/* Part of SWI-Prolog Author: Markus Triska E-mail: triska@metalevel.at WWW: http://www.swi-prolog.org Copyright (C): 2014-2016 Markus Triska All rights reserved. Redistribution and use in source and binary forms, with or without modification, are permitted provided that the following conditions are met: 1. Redistributions of source code must retain the above copyright notice, this list of conditions and the following disclaimer. 2. Redistributions in binary form must reproduce the above copyright notice, this list of conditions and the following disclaimer in the documentation and/or other materials provided with the distribution. THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. */ /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - CLP(B): Constraint Logic Programming over Boolean variables. - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */ :- module(clpb, [ op(300, fy, ~), op(500, yfx, #), sat/1, taut/2, labeling/1, sat_count/2, weighted_maximum/3, random_labeling/2 ]). :- use_module(library(error)). :- use_module(library(assoc)). :- use_module(library(apply_macros)). :- create_prolog_flag(clpb_monotonic, false, []). :- create_prolog_flag(clpb_residuals, default, []). /** CLP(B): Constraint Logic Programming over Boolean Variables ## Introduction {#clpb-intro} This library provides CLP(B), Constraint Logic Programming over Boolean variables. It can be used to model and solve combinatorial problems such as verification, allocation and covering tasks. CLP(B) is an instance of the general [CLP(_X_) scheme](<#clp>), extending logic programming with reasoning over specialised domains. The implementation is based on reduced and ordered Binary Decision Diagrams (BDDs). Benchmarks and usage examples of this library are available from: [__https://www.metalevel.at/clpb/__](https://www.metalevel.at/clpb/) We recommend the following reference (PDF: [https://www.metalevel.at/swiclpb.pdf](https://www.metalevel.at/swiclpb.pdf)) for citing this library in scientific publications: == @inproceedings{Triska2016, author = "Markus Triska", title = "The {Boolean} Constraint Solver of {SWI-Prolog}: System Description", booktitle = "FLOPS", series = "LNCS", volume = 9613, year = 2016, pages = "45--61" } == and the following URL to link to its documentation: http://eu.swi-prolog.org/man/clpb.html ## Boolean expressions {#clpb-exprs} A _Boolean expression_ is one of: | `0` | false | | `1` | true | | _variable_ | unknown truth value | | _atom_ | universally quantified variable | | ~ _Expr_ | logical NOT | | _Expr_ + _Expr_ | logical OR | | _Expr_ * _Expr_ | logical AND | | _Expr_ # _Expr_ | exclusive OR | | _Var_ ^ _Expr_ | existential quantification | | _Expr_ =:= _Expr_ | equality | | _Expr_ =\= _Expr_ | disequality (same as #) | | _Expr_ =< _Expr_ | less or equal (implication) | | _Expr_ >= _Expr_ | greater or equal | | _Expr_ < _Expr_ | less than | | _Expr_ > _Expr_ | greater than | | card(Is,Exprs) | _see below_ | | `+(Exprs)` | _see below_ | | `*(Exprs)` | _see below_ | where _Expr_ again denotes a Boolean expression. The Boolean expression card(Is,Exprs) is true iff the number of true expressions in the list `Exprs` is a member of the list `Is` of integers and integer ranges of the form `From-To`. `+(Exprs)` and `*(Exprs)` denote, respectively, the disjunction and conjunction of all elements in the list `Exprs` of Boolean expressions. Atoms denote parametric values that are universally quantified. All universal quantifiers appear implicitly in front of the entire expression. In residual goals, universally quantified variables always appear on the right-hand side of equations. Therefore, they can be used to express functional dependencies on input variables. ## Interface predicates {#clpb-interface} The most frequently used CLP(B) predicates are: * sat(+Expr) True iff the Boolean expression Expr is satisfiable. * taut(+Expr, -T) If Expr is a tautology with respect to the posted constraints, succeeds with *T = 1*. If Expr cannot be satisfied, succeeds with *T = 0*. Otherwise, it fails. * labeling(+Vs) Assigns truth values to the variables Vs such that all constraints are satisfied. The unification of a CLP(B) variable _X_ with a term _T_ is equivalent to posting the constraint sat(X=:=T). ## Examples {#clpb-examples} Here is an example session with a few queries and their answers: == ?- use_module(library(clpb)). true. ?- sat(X*Y). X = Y, Y = 1. ?- sat(X * ~X). false. ?- taut(X * ~X, T). T = 0, sat(X=:=X). ?- sat(X^Y^(X+Y)). sat(X=:=X), sat(Y=:=Y). ?- sat(X*Y + X*Z), labeling([X,Y,Z]). X = Z, Z = 1, Y = 0 ; X = Y, Y = 1, Z = 0 ; X = Y, Y = Z, Z = 1. ?- sat(X =< Y), sat(Y =< Z), taut(X =< Z, T). T = 1, sat(X=:=X*Y), sat(Y=:=Y*Z). ?- sat(1#X#a#b). sat(X=:=a#b). == The pending residual goals constrain remaining variables to Boolean expressions and are declaratively equivalent to the original query. The last example illustrates that when applicable, remaining variables are expressed as functions of universally quantified variables. ## Obtaining BDDs {#clpb-residual-goals} By default, CLP(B) residual goals appear in (approximately) algebraic normal form (ANF). This projection is often computationally expensive. You can set the Prolog flag `clpb_residuals` to the value `bdd` to see the BDD representation of all constraints. This results in faster projection to residual goals, and is also useful for learning more about BDDs. For example: == ?- set_prolog_flag(clpb_residuals, bdd). true. ?- sat(X#Y). node(3)- (v(X, 0)->node(2);node(1)), node(1)- (v(Y, 1)->true;false), node(2)- (v(Y, 1)->false;true). == Note that this representation cannot be pasted back on the toplevel, and its details are subject to change. Use copy_term/3 to obtain such answers as Prolog terms. The variable order of the BDD is determined by the order in which the variables first appear in constraints. To obtain different orders, you can for example use: == ?- sat(+[1,Y,X]), sat(X#Y). node(3)- (v(Y, 0)->node(2);node(1)), node(1)- (v(X, 1)->true;false), node(2)- (v(X, 1)->false;true). == ## Enabling monotonic CLP(B) {#clpb-monotonic} In the default execution mode, CLP(B) constraints are _not_ monotonic. This means that _adding_ constraints can yield new solutions. For example: == ?- sat(X=:=1), X = 1+0. false. ?- X = 1+0, sat(X=:=1), X = 1+0. X = 1+0. == This behaviour is highly problematic from a logical point of view, and it may render [**declarative debugging**](https://www.metalevel.at/prolog/debugging) techniques inapplicable. Set the flag `clpb_monotonic` to `true` to make CLP(B) *monotonic*. If this mode is enabled, then you must wrap CLP(B) variables with the functor `v/1`. For example: == ?- set_prolog_flag(clpb_monotonic, true). true. ?- sat(v(X)=:=1#1). X = 0. == @author [Markus Triska](https://www.metalevel.at) */ /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - Each CLP(B) variable belongs to exactly one BDD. Each CLP(B) variable gets an attribute (in module "clpb") of the form: index_root(Index,Root) where Index is the variable's unique integer index, and Root is the root of the BDD that the variable belongs to. Each CLP(B) variable also gets an attribute in module clpb_hash: an association table node(LID,HID) -> Node, to keep the BDD reduced. The association table of each variable must be rebuilt on occasion to remove nodes that are no longer reachable. We rebuild the association tables of involved variables after BDDs are merged to build a new root. This only serves to reclaim memory: Keeping a node in a local table even when it no longer occurs in any BDD does not affect the solver's correctness. However, apply_shortcut/4 relies on the invariant that every node that occurs in the relevant BDDs is also registered in the table of its branching variable. A root is a logical variable with a single attribute ("clpb_bdd") of the form: Sat-BDD where Sat is the SAT formula (in original form) that corresponds to BDD. Sat is necessary to rebuild the BDD after variable aliasing, and to project all remaining constraints to a list of sat/1 goals. Finally, a BDD is either: *) The integers 0 or 1, denoting false and true, respectively, or *) A node of the form node(ID, Var, Low, High, Aux) Where ID is the node's unique integer ID, Var is the node's branching variable, and Low and High are the node's low (Var = 0) and high (Var = 1) children. Aux is a free variable, one for each node, that can be used to attach attributes and store intermediate results. Variable aliasing is treated as a conjunction of corresponding SAT formulae. You should think of CLP(B) as a potentially vast collection of BDDs that can range from small to gigantic in size, and which can merge. - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */ /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - Type checking. - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */ is_sat(V) :- var(V), !, non_monotonic(V). is_sat(v(V)) :- var(V), !. is_sat(v(I)) :- integer(I), between(0, 1, I). is_sat(I) :- integer(I), between(0, 1, I). is_sat(A) :- atom(A). is_sat(~A) :- is_sat(A). is_sat(A*B) :- is_sat(A), is_sat(B). is_sat(A+B) :- is_sat(A), is_sat(B). is_sat(A#B) :- is_sat(A), is_sat(B). is_sat(A=:=B) :- is_sat(A), is_sat(B). is_sat(A=\=B) :- is_sat(A), is_sat(B). is_sat(A==B) :- is_sat(A), is_sat(B). is_sat(AB) :- is_sat(A), is_sat(B). is_sat(+(Ls)) :- must_be(list, Ls), maplist(is_sat, Ls). is_sat(*(Ls)) :- must_be(list, Ls), maplist(is_sat, Ls). is_sat(X^F) :- var(X), is_sat(F). is_sat(card(Is,Fs)) :- must_be(list(ground), Is), must_be(list, Fs), maplist(is_sat, Fs). non_monotonic(X) :- ( var_index(X, _) -> % OK: already constrained to a CLP(B) variable true ; current_prolog_flag(clpb_monotonic, true) -> instantiation_error(X) ; true ). /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - Rewriting to canonical expressions. Atoms are converted to variables with a special attribute. A global lookup table maintains the correspondence between atoms and their variables throughout different sat/1 goals. - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */ % elementary sat_rewrite(V, V) :- var(V), !. sat_rewrite(I, I) :- integer(I), !. sat_rewrite(A, V) :- atom(A), !, clpb_atom_var(A, V). sat_rewrite(v(V), V). sat_rewrite(P0*Q0, P*Q) :- sat_rewrite(P0, P), sat_rewrite(Q0, Q). sat_rewrite(P0+Q0, P+Q) :- sat_rewrite(P0, P), sat_rewrite(Q0, Q). sat_rewrite(P0#Q0, P#Q) :- sat_rewrite(P0, P), sat_rewrite(Q0, Q). sat_rewrite(X^F0, X^F) :- sat_rewrite(F0, F). sat_rewrite(card(Is,Fs0), card(Is,Fs)) :- maplist(sat_rewrite, Fs0, Fs). % synonyms sat_rewrite(~P, R) :- sat_rewrite(1 # P, R). sat_rewrite(P =:= Q, R) :- sat_rewrite(~P # Q, R). sat_rewrite(P =\= Q, R) :- sat_rewrite(P # Q, R). sat_rewrite(P =< Q, R) :- sat_rewrite(~P + Q, R). sat_rewrite(P >= Q, R) :- sat_rewrite(Q =< P, R). sat_rewrite(P < Q, R) :- sat_rewrite(~P * Q, R). sat_rewrite(P > Q, R) :- sat_rewrite(Q < P, R). sat_rewrite(+(Ls), R) :- foldl(or, Ls, 0, F), sat_rewrite(F, R). sat_rewrite(*(Ls), R) :- foldl(and, Ls, 1, F), sat_rewrite(F, R). or(A, B, B + A). and(A, B, B * A). must_be_sat(Sat) :- must_be(acyclic, Sat), ( is_sat(Sat) -> true ; no_truth_value(Sat) ). no_truth_value(Term) :- domain_error(clpb_expr, Term). parse_sat(Sat0, Sat) :- must_be_sat(Sat0), sat_rewrite(Sat0, Sat), term_variables(Sat, Vs), maplist(enumerate_variable, Vs). enumerate_variable(V) :- ( var_index_root(V, _, _) -> true ; clpb_next_id('$clpb_next_var', Index), put_attr(V, clpb, index_root(Index,_)), put_empty_hash(V) ). var_index(V, I) :- var_index_root(V, I, _). var_index_root(V, I, Root) :- get_attr(V, clpb, index_root(I,Root)). put_empty_hash(V) :- empty_assoc(H0), put_attr(V, clpb_hash, H0). sat_roots(Sat, Roots) :- term_variables(Sat, Vs), maplist(var_index_root, Vs, _, Roots0), term_variables(Roots0, Roots). %% sat(+Expr) is semidet. % % True iff Expr is a satisfiable Boolean expression. sat(Sat0) :- ( phrase(sat_ands(Sat0), Ands), Ands = [_,_|_] -> maplist(sat, Ands) ; parse_sat(Sat0, Sat), sat_bdd(Sat, BDD), sat_roots(Sat, Roots), roots_and(Roots, Sat0-BDD, And-BDD1), maplist(del_bdd, Roots), maplist(=(Root), Roots), root_put_formula_bdd(Root, And, BDD1), is_bdd(BDD1), satisfiable_bdd(BDD1) ). /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - Posting many small sat/1 constraints is better than posting a huge conjunction (or negated disjunction), because unneeded nodes are removed from node tables after BDDs are merged. This is not possible in sat_bdd/2 because the nodes may occur in other BDDs. A better version of sat_bdd/2 or a proper implementation of a unique table including garbage collection would make this obsolete and also improve taut/2 and sat_count/2 in such cases. - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */ sat_ands(X) --> ( { var(X) } -> [X] ; { X = (A*B) } -> sat_ands(A), sat_ands(B) ; { X = *(Ls) } -> sat_ands_(Ls) ; { X = ~Y } -> not_ors(Y) ; [X] ). sat_ands_([]) --> []. sat_ands_([L|Ls]) --> [L], sat_ands_(Ls). not_ors(X) --> ( { var(X) } -> [~X] ; { X = (A+B) } -> not_ors(A), not_ors(B) ; { X = +(Ls) } -> not_ors_(Ls) ; [~X] ). not_ors_([]) --> []. not_ors_([L|Ls]) --> [~L], not_ors_(Ls). del_bdd(Root) :- del_attr(Root, clpb_bdd). root_get_formula_bdd(Root, F, BDD) :- get_attr(Root, clpb_bdd, F-BDD). root_put_formula_bdd(Root, F, BDD) :- put_attr(Root, clpb_bdd, F-BDD). roots_and(Roots, Sat0-BDD0, Sat-BDD) :- foldl(root_and, Roots, Sat0-BDD0, Sat-BDD), rebuild_hashes(BDD). root_and(Root, Sat0-BDD0, Sat-BDD) :- ( root_get_formula_bdd(Root, F, B) -> Sat = F*Sat0, bdd_and(B, BDD0, BDD) ; Sat = Sat0, BDD = BDD0 ). bdd_and(NA, NB, And) :- apply(*, NA, NB, And), is_bdd(And). %% taut(+Expr, -T) is semidet % % Tautology check. Succeeds with T = 0 if the Boolean expression Expr % cannot be satisfied, and with T = 1 if Expr is always true with % respect to the current constraints. Fails otherwise. taut(Sat0, T) :- parse_sat(Sat0, Sat), ( T = 0, \+ sat(Sat) -> true ; T = 1, tautology(Sat) -> true ; false ). /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - The algebraic equivalence: tautology(F) <=> \+ sat(~F) does NOT hold in CLP(B) because the quantifiers of universally quantified variables always implicitly appear in front of the *entire* expression. Thus we have for example: X+a is not a tautology, but ~(X+a), meaning forall(a, ~(X+a)), is unsatisfiable: sat(~(X+a)) = sat(~X * ~a) = sat(~X), sat(~a) = X=0, false The actual negation of X+a, namely ~forall(A,X+A), in terms of CLP(B): ~ ~exists(A, ~(X+A)), is of course satisfiable: ?- sat(~ ~A^ ~(X+A)). %@ X = 0, %@ sat(A=:=A). Instead, of such rewriting, we test whether the BDD of the negated formula is 0. Critically, this avoids constraint propagation. - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */ tautology(Sat) :- ( phrase(sat_ands(Sat), Ands), Ands = [_,_|_] -> maplist(tautology, Ands) ; catch((sat_roots(Sat, Roots), roots_and(Roots, _-1, _-Ands), sat_bdd(1#Sat, BDD), bdd_and(BDD, Ands, B), B == 0, % reset all attributes throw(tautology)), tautology, true) ). satisfiable_bdd(BDD) :- ( BDD == 0 -> false ; BDD == 1 -> true ; ( bdd_nodes(var_unbound, BDD, Nodes) -> bdd_variables_classification(BDD, Nodes, Classes), partition(var_class, Classes, Eqs, Bs, Ds), domain_consistency(Eqs, Goal), aliasing_consistency(Bs, Ds, Goals), maplist(unification, [Goal|Goals]) ; % if any variable is instantiated, we do not perform % any propagation for now true ) ). var_class(_=_, <). var_class(further_branching(_,_), =). var_class(negative_decisive(_), >). unification(true). unification(A=B) :- A = B. % safe_goal/1 detects safety of this call var_unbound(Node) :- node_var_low_high(Node, Var, _, _), var(Var). universal_var(Var) :- get_attr(Var, clpb_atom, _). /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - By aliasing consistency, we mean that all unifications X=Y, where taut(X=:=Y, 1) holds, are posted. To detect this, we distinguish two kinds of variables among those variables that are not skipped in any branch: further-branching and negative-decisive. X is negative-decisive iff every node where X appears as a branching variable has 0 as one of its children. X is further-branching iff 1 is not a direct child of any node where X appears as a branching variable. Any potential aliasing must involve one further-branching, and one negative-decisive variable. X=Y must hold if, for each low branch of nodes with X as branching variable, Y has high branch 0, and for each high branch of nodes involving X, Y has low branch 0. - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */ aliasing_consistency(Bs, Ds, Goals) :- phrase(aliasings(Bs, Ds), Goals). aliasings([], _) --> []. aliasings([further_branching(B,Nodes)|Bs], Ds) --> { var_index(B, BI) }, aliasings_(Ds, B, BI, Nodes), aliasings(Bs, Ds). aliasings_([], _, _, _) --> []. aliasings_([negative_decisive(D)|Ds], B, BI, Nodes) --> { var_index(D, DI) }, ( { DI > BI, always_false(high, DI, Nodes), always_false(low, DI, Nodes), var_or_atom(D, DA), var_or_atom(B, BA) } -> [DA=BA] ; [] ), aliasings_(Ds, B, BI, Nodes). var_or_atom(Var, VA) :- ( get_attr(Var, clpb_atom, VA) -> true ; VA = Var ). always_false(Which, DI, Nodes) :- phrase(nodes_always_false(Nodes, Which, DI), Opposites), maplist(with_aux(unvisit), Opposites). nodes_always_false([], _, _) --> []. nodes_always_false([Node|Nodes], Which, DI) --> { which_node_child(Which, Node, Child), opposite(Which, Opposite) }, opposite_always_false(Opposite, DI, Child), nodes_always_false(Nodes, Which, DI). which_node_child(low, Node, Child) :- node_var_low_high(Node, _, Child, _). which_node_child(high, Node, Child) :- node_var_low_high(Node, _, _, Child). opposite(low, high). opposite(high, low). opposite_always_false(Opposite, DI, Node) --> ( { node_visited(Node) } -> [] ; { node_var_low_high(Node, Var, Low, High), with_aux(put_visited, Node), var_index(Var, VI) }, [Node], ( { VI =:= DI } -> { which_node_child(Opposite, Node, Child), Child == 0 } ; opposite_always_false(Opposite, DI, Low), opposite_always_false(Opposite, DI, High) ) ). further_branching(Node) :- node_var_low_high(Node, _, Low, High), Low \== 1, High \== 1. negative_decisive(Node) :- node_var_low_high(Node, _, Low, High), ( Low == 0 -> true ; High == 0 -> true ; false ). /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - Instantiate all variables that only admit a single Boolean value. This is the case if: The variable is not skipped in any branch leading to 1 (its being skipped means that it may be assigned either 0 or 1 and can thus not be fixed yet), and all nodes where it occurs as a branching variable have either lower or upper child fixed to 0 consistently. - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */ domain_consistency(Eqs, Goal) :- maplist(eq_a_b, Eqs, Vs, Values), Goal = (Vs = Values). % propagate all assignments at once eq_a_b(A=B, A, B). consistently_false_(Which, Node) :- which_node_child(Which, Node, Child), Child == 0. /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - In essentially one sweep of the BDD, all variables can be classified: Unification with 0 or 1, further branching and/or negative decisive. Strategy: Breadth-first traversal of the BDD, failing (and thus clearing all attributes) if the variable is skipped in some branch, and moving the frontier along each time. A formula is only satisfiable if it is a tautology after all (also implicitly) existentially quantified variables are projected away. However, we only need to check this explicitly if at least one universally quantified variable appears. Otherwise, we know that the formula is satisfiable at this point, because its BDD is not 0. - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */ bdd_variables_classification(BDD, Nodes, Classes) :- nodes_variables(Nodes, Vs0), variables_in_index_order(Vs0, Vs), ( partition(universal_var, Vs, [_|_], Es) -> foldl(existential, Es, BDD, 1) ; true ), phrase(variables_classification(Vs, [BDD]), Classes), maplist(with_aux(unvisit), Nodes). variables_classification([], _) --> []. variables_classification([V|Vs], Nodes0) --> { var_index(V, Index) }, ( { phrase(nodes_with_variable(Nodes0, Index), Nodes) } -> ( { maplist(consistently_false_(low), Nodes) } -> [V=1] ; { maplist(consistently_false_(high), Nodes) } -> [V=0] ; [] ), ( { maplist(further_branching, Nodes) } -> [further_branching(V, Nodes)] ; [] ), ( { maplist(negative_decisive, Nodes) } -> [negative_decisive(V)] ; [] ), { maplist(with_aux(unvisit), Nodes) }, variables_classification(Vs, Nodes) ; variables_classification(Vs, Nodes0) ). nodes_with_variable([], _) --> []. nodes_with_variable([Node|Nodes], VI) --> { Node \== 1 }, ( { node_visited(Node) } -> nodes_with_variable(Nodes, VI) ; { with_aux(put_visited, Node), node_var_low_high(Node, OVar, Low, High), var_index(OVar, OVI) }, { OVI =< VI }, ( { OVI =:= VI } -> [Node] ; nodes_with_variable([Low,High], VI) ), nodes_with_variable(Nodes, VI) ). /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - Node management. Always use an existing node, if there is one. - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */ make_node(Var, Low, High, Node) :- ( Low == High -> Node = Low ; low_high_key(Low, High, Key), ( lookup_node(Var, Key, Node) -> true ; clpb_next_id('$clpb_next_node', ID), Node = node(ID,Var,Low,High,_Aux), register_node(Var, Key, Node) ) ). make_node(Var, Low, High, Node) --> % make it conveniently usable within DCGs { make_node(Var, Low, High, Node) }. % The key of a node for hashing is determined by the IDs of its % children. low_high_key(Low, High, node(LID,HID)) :- node_id(Low, LID), node_id(High, HID). rebuild_hashes(BDD) :- bdd_nodes(nodevar_put_empty_hash, BDD, Nodes), maplist(re_register_node, Nodes). nodevar_put_empty_hash(Node) :- node_var_low_high(Node, Var, _, _), empty_assoc(H0), put_attr(Var, clpb_hash, H0). re_register_node(Node) :- node_var_low_high(Node, Var, Low, High), low_high_key(Low, High, Key), register_node(Var, Key, Node). register_node(Var, Key, Node) :- get_attr(Var, clpb_hash, H0), put_assoc(Key, H0, Node, H), put_attr(Var, clpb_hash, H). lookup_node(Var, Key, Node) :- get_attr(Var, clpb_hash, H0), get_assoc(Key, H0, Node). node_id(0, false). node_id(1, true). node_id(node(ID,_,_,_,_), ID). node_aux(Node, Aux) :- arg(5, Node, Aux). node_var_low_high(Node, Var, Low, High) :- Node = node(_,Var,Low,High,_). /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - sat_bdd/2 converts a SAT formula in canonical form to an ordered and reduced BDD. - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */ sat_bdd(V, Node) :- var(V), !, make_node(V, 0, 1, Node). sat_bdd(I, I) :- integer(I), !. sat_bdd(V^Sat, Node) :- !, sat_bdd(Sat, BDD), existential(V, BDD, Node). sat_bdd(card(Is,Fs), Node) :- !, counter_network(Is, Fs, Node). sat_bdd(Sat, Node) :- !, Sat =.. [F,A,B], sat_bdd(A, NA), sat_bdd(B, NB), apply(F, NA, NB, Node). existential(V, BDD, Node) :- var_index(V, Index), bdd_restriction(BDD, Index, 0, NA), bdd_restriction(BDD, Index, 1, NB), apply(+, NA, NB, Node). /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - Counter network for card(Is,Fs). - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */ counter_network(Cs, Fs, Node) :- same_length([_|Fs], Indicators), fill_indicators(Indicators, 0, Cs), phrase(formulas_variables(Fs, Vars0), ExBDDs), maplist(unvisit, Vars0), % The counter network is built bottom-up, so variables with % highest index must be processed first. variables_in_index_order(Vars0, Vars1), reverse(Vars1, Vars), counter_network_(Vars, Indicators, Node0), foldl(existential_and, ExBDDs, Node0, Node). % Introduce fresh variables for expressions that are not variables. % These variables are later existentially quantified to remove them. % Also, new variables are introduced for variables that are used more % than once, as in card([0,1],[X,X,Y]), to keep the BDD ordered. formulas_variables([], []) --> []. formulas_variables([F|Fs], [V|Vs]) --> ( { var(F), \+ is_visited(F) } -> { V = F, put_visited(F) } ; { enumerate_variable(V), sat_rewrite(V =:= F, Sat), sat_bdd(Sat, BDD) }, [V-BDD] ), formulas_variables(Fs, Vs). counter_network_([], [Node], Node). counter_network_([Var|Vars], [I|Is0], Node) :- foldl(indicators_pairing(Var), Is0, Is, I, _), counter_network_(Vars, Is, Node). indicators_pairing(Var, I, Node, Prev, I) :- make_node(Var, Prev, I, Node). fill_indicators([], _, _). fill_indicators([I|Is], Index0, Cs) :- ( memberchk(Index0, Cs) -> I = 1 ; member(A-B, Cs), between(A, B, Index0) -> I = 1 ; I = 0 ), Index1 is Index0 + 1, fill_indicators(Is, Index1, Cs). existential_and(Ex-BDD, Node0, Node) :- bdd_and(BDD, Node0, Node1), existential(Ex, Node1, Node), % remove attributes to avoid residual goals for variables that % are only used as substitutes for formulas del_attrs(Ex). /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - Compute F(NA, NB). We use a DCG to thread through an implicit argument G0, an association table F(IDA,IDB) -> Node, used for memoization. - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */ apply(F, NA, NB, Node) :- empty_assoc(G0), phrase(apply(F, NA, NB, Node), [G0], _). apply(F, NA, NB, Node) --> ( { integer(NA), integer(NB) } -> { once(bool_op(F, NA, NB, Node)) } ; { apply_shortcut(F, NA, NB, Node) } -> [] ; { node_id(NA, IDA), node_id(NB, IDB), Key =.. [F,IDA,IDB] }, ( state(G0), { get_assoc(Key, G0, Node) } -> [] ; apply_(F, NA, NB, Node), state(G0, G), { put_assoc(Key, G0, Node, G) } ) ). apply_shortcut(+, NA, NB, Node) :- ( NA == 0 -> Node = NB ; NA == 1 -> Node = 1 ; NB == 0 -> Node = NA ; NB == 1 -> Node = 1 ; false ). apply_shortcut(*, NA, NB, Node) :- ( NA == 0 -> Node = 0 ; NA == 1 -> Node = NB ; NB == 0 -> Node = 0 ; NB == 1 -> Node = NA ; false ). apply_(F, NA, NB, Node) --> { var_less_than(NA, NB), !, node_var_low_high(NA, VA, LA, HA) }, apply(F, LA, NB, Low), apply(F, HA, NB, High), make_node(VA, Low, High, Node). apply_(F, NA, NB, Node) --> { node_var_low_high(NA, VA, LA, HA), node_var_low_high(NB, VB, LB, HB), VA == VB }, !, apply(F, LA, LB, Low), apply(F, HA, HB, High), make_node(VA, Low, High, Node). apply_(F, NA, NB, Node) --> % NB < NA { node_var_low_high(NB, VB, LB, HB) }, apply(F, NA, LB, Low), apply(F, NA, HB, High), make_node(VB, Low, High, Node). node_varindex(Node, VI) :- node_var_low_high(Node, V, _, _), var_index(V, VI). var_less_than(NA, NB) :- ( integer(NB) -> true ; node_varindex(NA, VAI), node_varindex(NB, VBI), VAI < VBI ). bool_op(+, 0, 0, 0). bool_op(+, 0, 1, 1). bool_op(+, 1, 0, 1). bool_op(+, 1, 1, 1). bool_op(*, 0, 0, 0). bool_op(*, 0, 1, 0). bool_op(*, 1, 0, 0). bool_op(*, 1, 1, 1). bool_op(#, 0, 0, 0). bool_op(#, 0, 1, 1). bool_op(#, 1, 0, 1). bool_op(#, 1, 1, 0). /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - Access implicit state in DCGs. - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */ state(S) --> state(S, S). state(S0, S), [S] --> [S0]. /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - Unification. X = Expr is equivalent to sat(X =:= Expr). Current limitation: =================== The current interface of attributed variables is not general enough to express what we need. For example, ?- sat(A + B), A = A + 1. should be equivalent to ?- sat(A + B), sat(A =:= A + 1). However, attr_unify_hook/2 is only called *after* the unification of A with A + 1 has already taken place and turned A into a cyclic ground term, raised an error or failed (depending on the flag occurs_check), making it impossible to reason about the variable A in the unification hook. Therefore, a more general interface for attributed variables should replace the current one. In particular, unification filters should be able to reason about terms before they are unified with anything. - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */ attr_unify_hook(index_root(I,Root), Other) :- ( integer(Other) -> ( between(0, 1, Other) -> root_get_formula_bdd(Root, Sat, BDD0), bdd_restriction(BDD0, I, Other, BDD), root_put_formula_bdd(Root, Sat, BDD), satisfiable_bdd(BDD) ; no_truth_value(Other) ) ; atom(Other) -> root_get_formula_bdd(Root, Sat0, _), parse_sat(Sat0, Sat), sat_bdd(Sat, BDD), root_put_formula_bdd(Root, Sat0, BDD), is_bdd(BDD), satisfiable_bdd(BDD) ; % due to variable aliasing, any BDDs may now be unordered, % so we need to rebuild the new BDD from the conjunction. root_get_formula_bdd(Root, Sat0, _), Sat = Sat0*OtherSat, ( var(Other), var_index_root(Other, _, OtherRoot), OtherRoot \== Root -> root_get_formula_bdd(OtherRoot, OtherSat, _), parse_sat(Sat, Sat1), sat_bdd(Sat1, BDD1), And = Sat, sat_roots(Sat, Roots) ; parse_sat(Other, OtherSat), sat_roots(Sat, Roots), maplist(root_rebuild_bdd, Roots), roots_and(Roots, 1-1, And-BDD1) ), maplist(del_bdd, Roots), maplist(=(NewRoot), Roots), root_put_formula_bdd(NewRoot, And, BDD1), is_bdd(BDD1), satisfiable_bdd(BDD1) ). root_rebuild_bdd(Root) :- ( root_get_formula_bdd(Root, F0, _) -> parse_sat(F0, Sat), sat_bdd(Sat, BDD), root_put_formula_bdd(Root, F0, BDD) ; true ). /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - Support for project_attributes/2. This is called by the toplevel as project_attributes(+QueryVars, +AttrVars) in order to project all remaining constraints onto QueryVars. All CLP(B) variables that do not occur in QueryVars or AttrVars need to be existentially quantified, so that they do not occur in residual goals. This is very easy to do in the case of CLP(B). - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */ project_attributes(QueryVars0, AttrVars) :- append(QueryVars0, AttrVars, QueryVars1), include(clpb_variable, QueryVars1, QueryVars), maplist(var_index_root, QueryVars, _, Roots0), sort(Roots0, Roots), maplist(remove_hidden_variables(QueryVars), Roots). clpb_variable(Var) :- var_index(Var, _). /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - All CLP(B) variables occurring in BDDs but not in query variables become existentially quantified. This must also be reflected in the formula. In addition, an attribute is attached to these variables to suppress superfluous sat(V=:=V) goals. - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */ remove_hidden_variables(QueryVars, Root) :- root_get_formula_bdd(Root, Formula, BDD0), maplist(put_visited, QueryVars), bdd_variables(BDD0, HiddenVars0), exclude(universal_var, HiddenVars0, HiddenVars), maplist(unvisit, QueryVars), foldl(existential, HiddenVars, BDD0, BDD), foldl(quantify_existantially, HiddenVars, Formula, ExFormula), root_put_formula_bdd(Root, ExFormula, BDD). quantify_existantially(E, E0, E^E0) :- put_attr(E, clpb_omit_boolean, true). /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - BDD restriction. - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */ bdd_restriction(Node, VI, Value, Res) :- empty_assoc(G0), phrase(bdd_restriction_(Node, VI, Value, Res), [G0], _), is_bdd(Res). bdd_restriction_(Node, VI, Value, Res) --> ( { integer(Node) } -> { Res = Node } ; { node_var_low_high(Node, Var, Low, High) } -> ( { integer(Var) } -> ( { Var =:= 0 } -> bdd_restriction_(Low, VI, Value, Res) ; { Var =:= 1 } -> bdd_restriction_(High, VI, Value, Res) ; { no_truth_value(Var) } ) ; { var_index(Var, I0), node_id(Node, ID) }, ( { I0 =:= VI } -> ( { Value =:= 0 } -> { Res = Low } ; { Value =:= 1 } -> { Res = High } ) ; { I0 > VI } -> { Res = Node } ; state(G0), { get_assoc(ID, G0, Res) } -> [] ; bdd_restriction_(Low, VI, Value, LRes), bdd_restriction_(High, VI, Value, HRes), make_node(Var, LRes, HRes, Res), state(G0, G), { put_assoc(ID, G0, Res, G) } ) ) ; { domain_error(node, Node) } ). /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - Relating a BDD to its elements (nodes and variables). Note that BDDs can become quite big (easily millions of nodes), and memory space is a major bottleneck for many problems. If possible, we therefore do not duplicate the entire BDD in memory (as in bdd_ites/2), but only extract its features as needed. - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */ bdd_nodes(BDD, Ns) :- bdd_nodes(ignore_node, BDD, Ns). ignore_node(_). % VPred is a unary predicate that is called for each node that has a % branching variable (= each inner node). bdd_nodes(VPred, BDD, Ns) :- phrase(bdd_nodes_(VPred, BDD), Ns), maplist(with_aux(unvisit), Ns). bdd_nodes_(VPred, Node) --> ( { node_visited(Node) } -> [] ; { call(VPred, Node), with_aux(put_visited, Node), node_var_low_high(Node, _, Low, High) }, [Node], bdd_nodes_(VPred, Low), bdd_nodes_(VPred, High) ). node_visited(Node) :- integer(Node). node_visited(Node) :- with_aux(is_visited, Node). bdd_variables(BDD, Vs) :- bdd_nodes(BDD, Nodes), nodes_variables(Nodes, Vs). nodes_variables(Nodes, Vs) :- phrase(nodes_variables_(Nodes), Vs), maplist(unvisit, Vs). nodes_variables_([]) --> []. nodes_variables_([Node|Nodes]) --> { node_var_low_high(Node, Var, _, _) }, ( { integer(Var) } -> [] ; { is_visited(Var) } -> [] ; { put_visited(Var) }, [Var] ), nodes_variables_(Nodes). unvisit(V) :- del_attr(V, clpb_visited). is_visited(V) :- get_attr(V, clpb_visited, true). put_visited(V) :- put_attr(V, clpb_visited, true). with_aux(Pred, Node) :- node_aux(Node, Aux), call(Pred, Aux). /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - Internal consistency checks. To enable these checks, set the flag clpb_validation to true. - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */ is_bdd(BDD) :- ( current_prolog_flag(clpb_validation, true) -> bdd_ites(BDD, ITEs), pairs_values(ITEs, Ls0), sort(Ls0, Ls1), ( same_length(Ls0, Ls1) -> true ; domain_error(reduced_ites, (ITEs,Ls0,Ls1)) ), ( member(ITE, ITEs), \+ registered_node(ITE) -> domain_error(registered_node, ITE) ; true ), ( member(I, ITEs), \+ ordered(I) -> domain_error(ordered_node, I) ; true ) ; true ). ordered(_-ite(Var,High,Low)) :- ( var_index(Var, VI) -> greater_varindex_than(High, VI), greater_varindex_than(Low, VI) ; true ). greater_varindex_than(Node, VI) :- ( integer(Node) -> true ; node_var_low_high(Node, Var, _, _), ( var_index(Var, OI) -> OI > VI ; true ) ). registered_node(Node-ite(Var,High,Low)) :- ( var(Var) -> low_high_key(Low, High, Key), lookup_node(Var, Key, Node0), Node == Node0 ; true ). bdd_ites(BDD, ITEs) :- bdd_nodes(BDD, Nodes), maplist(node_ite, Nodes, ITEs). node_ite(Node, Node-ite(Var,High,Low)) :- node_var_low_high(Node, Var, Low, High). %% labeling(+Vs) is multi. % % Enumerate concrete solutions. Assigns truth values to the Boolean % variables Vs such that all stated constraints are satisfied. labeling(Vs0) :- must_be(list, Vs0), maplist(labeling_var, Vs0), variables_in_index_order(Vs0, Vs), maplist(indomain, Vs). labeling_var(V) :- var(V), !. labeling_var(V) :- V == 0, !. labeling_var(V) :- V == 1, !. labeling_var(V) :- domain_error(clpb_variable, V). variables_in_index_order(Vs0, Vs) :- maplist(var_with_index, Vs0, IVs0), keysort(IVs0, IVs), pairs_values(IVs, Vs). var_with_index(V, I-V) :- ( var_index_root(V, I, _) -> true ; I = 0 ). indomain(0). indomain(1). %% sat_count(+Expr, -Count) is det. % % Count the number of admissible assignments. Count is the number of % different assignments of truth values to the variables in the % Boolean expression Expr, such that Expr is true and all posted % constraints are satisfiable. % % A common form of invocation is `sat_count(+[1|Vs], Count)`: This % counts the number of admissible assignments to `Vs` without imposing % any further constraints. % % Examples: % % == % ?- sat(A =< B), Vs = [A,B], sat_count(+[1|Vs], Count). % Vs = [A, B], % Count = 3, % sat(A=:=A*B). % % ?- length(Vs, 120), % sat_count(+Vs, CountOr), % sat_count(*(Vs), CountAnd). % Vs = [...], % CountOr = 1329227995784915872903807060280344575, % CountAnd = 1. % == sat_count(Sat0, N) :- catch((parse_sat(Sat0, Sat), sat_bdd(Sat, BDD), sat_roots(Sat, Roots), roots_and(Roots, _-BDD, _-BDD1), % we mark variables that occur in Sat0 as visited ... term_variables(Sat0, Vs), maplist(put_visited, Vs), % ... so that they do not appear in Vs1 ... bdd_variables(BDD1, Vs1), partition(universal_var, Vs1, Univs, Exis), % ... and then remove remaining variables: foldl(universal, Univs, BDD1, BDD2), foldl(existential, Exis, BDD2, BDD3), variables_in_index_order(Vs, IVs), foldl(renumber_variable, IVs, 1, VNum), bdd_count(BDD3, VNum, Count0), var_u(BDD3, VNum, P), % Do not unify N directly, because we are not prepared % for propagation here in case N is a CLP(B) variable. N0 is 2^(P - 1)*Count0, % reset all attributes and Aux variables throw(count(N0))), count(N0), N = N0). universal(V, BDD, Node) :- var_index(V, Index), bdd_restriction(BDD, Index, 0, NA), bdd_restriction(BDD, Index, 1, NB), apply(*, NA, NB, Node). renumber_variable(V, I0, I) :- put_attr(V, clpb, index_root(I0,_)), I is I0 + 1. bdd_count(Node, VNum, Count) :- ( integer(Node) -> Count = Node ; node_aux(Node, Count), ( integer(Count) -> true ; node_var_low_high(Node, V, Low, High), bdd_count(Low, VNum, LCount), bdd_count(High, VNum, HCount), bdd_pow(Low, V, VNum, LPow), bdd_pow(High, V, VNum, HPow), Count is LPow*LCount + HPow*HCount ) ). bdd_pow(Node, V, VNum, Pow) :- var_index(V, Index), var_u(Node, VNum, P), Pow is 2^(P - Index - 1). var_u(Node, VNum, Index) :- ( integer(Node) -> Index = VNum ; node_varindex(Node, Index) ). /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - Pick a solution in such a way that each solution is equally likely. - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */ %% random_labeling(+Seed, +Vs) is det. % % Select a single random solution. An admissible assignment of truth % values to the Boolean variables in Vs is chosen in such a way that % each admissible assignment is equally likely. Seed is an integer, % used as the initial seed for the random number generator. single_bdd(Vars0) :- maplist(monotonic_variable, Vars0, Vars), % capture all variables with a single BDD sat(+[1|Vars]). random_labeling(Seed, Vars) :- must_be(list, Vars), set_random(seed(Seed)), ( ground(Vars) -> true ; catch((single_bdd(Vars), once((member(Var, Vars),var(Var))), var_index_root(Var, _, Root), root_get_formula_bdd(Root, _, BDD), bdd_variables(BDD, Vs), variables_in_index_order(Vs, IVs), foldl(renumber_variable, IVs, 1, VNum), phrase(random_bindings(VNum, BDD), Bs), maplist(del_attrs, Vs), % reset all attribute modifications throw(randsol(Vars, Bs))), randsol(Vars, Bs), true), maplist(call, Bs), % set remaining variables to 0 or 1 with equal probability include(var, Vars, Remaining), maplist(maybe_zero, Remaining) ). maybe_zero(Var) :- ( maybe -> Var = 0 ; Var = 1 ). random_bindings(_, Node) --> { Node == 1 }, !. random_bindings(VNum, Node) --> { node_var_low_high(Node, Var, Low, High), bdd_count(Node, VNum, Total), bdd_count(Low, VNum, LCount) }, ( { maybe(LCount, Total) } -> [Var=0], random_bindings(VNum, Low) ; [Var=1], random_bindings(VNum, High) ). /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - Find solutions with maximum weight. - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */ %% weighted_maximum(+Weights, +Vs, -Maximum) is multi. % % Enumerate weighted optima over admissible assignments. Maximize a % linear objective function over Boolean variables Vs with integer % coefficients Weights. This predicate assigns 0 and 1 to the % variables in Vs such that all stated constraints are satisfied, and % Maximum is the maximum of sum(Weight_i*V_i) over all admissible % assignments. On backtracking, all admissible assignments that % attain the optimum are generated. % % This predicate can also be used to _minimize_ a linear Boolean % program, since negative integers can appear in Weights. % % Example: % % == % ?- sat(A#B), weighted_maximum([1,2,1], [A,B,C], Maximum). % A = 0, B = 1, C = 1, Maximum = 3. % == weighted_maximum(Ws, Vars, Max) :- must_be(list(integer), Ws), must_be(list(var), Vars), single_bdd(Vars), Vars = [Var|_], var_index_root(Var, _, Root), root_get_formula_bdd(Root, _, BDD0), bdd_variables(BDD0, Vs), % existentially quantify variables that are not considered maplist(put_visited, Vars), exclude(is_visited, Vs, Unvisited), maplist(unvisit, Vars), foldl(existential, Unvisited, BDD0, BDD), maplist(var_with_index, Vars, IVs), pairs_keys_values(Pairs0, IVs, Ws), keysort(Pairs0, Pairs1), pairs_keys_values(Pairs1, IVs1, WeightsIndexOrder), pairs_values(IVs1, VarsIndexOrder), % Pairs is a list of Var-Weight terms, in index order of Vars pairs_keys_values(Pairs, VarsIndexOrder, WeightsIndexOrder), bdd_maximum(BDD, Pairs, Max), max_labeling(BDD, Pairs). max_labeling(1, Pairs) :- max_upto(Pairs, _, _). max_labeling(node(_,Var,Low,High,Aux), Pairs0) :- max_upto(Pairs0, Var, Pairs), get_attr(Aux, clpb_max, max(_,Dir)), direction_labeling(Dir, Var, Low, High, Pairs). max_upto([], _, _). max_upto([Var0-Weight|VWs0], Var, VWs) :- ( Var == Var0 -> VWs = VWs0 ; Weight =:= 0 -> ( Var0 = 0 ; Var0 = 1 ), max_upto(VWs0, Var, VWs) ; Weight < 0 -> Var0 = 0, max_upto(VWs0, Var, VWs) ; Var0 = 1, max_upto(VWs0, Var, VWs) ). direction_labeling(low, 0, Low, _, Pairs) :- max_labeling(Low, Pairs). direction_labeling(high, 1, _, High, Pairs) :- max_labeling(High, Pairs). bdd_maximum(1, Pairs, Max) :- pairs_values(Pairs, Weights0), include(<(0), Weights0, Weights), sum_list(Weights, Max). bdd_maximum(node(_,Var,Low,High,Aux), Pairs0, Max) :- ( get_attr(Aux, clpb_max, max(Max,_)) -> true ; ( skip_to_var(Var, Weight, Pairs0, Pairs), ( Low == 0 -> bdd_maximum_(High, Pairs, MaxHigh, MaxToHigh), Max is MaxToHigh + MaxHigh + Weight, Dir = high ; High == 0 -> bdd_maximum_(Low, Pairs, MaxLow, MaxToLow), Max is MaxToLow + MaxLow, Dir = low ; bdd_maximum_(Low, Pairs, MaxLow, MaxToLow), bdd_maximum_(High, Pairs, MaxHigh, MaxToHigh), Max0 is MaxToLow + MaxLow, Max1 is MaxToHigh + MaxHigh + Weight, Max is max(Max0,Max1), ( Max0 =:= Max1 -> Dir = _Any ; Max0 < Max1 -> Dir = high ; Dir = low ) ), store_maximum(Aux, Max, Dir) ) ). bdd_maximum_(Node, Pairs, Max, MaxTo) :- bdd_maximum(Node, Pairs, Max), between_weights(Node, Pairs, MaxTo). store_maximum(Aux, Max, Dir) :- put_attr(Aux, clpb_max, max(Max,Dir)). between_weights(Node, Pairs0, MaxTo) :- ( Node == 1 -> MaxTo = 0 ; node_var_low_high(Node, Var, _, _), phrase(skip_to_var_(Var, _, Pairs0, _), Weights0), include(<(0), Weights0, Weights), sum_list(Weights, MaxTo) ). skip_to_var(Var, Weight, Pairs0, Pairs) :- phrase(skip_to_var_(Var, Weight, Pairs0, Pairs), _). skip_to_var_(Var, Weight, [Var0-Weight0|VWs0], VWs) --> ( { Var == Var0 } -> { Weight = Weight0, VWs0 = VWs } ; ( { Weight0 =< 0 } -> [] ; [Weight0] ), skip_to_var_(Var, Weight, VWs0, VWs) ). /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - Projection to residual goals. - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */ attribute_goals(Var) --> { var_index_root(Var, _, Root) }, ( { root_get_formula_bdd(Root, Formula, BDD) } -> { del_bdd(Root) }, ( { current_prolog_flag(clpb_residuals, bdd) } -> { bdd_nodes(BDD, Nodes), phrase(nodes(Nodes), Ns) }, [clpb:'$clpb_bdd'(Ns)] ; { prepare_global_variables(BDD), phrase(sat_ands(Formula), Ands0), ands_fusion(Ands0, Ands), maplist(formula_anf, Ands, ANFs0), sort(ANFs0, ANFs1), exclude(eq_1, ANFs1, ANFs2), variables_separation(ANFs2, ANFs) }, sats(ANFs) ), ( { get_attr(Var, clpb_atom, Atom) } -> [clpb:sat(Var=:=Atom)] ; [] ), % formula variables not occurring in the BDD should be booleans { bdd_variables(BDD, Vs), maplist(del_clpb, Vs), term_variables(Formula, RestVs0), include(clpb_variable, RestVs0, RestVs) }, booleans(RestVs) ; boolean(Var) % the variable may have occurred only in taut/2 ). del_clpb(Var) :- del_attr(Var, clpb). /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - To make residual projection work with recorded constraints, the global counters must be adjusted so that new variables and nodes also get new IDs. Also, clpb_next_id/2 is used to actually create these counters, because creating them with b_setval/2 would make them [] on backtracking, which is quite unfortunate in itself. - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */ prepare_global_variables(BDD) :- clpb_next_id('$clpb_next_var', V0), clpb_next_id('$clpb_next_node', N0), bdd_nodes(BDD, Nodes), foldl(max_variable_node, Nodes, V0-N0, MaxV0-MaxN0), MaxV is MaxV0 + 1, MaxN is MaxN0 + 1, b_setval('$clpb_next_var', MaxV), b_setval('$clpb_next_node', MaxN). max_variable_node(Node, V0-N0, V-N) :- node_id(Node, N1), node_varindex(Node, V1), N is max(N0,N1), V is max(V0,V1). /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - Fuse formulas that share the same variables into single conjunctions. - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */ ands_fusion(Ands0, Ands) :- maplist(with_variables, Ands0, Pairs0), keysort(Pairs0, Pairs), group_pairs_by_key(Pairs, Groups), pairs_values(Groups, Andss), maplist(list_to_conjunction, Andss, Ands). with_variables(F, Vs-F) :- term_variables(F, Vs0), variables_in_index_order(Vs0, Vs). /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - If possible, separate variables into different sat/1 goals. A formula F can be split in two if for two of its variables A and B, taut((A^F)*(B^F) =:= F, 1) holds. In the first conjunct, A does not occur, and in the second, B does not occur. We separate variables until that is no longer possible. There may be a better way to do this. - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */ variables_separation(Fs0, Fs) :- separation_fixpoint(Fs0, [], Fs). separation_fixpoint(Fs0, Ds0, Fs) :- phrase(variables_separation_(Fs0, Ds0, Rest), Fs1), partition(anf_done, Fs1, Ds1, Fs2), maplist(arg(1), Ds1, Ds2), maplist(arg(1), Fs2, Fs3), append(Ds0, Ds2, Ds3), append(Rest, Fs3, Fs4), sort(Fs4, Fs5), sort(Ds3, Ds4), ( Fs5 == [] -> Fs = Ds4 ; separation_fixpoint(Fs5, Ds4, Fs) ). anf_done(done(_)). variables_separation_([], _, []) --> []. variables_separation_([F0|Fs0], Ds, Rest) --> ( { member(Done, Ds), F0 == Done } -> variables_separation_(Fs0, Ds, Rest) ; { sat_rewrite(F0, F), sat_bdd(F, BDD), bdd_variables(BDD, Vs0), exclude(universal_var, Vs0, Vs), maplist(existential_(BDD), Vs, Nodes), phrase(pairs(Nodes), Pairs), group_pairs_by_key(Pairs, Groups), phrase(groups_separation(Groups, BDD), ANFs) }, ( { ANFs = [_|_] } -> list(ANFs), { Rest = Fs0 } ; [done(F0)], variables_separation_(Fs0, Ds, Rest) ) ). existential_(BDD, V, Node) :- existential(V, BDD, Node). groups_separation([], _) --> []. groups_separation([BDD1-BDDs|Groups], OrigBDD) --> { phrase(separate_pairs(BDDs, BDD1, OrigBDD), Nodes) }, ( { Nodes = [_|_] } -> nodes_anfs([BDD1|Nodes]) ; [] ), groups_separation(Groups, OrigBDD). separate_pairs([], _, _) --> []. separate_pairs([BDD2|Ps], BDD1, OrigBDD) --> ( { apply(*, BDD1, BDD2, And), And == OrigBDD } -> [BDD2] ; [] ), separate_pairs(Ps, BDD1, OrigBDD). nodes_anfs([]) --> []. nodes_anfs([N|Ns]) --> { node_anf(N, ANF) }, [anf(ANF)], nodes_anfs(Ns). pairs([]) --> []. pairs([V|Vs]) --> pairs_(Vs, V), pairs(Vs). pairs_([], _) --> []. pairs_([B|Bs], A) --> [A-B], pairs_(Bs, A). /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - Set the Prolog flag clpb_residuals to bdd to obtain the BDD nodes as residuals. Note that they cannot be used as regular goals. - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */ nodes([]) --> []. nodes([Node|Nodes]) --> { node_var_low_high(Node, Var0, Low, High), var_or_atom(Var0, Var), maplist(node_projection, [Node,High,Low], [ID,HID,LID]), var_index(Var0, VI) }, [ID-(v(Var,VI) -> HID ; LID)], nodes(Nodes). node_projection(Node, Projection) :- node_id(Node, ID), ( integer(ID) -> Projection = node(ID) ; Projection = ID ). /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - By default, residual goals are sat/1 calls of the remaining formulas, using (mostly) algebraic normal form. - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */ sats([]) --> []. sats([A|As]) --> [clpb:sat(A)], sats(As). booleans([]) --> []. booleans([B|Bs]) --> boolean(B), { del_clpb(B) }, booleans(Bs). boolean(Var) --> ( { get_attr(Var, clpb_omit_boolean, true) } -> [] ; [clpb:sat(Var =:= Var)] ). /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - Relate a formula to its algebraic normal form (ANF). - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */ formula_anf(Formula0, ANF) :- parse_sat(Formula0, Formula), sat_bdd(Formula, Node), node_anf(Node, ANF). node_anf(Node, ANF) :- node_xors(Node, Xors0), maplist(maplist(monotonic_variable), Xors0, Xors), maplist(list_to_conjunction, Xors, Conjs), ( Conjs = [Var,C|Rest], clpb_var(Var) -> foldl(xor, Rest, C, RANF), ANF = (Var =\= RANF) ; Conjs = [One,Var,C|Rest], One == 1, clpb_var(Var) -> foldl(xor, Rest, C, RANF), ANF = (Var =:= RANF) ; Conjs = [C|Cs], foldl(xor, Cs, C, ANF) ). monotonic_variable(Var0, Var) :- ( var(Var0), current_prolog_flag(clpb_monotonic, true) -> Var = v(Var0) ; Var = Var0 ). clpb_var(Var) :- var(Var), !. clpb_var(v(_)). list_to_conjunction([], 1). list_to_conjunction([L|Ls], Conj) :- foldl(and, Ls, L, Conj). xor(A, B, B # A). eq_1(V) :- V == 1. node_xors(Node, Xors) :- phrase(xors(Node), Xors0), % we remove elements that occur an even number of times (A#A --> 0) maplist(sort, Xors0, Xors1), pairs_keys_values(Pairs0, Xors1, _), keysort(Pairs0, Pairs), group_pairs_by_key(Pairs, Groups), exclude(even_occurrences, Groups, Odds), pairs_keys(Odds, Xors2), maplist(exclude(eq_1), Xors2, Xors). even_occurrences(_-Ls) :- length(Ls, L), L mod 2 =:= 0. xors(Node) --> ( { Node == 0 } -> [] ; { Node == 1 } -> [[1]] ; { node_var_low_high(Node, Var0, Low, High), var_or_atom(Var0, Var), node_xors(Low, Ls0), node_xors(High, Hs0), maplist(with_var(Var), Ls0, Ls), maplist(with_var(Var), Hs0, Hs) }, list(Ls0), list(Ls), list(Hs) ). list([]) --> []. list([L|Ls]) --> [L], list(Ls). with_var(Var, Ls, [Var|Ls]). /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - Global variables for unique node and variable IDs and atoms. - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */ make_clpb_var('$clpb_next_var') :- nb_setval('$clpb_next_var', 0). make_clpb_var('$clpb_next_node') :- nb_setval('$clpb_next_node', 0). make_clpb_var('$clpb_atoms') :- empty_assoc(E), nb_setval('$clpb_atoms', E). :- multifile user:exception/3. user:exception(undefined_global_variable, Name, retry) :- make_clpb_var(Name), !. clpb_next_id(Var, ID) :- b_getval(Var, ID), Next is ID + 1, b_setval(Var, Next). clpb_atom_var(Atom, Var) :- b_getval('$clpb_atoms', A0), ( get_assoc(Atom, A0, Var) -> true ; put_attr(Var, clpb_atom, Atom), put_attr(Var, clpb_omit_boolean, true), put_assoc(Atom, A0, Var, A), b_setval('$clpb_atoms', A) ). /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - The variable attributes below are not used as constraints by this library. Project remaining attributes to empty lists of residuals. Because accessing these hooks is basically a cross-module call, we must declare them public. - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */ :- public clpb_hash:attr_unify_hook/2, clpb_bdd:attribute_goals//1, clpb_hash:attribute_goals//1, clpb_omit_boolean:attr_unify_hook/2, clpb_omit_boolean:attribute_goals//1, clpb_atom:attr_unify_hook/2, clpb_atom:attribute_goals//1. clpb_hash:attr_unify_hook(_,_). % this unification is always admissible /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - If a universally quantified variable is unified to a Boolean value, it indicates that the formula does not hold for the other value, so it is false. - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */ clpb_atom:attr_unify_hook(_, _) :- false. clpb_omit_boolean:attr_unify_hook(_,_). clpb_bdd:attribute_goals(_) --> []. clpb_hash:attribute_goals(_) --> []. clpb_omit_boolean:attribute_goals(_) --> []. clpb_atom:attribute_goals(_) --> []. % clpb_hash:attribute_goals(Var) --> % { get_attr(Var, clpb_hash, Assoc), % assoc_to_list(Assoc, List0), % maplist(node_portray, List0, List) }, [Var-List]. % node_portray(Key-Node, Key-Node-ite(Var,High,Low)) :- % node_var_low_high(Node, Var, Low, High). /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - Messages - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */ :- multifile prolog:message//1. prolog:message(clpb(bounded)) --> ['Using CLP(B) with bounded arithmetic may yield wrong results.'-[]]. warn_if_bounded_arithmetic :- ( current_prolog_flag(bounded, true) -> print_message(warning, clpb(bounded)) ; true ). :- initialization(warn_if_bounded_arithmetic). /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - Sanbox declarations - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */ :- multifile sandbox:safe_global_variable/1, sandbox:safe_primitive/1. sandbox:safe_global_variable('$clpb_next_var'). sandbox:safe_global_variable('$clpb_next_node'). sandbox:safe_global_variable('$clpb_atoms'). sandbox:safe_primitive(set_prolog_flag(clpb_residuals, _)).