4 #/\ R#=V+2) #\/ (M#=2 #/\ V in 5..5 #/\ R#=V+3) #\/ (M#=3 #/\ R#=V)).
or equivalently as:
calendar(M, V, R) :- case(f(A,B,C), [f(M,V,R)], [node(0, A, [(1..1)-1, (2..2)-2, (3..3)-3]), node(1, B, [(1..3)-[scalar_product([1,-1],[B,C],#=<,-2), scalar_product([1,-1],[C,B],#=<, 2)]-4, (4..5)-[scalar_product([1,-1],[B,C],#=<,-3), scalar_product([1,-1],[C,B],#=<, 3)]-4]), node(2, B, [(1..2)-[scalar_product([1,-1],[B,C],#=<, 0), scalar_product([1,-1],[C,B],#=<, 0)]-4, (3..4)-[scalar_product([1,-1],[B,C],#=<,-2), scalar_product([1,-1],[C,B],#=<, 2)]-4, (5..5)-[scalar_product([1,-1],[B,C],#=<,-3), scalar_product([1,-1],[C,B],#=<, 3)]-4]), node(3, B, [(1..8)-[scalar_product([1,-1],[B,C],#=<, 0), scalar_product([1,-1],[C,B],#=<, 0)]-4]), node(4, C, [(1..8)])]).
Note that equality must be modeled as the conjunction of inequalities,
as only constraints of the form scalar_product(+Coeffs,
+Xs, #=<, +Bound)
are allowed as side constraints.
The DAG of the calendar constraint has the following shape:
calendar/3
.
A couple of sample queries:
| ?- M in 1..3, V in 1..8, R in 1..8, calendar(M, V, R). M in 1..3, V in 1..8, R in 1..8 | ?- M in 1..3, V in 1..8, R in 1..8, calendar(M, V, R), M #= 1. M = 1, V in 1..5, R in 1..8 | ?- M in 1..3, V in 1..8, R in 1..8, calendar(M, V, R), M #= 2, V #> 4. M = 2, V = 5, R = 8