CLP(FD): Constraint Logic Programming over Finite Domains

Development of this library has moved to SICStus Prolog.

Please see CLP(Z) for more information.

Introduction

This library provides CLP(FD): Constraint Logic Programming over Finite Domains. This is an instance of the general CLP(X) scheme, extending logic programming with reasoning over specialised domains.

CLP(FD) lets us reason about integers in a way that honors the relational nature of Prolog. Read The Power of Prolog to understand how this library is meant to be used in practice.

There are two major use cases of CLP(FD) constraints:

declarative integer arithmetic
solving combinatorial problems such as planning, scheduling and allocation tasks.

The predicates of this library can be classified as:

arithmetic constraints like #=/2, #>/2 and #\=/2
the membership constraints in/2 and ins/2
the enumeration predicates indomain/1, label/1 and labeling/2
combinatorial constraints like all_distinct/1 and global_cardinality/2
reification predicates such as #<==>/2
reflection predicates such as fd_dom/2

In most cases, arithmetic constraints are the only predicates you will ever need from this library. When reasoning over integers, simply replace low-level arithmetic predicates like (is)/2 and (>)/2 by the corresponding CLP(FD) constraints like #=/2 and #>/2 to honor and preserve declarative properties of your programs. For satisfactory performance, arithmetic constraints are implicitly rewritten at compilation time so that low-level fallback predicates are automatically used whenever possible.

Almost all Prolog programs also reason about integers. Therefore, it is highly advisable that you make CLP(FD) constraints available in all your programs. One way to do this is to put the following directive in your ~/.swiplrc initialisation file:

:- use_module(library(clpfd)).

All example programs that appear in the CLP(FD) documentation assume that you have done this.

Important concepts and principles of this library are illustrated by means of usage examples that are available in a public git repository: github.com/triska/clpfd

If you are used to the complicated operational considerations that low-level arithmetic primitives necessitate, then moving to CLP(FD) constraints may, due to their power and convenience, at first feel to you excessive and almost like cheating. It isn't. Constraints are an integral part of all popular Prolog systems, and they are designed to help you eliminate and avoid the use of low-level and less general primitives by providing declarative alternatives that are meant to be used instead.

When teaching Prolog, CLP(FD) constraints should be introduced before explaining low-level arithmetic predicates and their procedural idiosyncrasies. This is because constraints are easy to explain, understand and use due to their purely relational nature. In contrast, the modedness and directionality of low-level arithmetic primitives are impure limitations that are better deferred to more advanced lectures.

We recommend the following reference (PDF: metalevel.at/swiclpfd.pdf) for citing this library in scientific publications:

@inproceedings{Triska12,
  author    = {Markus Triska},
  title     = {The Finite Domain Constraint Solver of {SWI-Prolog}},
  booktitle = {FLOPS},
  series    = {LNCS},
  volume    = {7294},
  year      = {2012},
  pages     = {307-316}
}

More information about CLP(FD) constraints and their implementation is contained in: metalevel.at/drt.pdf

The best way to discuss applying, improving and extending CLP(FD) constraints is to use the dedicated clpfd tag on stackoverflow.com. Several of the world's foremost CLP(FD) experts regularly participate in these discussions and will help you for free on this platform.

Arithmetic constraints

In modern Prolog systems, arithmetic constraints subsume and supersede low-level predicates over integers. The main advantage of arithmetic constraints is that they are true relations and can be used in all directions. For most programs, arithmetic constraints are the only predicates you will ever need from this library.

The most important arithmetic constraint is #=/2, which subsumes both (is)/2 and (=:=)/2 over integers. Use #=/2 to make your programs more general.

In total, the arithmetic constraints are:

Expr1 `#=` Expr2	Expr1 equals Expr2
Expr1 `#\=` Expr2	Expr1 is not equal to Expr2
Expr1 `#>=` Expr2	Expr1 is greater than or equal to Expr2
Expr1 `#=<` Expr2	Expr1 is less than or equal to Expr2
Expr1 `#>` Expr2	Expr1 is greater than Expr2
Expr1 `#<` Expr2	Expr1 is less than Expr2

Expr1 and Expr2 denote arithmetic expressions, which are:

integer	Given value
variable	Unknown integer
?(variable)	Unknown integer
-Expr	Unary minus
Expr + Expr	Addition
Expr * Expr	Multiplication
Expr - Expr	Subtraction
Expr ^ Expr	Exponentiation
`min(Expr,Expr)`	Minimum of two expressions
`max(Expr,Expr)`	Maximum of two expressions
Expr `mod` Expr	Modulo induced by floored division
Expr `rem` Expr	Modulo induced by truncated division
`abs(Expr)`	Absolute value
Expr // Expr	Truncated integer division
Expr div Expr	Floored integer division

where Expr again denotes an arithmetic expression.

The bitwise operations (\)/1, (/\)/2, (\/)/2, (>>)/2, (<<)/2, lsb/1, msb/1, popcount/1 and (xor)/2 are also supported.

Declarative integer arithmetic

The arithmetic constraints #=/2, #>/2 etc. are meant to be used instead of the primitives (is)/2, (=:=)/2, (>)/2 etc. over integers. Almost all Prolog programs also reason about integers. Therefore, it is recommended that you put the following directive in your ~/.swiplrc initialisation file to make CLP(FD) constraints available in all your programs:

:- use_module(library(clpfd)).

Throughout the following, it is assumed that you have done this.

The most basic use of CLP(FD) constraints is evaluation of arithmetic expressions involving integers. For example:

?- X #= 1+2.
X = 3.

This could in principle also be achieved with the lower-level predicate (is)/2. However, an important advantage of arithmetic constraints is their purely relational nature: Constraints can be used in all directions, also if one or more of their arguments are only partially instantiated. For example:

?- 3 #= Y+2.
Y = 1.

This relational nature makes CLP(FD) constraints easy to explain and use, and well suited for beginners and experienced Prolog programmers alike. In contrast, when using low-level integer arithmetic, we get:

?- 3 is Y+2.
ERROR: is/2: Arguments are not sufficiently instantiated

?- 3 =:= Y+2.
ERROR: =:=/2: Arguments are not sufficiently instantiated

Due to the necessary operational considerations, the use of these low-level arithmetic predicates is considerably harder to understand and should therefore be deferred to more advanced lectures.

For supported expressions, CLP(FD) constraints are drop-in replacements of these low-level arithmetic predicates, often yielding more general programs. See n_factorial/2 for an example.

This library uses goal_expansion/2 to automatically rewrite constraints at compilation time so that low-level arithmetic predicates are automatically used whenever possible. For example, the predicate:

positive_integer(N) :- N #>= 1.

is executed as if it were written as:

positive_integer(N) :-
        (   integer(N)
        ->  N >= 1
        ;   N #>= 1
        ).

This illustrates why the performance of CLP(FD) constraints is almost always completely satisfactory when they are used in modes that can be handled by low-level arithmetic. To disable the automatic rewriting, set the Prolog flag clpfd_goal_expansion to false.

Example: Factorial relation

We illustrate the benefit of using #=/2 for more generality with a simple example.

Consider first a rather conventional definition of n_factorial/2, relating each natural number N to its factorial F:

n_factorial(0, 1).
n_factorial(N, F) :-
        N #> 0,
        N1 #= N - 1,
        n_factorial(N1, F1),
        F #= N * F1.

This program uses CLP(FD) constraints instead of low-level arithmetic throughout, and everything that would have worked with low-level arithmetic also works with CLP(FD) constraints, retaining roughly the same performance. For example:

?- n_factorial(47, F).
F = 258623241511168180642964355153611979969197632389120000000000 ;
false.

Now the point: Due to the increased flexibility and generality of CLP(FD) constraints, we are free to reorder the goals as follows:

n_factorial(0, 1).
n_factorial(N, F) :-
        N #> 0,
        N1 #= N - 1,
        F #= N * F1,
        n_factorial(N1, F1).

In this concrete case, termination properties of the predicate are improved. For example, the following queries now both terminate:

?- n_factorial(N, 1).
N = 0 ;
N = 1 ;
false.

?- n_factorial(N, 3).
false.

To make the predicate terminate if any argument is instantiated, add the (implied) constraint `F #\= 0` before the recursive call. Otherwise, the query n_factorial(N, 0) is the only non-terminating case of this kind.

The value of CLP(FD) constraints does not lie in completely freeing us from all procedural phenomena. For example, the two programs do not even have the same termination properties in all cases. Instead, the primary benefit of CLP(FD) constraints is that they allow you to try different execution orders and apply declarative debugging techniques at all! Reordering goals (and clauses) can significantly impact the performance of Prolog programs, and you are free to try different variants if you use declarative approaches. Moreover, since all CLP(FD) constraints always terminate, placing them earlier can at most improve, never worsen, the termination properties of your programs. An additional benefit of CLP(FD) constraints is that they eliminate the complexity of introducing (is)/2 and (=:=)/2 to beginners, since both predicates are subsumed by #=/2 when reasoning over integers.

In the case above, the clauses are mutually exclusive if the first argument is sufficiently instantiated. To make the predicate deterministic in such cases while retaining its generality, you can use zcompare/3 to reify a comparison, making the different cases distinguishable by pattern matching. For example, in this concrete case and others like it, you can use zcompare(Comp, 0, N) to obtain as Comp the symbolic outcome (<, =, >) of 0 compared to N.

Combinatorial constraints

In addition to subsuming and replacing low-level arithmetic predicates, CLP(FD) constraints are often used to solve combinatorial problems such as planning, scheduling and allocation tasks. Among the most frequently used combinatorial constraints are all_distinct/1, global_cardinality/2 and cumulative/2. This library also provides several other constraints like disjoint2/1 and automaton/8, which are useful in more specialized applications.

Domains

Each CLP(FD) variable has an associated set of admissible integers, which we call the variable's domain. Initially, the domain of each CLP(FD) variable is the set of all integers. CLP(FD) constraints like #=/2, #>/2 and #\=/2 can at most reduce, and never extend, the domains of their arguments. The constraints in/2 and ins/2 let us explicitly state domains of CLP(FD) variables. The process of determining and adjusting domains of variables is called constraint propagation, and it is performed automatically by this library. When the domain of a variable contains only one element, then the variable is automatically unified to that element.

Domains are taken into account when further constraints are stated, and by enumeration predicates like labeling/2.

Example: Sudoku

As another example, consider Sudoku: It is a popular puzzle over integers that can be easily solved with CLP(FD) constraints.

sudoku(Rows) :-
        length(Rows, 9), maplist(same_length(Rows), Rows),
        append(Rows, Vs), Vs ins 1..9,
        maplist(all_distinct, Rows),
        transpose(Rows, Columns),
        maplist(all_distinct, Columns),
        Rows = [As,Bs,Cs,Ds,Es,Fs,Gs,Hs,Is],
        blocks(As, Bs, Cs),
        blocks(Ds, Es, Fs),
        blocks(Gs, Hs, Is).

blocks([], [], []).
blocks([N1,N2,N3|Ns1], [N4,N5,N6|Ns2], [N7,N8,N9|Ns3]) :-
        all_distinct([N1,N2,N3,N4,N5,N6,N7,N8,N9]),
        blocks(Ns1, Ns2, Ns3).

problem(1, [[_,_,_,_,_,_,_,_,_],
            [_,_,_,_,_,3,_,8,5],
            [_,_,1,_,2,_,_,_,_],
            [_,_,_,5,_,7,_,_,_],
            [_,_,4,_,_,_,1,_,_],
            [_,9,_,_,_,_,_,_,_],
            [5,_,_,_,_,_,_,7,3],
            [_,_,2,_,1,_,_,_,_],
            [_,_,_,_,4,_,_,_,9]]).

Sample query:

?- problem(1, Rows), sudoku(Rows), maplist(writeln, Rows).
[9,8,7,6,5,4,3,2,1]
[2,4,6,1,7,3,9,8,5]
[3,5,1,9,2,8,7,4,6]
[1,2,8,5,3,7,6,9,4]
[6,3,4,8,9,2,1,5,7]
[7,9,5,4,6,1,8,3,2]
[5,1,9,2,8,6,4,7,3]
[4,7,2,3,1,9,5,6,8]
[8,6,3,7,4,5,2,1,9]
Rows = [[9, 8, 7, 6, 5, 4, 3, 2|...], ... , [...|...]].

In this concrete case, the constraint solver is strong enough to find the unique solution without any search.

Residual goals

Here is an example session with a few queries and their answers:

?- X #> 3.
X in 4..sup.

?- X #\= 20.
X in inf..19\/21..sup.

?- 2*X #= 10.
X = 5.

?- X*X #= 144.
X in -12\/12.

?- 4*X + 2*Y #= 24, X + Y #= 9, [X,Y] ins 0..sup.
X = 3,
Y = 6.

?- X #= Y #<==> B, X in 0..3, Y in 4..5.
B = 0,
X in 0..3,
Y in 4..5.

The answers emitted by the toplevel are called residual programs, and the goals that comprise each answer are called residual goals. In each case above, and as for all pure programs, the residual program is declaratively equivalent to the original query. From the residual goals, it is clear that the constraint solver has deduced additional domain restrictions in many cases.

To inspect residual goals, it is best to let the toplevel display them for us. Wrap the call of your predicate into call_residue_vars/2 to make sure that all constrained variables are displayed. To make the constraints a variable is involved in available as a Prolog term for further reasoning within your program, use copy_term/3. For example:

?- X #= Y + Z, X in 0..5, copy_term([X,Y,Z], [X,Y,Z], Gs).
Gs = [clpfd: (X in 0..5), clpfd: (Y+Z#=X)],
X in 0..5,
Y+Z#=X.

This library also provides reflection predicates (like fd_dom/2, fd_size/2 etc.) with which we can inspect a variable's current domain. These predicates can be useful if you want to implement your own labeling strategies.

Core relations and search

Using CLP(FD) constraints to solve combinatorial tasks typically consists of two phases:

First, all relevant constraints are stated.
Second, if the domain of each involved variable is finite, then enumeration predicates can be used to search for concrete solutions.

It is good practice to keep the modeling part, via a dedicated predicate called the core relation, separate from the actual search for solutions. This lets us observe termination and determinism properties of the core relation in isolation from the search, and more easily try different search strategies.

As an example of a constraint satisfaction problem, consider the cryptoarithmetic puzzle SEND + MORE = MONEY, where different letters denote distinct integers between 0 and 9. It can be modeled in CLP(FD) as follows:

puzzle([S,E,N,D] + [M,O,R,E] = [M,O,N,E,Y]) :-
        Vars = [S,E,N,D,M,O,R,Y],
        Vars ins 0..9,
        all_different(Vars),
                  S*1000 + E*100 + N*10 + D +
                  M*1000 + O*100 + R*10 + E #=
        M*10000 + O*1000 + N*100 + E*10 + Y,
        M #\= 0, S #\= 0.

Notice that we are not using labeling/2 in this predicate, so that we can first execute and observe the modeling part in isolation. Sample query and its result (actual variables replaced for readability):

?- puzzle(As+Bs=Cs).
As = [9, A2, A3, A4],
Bs = [1, 0, B3, A2],
Cs = [1, 0, A3, A2, C5],
A2 in 4..7,
all_different([9, A2, A3, A4, 1, 0, B3, C5]),
91*A2+A4+10*B3#=90*A3+C5,
A3 in 5..8,
A4 in 2..8,
B3 in 2..8,
C5 in 2..8.

From this answer, we see that this core relation terminates and is in fact deterministic. Moreover, we see from the residual goals that the constraint solver has deduced more stringent bounds for all variables. Such observations are only possible if modeling and search parts are cleanly separated.

Labeling can then be used to search for solutions in a separate predicate or goal:

?- puzzle(As+Bs=Cs), label(As).
As = [9, 5, 6, 7],
Bs = [1, 0, 8, 5],
Cs = [1, 0, 6, 5, 2] ;
false.

In this case, it suffices to label a subset of variables to find the puzzle's unique solution, since the constraint solver is strong enough to reduce the domains of remaining variables to singleton sets. In general though, it is necessary to label all variables to obtain ground solutions.

Example: Eight queens puzzle

We illustrate the concepts of the preceding sections by means of the so-called eight queens puzzle. The task is to place 8 queens on an 8x8 chessboard such that none of the queens is under attack. This means that no two queens share the same row, column or diagonal.

To express this puzzle via CLP(FD) constraints, we must first pick a suitable representation. Since CLP(FD) constraints reason over integers, we must find a way to map the positions of queens to integers. Several such mappings are conceivable, and it is not immediately obvious which we should use. On top of that, different constraints can be used to express the desired relations. For such reasons, modeling combinatorial problems via CLP(FD) constraints often necessitates some creativity and has been described as more of an art than a science.

In our concrete case, we observe that there must be exactly one queen per column. The following representation therefore suggests itself: We are looking for 8 integers, one for each column, where each integer denotes the row of the queen that is placed in the respective column, and which are subject to certain constraints.

In fact, let us now generalize the task to the so-called N queens puzzle, which is obtained by replacing 8 by N everywhere it occurs in the above description. We implement the above considerations in the core relation n_queens/2, where the first argument is the number of queens (which is identical to the number of rows and columns of the generalized chessboard), and the second argument is a list of N integers that represents a solution in the form described above.

n_queens(N, Qs) :-
        length(Qs, N),
        Qs ins 1..N,
        safe_queens(Qs).

safe_queens([]).
safe_queens([Q|Qs]) :- safe_queens(Qs, Q, 1), safe_queens(Qs).

safe_queens([], _, _).
safe_queens([Q|Qs], Q0, D0) :-
        Q0 #\= Q,
        abs(Q0 - Q) #\= D0,
        D1 #= D0 + 1,
        safe_queens(Qs, Q0, D1).

Note that all these predicates can be used in all directions: We can use them to find solutions, test solutions and complete partially instantiated solutions.

The original task can be readily solved with the following query:

?- n_queens(8, Qs), label(Qs).
Qs = [1, 5, 8, 6, 3, 7, 2, 4] .

Using suitable labeling strategies, we can easily find solutions with 80 queens and more:

?- n_queens(80, Qs), labeling([ff], Qs).
Qs = [1, 3, 5, 44, 42, 4, 50, 7, 68|...] .

?- time((n_queens(90, Qs), labeling([ff], Qs))).
% 5,904,401 inferences, 0.722 CPU in 0.737 seconds (98% CPU)
Qs = [1, 3, 5, 50, 42, 4, 49, 7, 59|...] .

Experimenting with different search strategies is easy because we have separated the core relation from the actual search.

Optimisation

We can use labeling/2 to minimize or maximize the value of a CLP(FD) expression, and generate solutions in increasing or decreasing order of the value. See the labeling options min(Expr) and max(Expr), respectively.

Again, to easily try different labeling options in connection with optimisation, we recommend to introduce a dedicated predicate for posting constraints, and to use labeling/2 in a separate goal. This way, we can observe properties of the core relation in isolation, and try different labeling options without recompiling our code.

If necessary, we can use once/1 to commit to the first optimal solution. However, it is often very valuable to see alternative solutions that are also optimal, so that we can choose among optimal solutions by other criteria. For the sake of purity and completeness, we recommend to avoid once/1 and other constructs that lead to impurities in CLP(FD) programs.

Related to optimisation with CLP(FD) constraints are library(simplex) and CLP(Q) which reason about linear constraints over rational numbers.

Reification

The constraints in/2, #=/2, #\=/2, #</2, #>/2, #=</2, and #>=/2 can be reified, which means reflecting their truth values into Boolean values represented by the integers 0 and 1. Let P and Q denote reifiable constraints or Boolean variables, then:

#\ Q	True iff Q is false
P #\/ Q	True iff either P or Q
P #/\ Q	True iff both P and Q
P #\ Q	True iff either P or Q, but not both
P #<==> Q	True iff P and Q are equivalent
P #==> Q	True iff P implies Q
P #<== Q	True iff Q implies P

The constraints of this table are reifiable as well.

When reasoning over Boolean variables, also consider using CLP(B) constraints as provided by library(clpb).

Enabling monotonic CLP(FD)

In the default execution mode, CLP(FD) constraints still exhibit some non-relational properties. For example, adding constraints can yield new solutions:

?-          X #= 2, X = 1+1.
false.

?- X = 1+1, X #= 2, X = 1+1.
X = 1+1.

This behaviour is highly problematic from a logical point of view, and it may render declarative debugging techniques inapplicable.

Set the Prolog flag clpfd_monotonic to true to make CLP(FD) monotonic: This means that adding new constraints cannot yield new solutions. When this flag is true, we must wrap variables that occur in arithmetic expressions with the functor (?)/1 or (#)/1. For example:

?- set_prolog_flag(clpfd_monotonic, true).
true.

?- #(X) #= #(Y) + #(Z).
#(Y)+ #(Z)#= #(X).

?-          X #= 2, X = 1+1.
ERROR: Arguments are not sufficiently instantiated

The wrapper can be omitted for variables that are already constrained to integers.

Custom constraints

We can define custom constraints. The mechanism to do this is not yet finalised, and we welcome suggestions and descriptions of use cases that are important to you.

As an example of how it can be done currently, let us define a new custom constraint oneground(X,Y,Z), where Z shall be 1 if at least one of X and Y is instantiated:

:- multifile clpfd:run_propagator/2.

oneground(X, Y, Z) :-
        clpfd:make_propagator(oneground(X, Y, Z), Prop),
        clpfd:init_propagator(X, Prop),
        clpfd:init_propagator(Y, Prop),
        clpfd:trigger_once(Prop).

clpfd:run_propagator(oneground(X, Y, Z), MState) :-
        (   integer(X) -> clpfd:kill(MState), Z = 1
        ;   integer(Y) -> clpfd:kill(MState), Z = 1
        ;   true
        ).

First, make_propagator/2 is used to transform a user-defined representation of the new constraint to an internal form. With init_propagator/2, this internal form is then attached to X and Y. From now on, the propagator will be invoked whenever the domains of X or Y are changed. Then, trigger_once/1 is used to give the propagator its first chance for propagation even though the variables' domains have not yet changed. Finally, run_propagator/2 is extended to define the actual propagator. As explained, this predicate is automatically called by the constraint solver. The first argument is the user-defined representation of the constraint as used in make_propagator/2, and the second argument is a mutable state that can be used to prevent further invocations of the propagator when the constraint has become entailed, by using kill/1. An example of using the new constraint:

?- oneground(X, Y, Z), Y = 5.
Y = 5,
Z = 1,
X in inf..sup.

Applications

CLP(FD) applications that we find particularly impressive and worth studying include:

Michael Hendricks uses CLP(FD) constraints for flexible reasoning about dates and times in the julian package.
Julien Cumin uses CLP(FD) constraints for integer arithmetic in Brachylog.

Acknowledgments

This library gives you a glimpse of what SICStus Prolog can do. The API is intentionally mostly compatible with that of SICStus Prolog, so that you can easily switch to a much more feature-rich and much faster CLP(FD) system when you need it. I thank Mats Carlsson, the designer and main implementor of SICStus Prolog, for his elegant example. I first encountered his system as part of the excellent GUPU teaching environment by Ulrich Neumerkel. Ulrich was also the first and most determined tester of the present system, filing hundreds of comments and suggestions for improvement. Tom Schrijvers has contributed several constraint libraries to SWI-Prolog, and I learned a lot from his coding style and implementation examples. Bart Demoen was a driving force behind the implementation of attributed variables in SWI-Prolog, and this library could not even have started without his prior work and contributions. Thank you all!

CLP(FD) predicate index

In the following, each CLP(FD) predicate is described in more detail.

We recommend the following link to refer to this manual:

http://eu.swi-prolog.org/man/clpfd.html

author: - Markus Triska */