The Course Logic Programming ID2213
-
Upload
tad-delacruz -
Category
Documents
-
view
14 -
download
1
description
Transcript of The Course Logic Programming ID2213
The Course Logic Programming ID2213
Thomas Sjöland [email protected]
SCS, Software and Computer Systems
ICT - School of Information and Communication Technology
KTH, The Royal Institute of Technology
Outline of lecturesW35:F1: Theory, Elementary Programs, unification Theory, Model Theory of LP, proof trees and search trees
W36:F2: Programming Style, recursion, equality primitives, representation Advanced Recursive Techniques, accumulators, diff-structures, ADT
W37:F3: Search Based Programming, cut and negation Concurrency, Search Based Programming, state space, puzzles, games
W38:F4: Logic programming and Grammars, parsing with DCG
W39:F5: Program Transformation. Higher-order programming. Metaprogramming, Expert Systems
W40:F6: Case study: A compiler for a simple stack machine
W41:F7: Case study: Support for reasoning about electronic circuits Red1: Project presentation 4 hours
W42: Written Examination
F1: Theory and simple programs
Sterling and Shapiro ch. 1,2,4,5,6Nilsson and Maluszynski ch.1,2,3,6
Theory for Logic Programming
Outline
Informal introduction to logic programming theoryData in logic programs: Individual constants, term
constructors, logical variables, compound terms, trees and lists
Equality theory, UnificationLogic Programs: Definite (Horn) ClausesModel theory (least Herbrand model, term interpretation)Proof theory and operational semantics of Prolog
(SLD-resolution, proof trees)Simple databases
Recursive rules
Logic Programming
Using proofs to compute To each proof you can order a computation To each computation you can order a proof
Representation of knowledge and computations - algorithms - functions - relations
Data in Logic Programs
Programs express facts about a world of objects Constants
Functors
NumbersCompounded structures (normally finite)
Lists
Trees
Objects in Logic ProgramsIndividual constants
a b foo 4711 -37 34.5 Functors
structure names of trees and graphssame syntax as non-numerical constants
Arity (number of arguments): term/4, a/0
Syntax example: term(a,b,-4711,other(b,a))
Logical Variables - Syntax
Syntax: begin with a capital letter (or '_')
X Y Z Foo _Bar _
Variables can occur wherever constants or structures occur
Range over logical objects
_ is "anonymous" or "void"
Programs are Theories
sets of relations (predicates) over objects
The classical form of a definition is as a clausal form where a positive literal P has exactly one occurrence:
P or not Q1 or ... or not Qn
This can be written as P if Q1 & ... & ... Qn.
If all goals Qi are true the clause becomes P.
Program = Definitions + Query
The general form of a relation definition is
P if (Q11 & ... & Q1n) or ...or (Qm1 & ... & Qmn).
1..m and 1..n are index sets large enough to cover all goal atoms, Qij
Program = Definitions + Query
Elementary literals (atoms)true, false, X=Ycannot be redefined (only used in queries and definition bodies)
Defined literals (p above)
Definite Clauses: Facts
Facts: statements of form P :- true.Also written simply as P.Example: brother(nils,karl).Means that the binary relation brother holds between individual constants nils and karl.
Definite Clauses: Rules
Rules: conditional formulae of the form
P :- Q1,....,Qn.
P is called the head and Q1,...,Qn the body of the clause and P, Q1,...,Qn are atomic formulas (relation symbols). Some of the Qi may be predefined relation symbols (=, <)
":-" is read as "if", "," is read as "and"
Definite Clauses: Rules, example
Example of a rule:
grandfather(X,Y) :- father(X,Z), father(Z,Y).
The binary relation grandfather holds between two individuals represented by variables X and Y if the relation father holds between X and some Z and between that Z and Y.
Clause Syntax
Example : p(17).p(X) :- X<8, q(X).p(X) :- q(X), X=s(Y), p(Y).
In english the above example could be stated as follows:- The property p holds for the constant 17. - The property p holds for a term denoted by the
variable X if X<8 and q holds for X. - The property p holds for X if q holds for X, X
equals a term s(Y) and p holds for Y.
Programs are Theories
Definitions are collections of facts and rules - sets of relations (predicates) over the objects
e.g. (for predicate p/2 using q/2 and r/2)
p(foo,bar).p(Foo,Bar) :- q(Foo,Baz), r(Baz,Bar).
Functions are special cases of relations (deterministic)
Query, Goal
formula to be verified (or falsified) Questions posed to the system are of the form
?- Q1,...,Qn.
for example ?- q(Foo,Baz), r(Baz,Bar).
If the system succeeds to prove the formula, the values of the variables (the bindings) that are the result are shown, otherwise the proof attempt has failed or possibly loops.
Note that more than one solution is possible.
How Prolog works A user query ?- p(Args). is proven using resolution - look for all definition clauses of p - pick one, save others as alternatives - match the arguments in Args with the
terms in the head of the clause, create necessary variable bindings
- if the matching unification fails, try next alternative
- else prove the goal clauses of the body from left to right
- if all proofs are done, the bindings are presented
Database for family relationships
parent(Parent, Child), male(Person) and female(Person)
parent(erik, jonas). male(erik).parent(erik, eva). male(jonas).parent(lena, jonas). female(lena).
?- parent(lena, jonas).
Yes
?-parent(X,Y).X=erik, Y=jonas; X=erik, Y=eva; X= lena, Y= jonas
?- parent(X, jonas).X=erik; X=lena
Logical variables - semanticsVariables can occur wherever constants or
structures occur.
Range over logical objects.
Bound to structures and to other variables.
The semantics is "single-assignment"- starts "unbound"
- once bound, stays the same in the whole proof
Example cont.: rules
father(Dad, Child):- parent(Dad, Child), male(Dad).mother(Mum, Child):- parent(Mum, Child), female(Mum).
?- father(X,Y). ?- mother(erik, jonas).X=erik, Y=jonas NoX=erik, Y=eva ?- mother(Erik, jonas). Yes. Why?
%sibling(Person1, Person2) :- ...sibling(X,Y) :- parent(Z,X), parent(Z,Y).
%cousin(Person1, Person2):- ...cousin(X,Y) :- parent(Z,X), parent(U,Y), sibling(Z,U).
syntactic sugar ';'
The symbol ';' can be used to avoid defining auxiliary predicates or to reduce the number of clauses.
';' is read as "or".
A clause of the formP :- Q1, (Q2 ; Q3), Q4.is the same as P :- Q1, Q, Q4.Q :- Q2.Q :- Q3.
Equality theory - Substitutions
X equals Y iffX is an unbound variable or Y is an unbound variableorX and Y are (bound to) the same constantorX and Y are terms with the same functor and aritye.g. X is term(Z1,..,Zn) and Y is term(U1,...,Un)and for all arguments 1=<i=<n: Zi equals Ui.
Substitutions
A substitution is a function Subst: Var -> Term
Substitutions can be applied to terms or substitutions and also to formulas
We may represent a substitution as a conjunction of simple equalities v=t where a variable v occurs on the left hand side at most once
or as a set {v/t | v=t} meaning a function thatreplaces v with t for each v/t in the set
Unifier
A unifier is a substitution suchthatst
(applying tosand totcreates identical terms)
Most general unifier
A unifier is more general than a unifier iff there exists another unifier such that
A unifier is the most general unifier of two termsiff is more general than any other unifier of the two
terms
Most general unifier, example
Example :
t(X,Y,Z) and t(U,V,W)
are unified by{X/a,Y/b,Z/c, U/a,V/b,W/c}
consider for instance the mgus in this case{X/U,Y/V,Z/W} and {U/X,V/Y,W/Z}
Unification procedure
An algorithm that constructs most general unifiers for two terms in an environment is a unification procedure.
Since the most general unifier is unique (modulo renaming of variables), unification can also be understood as a function
unify : Subst x Term x Term -> Subst
Theory and simple programs (cont) Operational Semantics, SLD
Sterling and Shapiro ch. 1,2,4,5,6Nilsson and Maluszynski ch.1,2,3,6
Example cont.: structured data
Use compound (not atomic) terms for the description of persons. parent(erik, jonas).
parent(erik, eva).
parent(lena, jonas).
male(person(erik,50,180cm)).
male(person(jonas,25,_)).
father(Dad,Child) :- Dad = person(DadName,_,_), Child=person(ChildName,_,_), parent(DadName, ChildName),
male(Dad).
?-father(person(_,50,_), person(X,_,_)). X=jonas (second solution: X = eva)
NB: how does the unification algorithm work here?
Logical Variables in ProgramsVariables and parameters are implicitly quantified
syntax: variables start with capital letterp(X,Y) :- q(X,Z), r(Z,Y).
is understood as forallX,Y:(p(X,Y) <-
exists Z:(q(X,Z), r(Z,Y)))
Parameters (X,Y) are often confusingly named "global variables" as opposed to "local variables" (Z)
but if X is global and Y is local, what is Y, if X=Y occurs in program?
Example cont.: recursive rules
ancestor(Ancestor, Descendant) :- parent(Ancestor, Descendant).
ancestor(Ancestor, Descendant) :- parent(Ancestor, Person), ancestor(Person, Descendant).
parent(ulf, erik). …
?- ancestor(X, Y).
Declarative vs procedural
A logic program can be understood in either of two ways:
it can be seen as a set of Horn clauses specifying facts about
data (a theory). This is the declarative or model-
theoretical reading of a logic program. What?
it can be viewed as a program describing a particular
execution (how to find a proof). This is the procedural
or proof-theoretical reading of a logic program. How?
Modus Ponens
P Q:-P --------------
Q
Proof methods with Horn clauses
Given a database: p :- q,r. q :- q1, q2. q1. q2. r.
Proof methods to prove p:
Forward chaining - use modus ponens to
accumulate known truths, starting from facts.
Backward chaining - prove p by proving q
and then proving r etc. (used in prolog)
Model Theory: Herbrand interpretation
When reading a program as a specification we need to determine the meaning of the symbols.
A term interpretation, or "Herbrand interpretation" is an association of a unique function to each functor occurring in the program and an association of sets of tuples of terms to relations.
An interpretation is a model for a program if all statements in the interpretation are true.
Model Theory: Least Herbrand Model
The least Herbrand model is the least term interpretation such that it is a model.
For definite clauses such a unique model always exists.
Least Herbrand Model computed
The model can be inductively built up from the relation symbols and the terms built from constants and terms in the program by constructing a fixpoint.
Use the monotone Tp-operator. (N&M p. 29 ch 2.4), ground(P) is the set of all ground instances of clauses in a program P (assume always at least one functor or constant and only finite structures).
Tp(I) := {A0 | A0:-A1,...,Am in ground(P) & {A1,...,Am} subset I }.
Start from the empty theory and determine the least fixpoint for I=Tp(I) U I.
Note that the model does not contain variables.
Constructing a model with Tp
p :- q,r. q :- q1, q2. q1. q2. r. s.
0: {} 1: {q1,q2,r,s} 2: {q1,q2,r,s,q} 3: {q1,q2,r,s,q,p} 4: {q1,q2,r,s,q,p} Done
The fixpoint is the model {q1,q2,r,s,q,p}
Constructing a model with Tp
p(X) :- q(X,Y),r(X). q(X,Y) :- q1(X,Y), q2(Y,X). q1(a,b). q2(b,a). r(a).
Herbrand universe: {a,b}
0: {} 1: {q1(a,b),q2(b,a),r(a)} 2: {q1(a,b),q2(b,a),r(a),q(a,b)} 3: {q1(a,b),q2(b,a),r(a),q(a,b),p(a)}
4: {q1(a,b),q2(b,a),r(a),q(a,b),p(a)} Done
The fixpoint is the model: {q1(a,b),q2(b,a),r(a),q(a,b),p(a)}
Infinite structures
Assuming that the least Herbrand Model defines the intended meaning of the program, unification must preserve the property that infinite (cyclic) terms are not allowed. This requires an occurs-check in the unification algorithm prohibiting for example X=f(X) from generating
X=f(f(f(f(f(f(f(........This is very inefficient so occurs-check is the responsibility
of the programmer. In critical cases a special test must be performed after the unification.
Note that SICStus Prolog uses rational trees in X=f(X)
Theoretically sound unification: unify_with_occurs_check/2
Proof theory:Execution is search for a proof or failure, generating an or-tree
restrictions on the variables are shown as bindings of the variables
Search trees and proof trees
proof
Search tree Proof tree
SLD-resolution rule
<- A1,..,Ai-1,Ai,Ai+1,...,Am B0 <- B1,...,Bn ------------------------------------------<- (A1,...,Ai-1,B1,...,Bn,Ai+1,...,Am)
Where P is a program,A1,...,Am are atomic formulas (goals),B0<- B1,...,Bn is a (renamed) clause in Pand=mgu(Ai,B0)
Goal and clause selection
A goal selection function specifies which goal Ai is selected by the SLD-rule.
The order in which clauses are chosen is determined with a clause selection rule.
Soundness of SLD-resolution
Any query (goal) that is provable with SLD-resolution is a logical consequence of the program.
Completeness of SLD-resolution
Any query (goal) that is (true) in the least Herbrand model is provable with SLD-resolution.In the case of an infinite SLD-tree, the selection
function has to be fair (as in breadth first search). For finite SLD-trees left-first-with-backtracking as used in Prolog gives a complete method.
Conclusion
LP can be used as a uniform language for representing databases, e.g. data structure and queries can be written in a single language
LP extends traditional databases by having recursive rules and structured data facilities
F2: Logic Programming Style
Sterling and Shapiro ch. 2,6,7,13 (except 2.4, 2.5, 3.6, 6.3, 7.6, 13.3, 13.4)Nilsson and Maluszynski ch.7 (except 7.3)
Outline
Programming techniquesArithmetic in PrologDifferent primitives for equality: =/2, ==/2
Recursive definitionsProcedural - declarativeImperative - logical style
binary trees, listsappend/3 reverse/2 quicksort/2
Specifying the use of a procedure
For serious projects it is good programming practice to specify the intended use of important procedures, such as the predicates intended to use in a library.
For instance this could be given as a comment of the following form:
% procedure foo(T1,T2,...Tn)%% Types: T1: type 1% T2: type 2% T3: type 3% ...% Tn: type n% Relation scheme:...% Modes uf use: (input arguments T1,T2) (output arguments T3,...,Tn)% Multiplicities of solution: deterministic (one solution only)
Built-in arithmetics
is/2 a built-in predicate for evaluation of arithmetical expressions
?- Value is Expression. - first, Expression is evaluated and, second, unified with Value
For example,
?- X = 2, Y is 1+X*3. - Y = 7?- X = 2, 4 is X*X. - yes?- Z is 1+x. - instantiation error, x is a constant?- Z is 1+X. - instantiation error, X is not instantiated?- 2 is 1+X. - instantiation error, X is not instantiated?- X=1+2. - Yes. X = 1+2.
Built-in arithmetics
is/2 evaluates expressions containing:
+ - * / // mod
- plus, minus, multiplication, division, integer division, remainder/\ \/ # \ << >>
- bitwise conjunction, disjunction, exclusive or, negation, shift to the left, shift to the right
abs(X), min(X), max(X), sin(X), cos(X), sqrt(X).
(for a complete list, see the SICStus manual)
Typical error: failing to unify floating point numbers.
NB: different "equals"= - unification= = ( \== ) - equality (inequality) of terms =:= ( =\= ) - arithmetic, boolean (not)equal is/2 - evaluation and unification
?- X=2, X=Y. - Yes. X=2, Y=2.?- X=2, X==Y. - No.?- X=:=2. - instantiation error ?- X=2, Y=2, X=:=Y. - Yes. X=2, Y=2.?- X=2+3, Y is X. - Yes. X=‘+’(2,3), Y=5.
Elementary programs
Sterling and Shapiro ch. 1,2,4,5,6Nilsson and Maluszynski ch.1,2,3,6
Composing recursive programs
think about declarative meaning of recursive data type (a definition)
write down recursive clause and base clauserun simple examples - check different goalscheck what is happening (do you get the
expected result?)
Composing recursive programs
Typical errors: missing (or erroneously failing) base
case error in data structure representationwrong arity of structuresmixing an element and a listpermuted arguments
Natural numbers
Unary syntaxFor example,
0 - denotes zeros(0) - denotes 1
... s(…s(s(0))…) - denotes n
Defining the natural numbers
natural_number(0).natural_number(s(X)) :-
natural_number(X).
Natural numbers
plusplus(0, X, X) :- natural_number(X).plus(s(X), Y, s(Z)) :- plus(X, Y, Z).
?- plus(s(0),0,s(0)). - checks 1+0=1Yes.
?- plus(X,s(0),s(s(0)). - checks X+1=2, (e.g. compute X=2-1)X=s(0).
?- plus(X, Y, s(s(0))). - checks X+Y=2, (e.g. generate all pairs of natural numbers, whose sum equals 2)
X=0, Y=s(s(0));
X=s(0), Y=s(0);
X=s(s(0)), Y=0.
Natural numbersless or equal
le(0, X) :- natural_number(X).le(s(X), s(Z)) :- le(X, Z).
multiplication
times(0, X, 0) :- natural_number(X).times(s(X), Y, Z) :-
plus(Y, Z1, Z), times(X, Y, Z1).
check how substitution works!
Recursive Arithmetic
sum([],0).sum([H|T],S) :- sum(T,V), S is H+V.
sum0([],0).sum0([H|T],S) :- sum0(T,V), S=H+V.
Binary treesSyntax (not built-in, create own compound terms)
For example, void - denotes empty tree tree(Element, Left, Right) - denotes a tree, where Element is
root and Left, Right are subtrees
tree(5,tree(8,void,void),tree(9,void,tree(3,void,void)))
defining a tree
binary_tree(void).binary_tree(tree(Element, Left, Right)) :-
binary_tree(Left), binary_tree(Right).
Binary trees
membershiptree_member(X,tree(X,_,_)).tree_member(X,tree(Y,Left ,_)):- tree_member(X,Left). tree_member(X,tree(Y,_,Right)):- tree_member(X,Right).
NB: X might be equal to Y in clauses 2 and 3!
Lists
Syntax
[Head|Tail] cons cell
Head is an element, Tail is a list '.'(Head,Tail)
[] empty list
simpler syntax[a | [] ] = [a] [a | [ b | [] ] ] = [a, b] [erik], [person(erik,_,_),jonas|[lena, eva]]
defining a listlist([]). - defines the basislist([X|Xs]) :- list(Xs). - defines the recursion
Lists
checking membershipmember(X, [X|Xs]).member(X, [Y|Ys]) :- member(X, Ys).
?- member(a, [b,c,a,d]). - checks membership
?- member(X, [b,c,a,d]). - takes an element from a list
?- member(b, Z). - generates a list containing b
Listsconcatenation of lists
append([], Xs, Xs).append([X|Xs], Y, [X|Zs]) :- append(Xs, Y, Zs).
?- append([a,b], [c], X). - addition of two lists
?- append(Xs, [a,d], [b,c,a,d]).
- finds a difference between lists
?- append(Xs, Ys, [a,b,c,d]).
- divides a list into two lists
Check SLD-tree!
Typical error: wrong "assembly" of a resulting list
Lists
reversing listsreverse([], []).reverse([H|T],R) :- reverse(T,S), append(S,[H],R).
?- reverse([a,b,c,d],R). - gives R=[d,c,b,a]
Check SLD-tree!
Typical error: wrong "assembly" of resulting listwrong_reverse([H|T],R):-
reverse(T,S), append(S,H,R).
Lists
sortingquicksort([X|Xs], Ys) :-
partition(Xs, X, Littles, Bigs),quicksort(Littles, Ls),quicksort(Bigs, Bs),append(Ls, [X|Bs], Ys).
quicksort([], []).
partition([Y|Ys], X, [Y|Ls], Bs) :- X>Y, partition(X, Ys, Ls, Bs).
partition([Y|Ys], X, Ls, [Y|Bs]) :- X=<Y, partition(X, Ys, Ls, Bs).
partition([], _, [], []).
DictionariesFinding and adding a value in a dictionary (an (ordered) binary tree) %lookup(+,?,?)
lookup(Key, tree(Key,Value, Left, Right), Value):- !.
lookup(Key, tree(Key1, Value1, Left, Right), Value) :- Key < Key1, lookup(Key, Left,
Value).
lookup(Key, tree(Key1, Value1, Left, Right), Value) :- Key > Key1, lookup(Key, Right,
Value).
?- lookup(1, D, fifi),lookup(2, D, mumu),lookup(1,D, X).D=tree(1, fifi, _C, tree(2, mumu, _B, _A)), X=fifi.
NB: for finding a key of a value, the traversal of tree should be implemented.
Composing recursive programsFinal exampleDefine a predicate unsort_deg(Xs, D) that, given a list of numbers, finds its unsort degree D.
The unsort degree of a list is the number of pairs of element positions in the list such that the first position precedes the second in the list, but the number occupying the first position is greater than the number occupying the second position.
Some examples: the unsort degree of the list [1, 2, 3, 4] is 0 the unsort degree of the list [2, 1, 4, 3] is 2 the unsort degree of the list [4, 3, 2, 1] is 6
Representing setsSets can be represented by the existing datatypes in a convenient way by
enforcing an order on a structure used to store the set. For instance using an ordered list (or tree) where each element has a unique occurrence and where all operations are assumed to take ordered unique lists as input and produce ordered unique lists.
If the sets are allowed to contain uninstantiated elements, however, we may have some problems with enforcing the requirement that the lists are ordered and unique, since the requirement may be violated in a later stage.
Consider for instance[X,Y,Z] as a representation of a set with three uninstantiated elements. Of course if X=Y is executed, the list no longer contains unique elements. Perhaps even more obvious is that we cannot ensure that the order of the elements is the one intended until the elements are at least partially known.
invariants are the responsibility of the programmer
Syntactic support
Using op/3 properties (priority, prefix,infix,postfix, associativity) of operators can be defined and then used. (see manual for details)
Predicates defined by the user are written with the same syntax as structures, for instance
:- op(950, xfy, [in]).foo(Y) :- X in Y, baz(Y in U,Z).
Advanced recursive techniques
Sterling and Shapiro ch. 7,8,13.3, 15Nilsson and Maluszynski ch. 7.3.
Outline
Programming with accumulating parameters
Programming with difference-structures
Queues with difference-structures
Abstract data types, separation of data definitions (types) and the program's logic
Accumulating parameters
reverse listsa) naive reverse (using append in each recursion step) reverse([], []). reverse([X|Xs], Ys) :-
reverse(Xs, Zs), append(Zs, [X], Ys).
b) reverse-accumulate reverse(Xs, Ys) :-
reverse(Xs, [], Ys). reverse([], Acc, Acc). reverse([X|Xs], Acc, Ys) :-
reverse(Xs, [X|Acc], Ys).
advice: draw simple SLD-tree and check substitutions!
Built-in arithmetics
ExampleDefine the predicate for computing the factorial of a given integer.
a) recursion
factorial(0, 1).factorial(N, F) :-
N > 0,N1 is N -1,factorial(N1, F1),F is N*F1.
Built-in arithmetics
ExampleDefine the predicate for computing the factorial of a given integer.
b) recursion with an accumulator
factorial(N, F) :- factorial(N, F, 1).factorial(0, F, F).factorial(N, F, F1) :-
N > 0,N1 is N -1,F2 is N*F1,factorial(N1, F, F2).
Built-in arithmeticsExample
Define the predicate for computing the sum of members of integer-list.
a) recursionsumlist([], 0).sumlist([I|Is], Sum) :-
sumlist(Is, Sum1),Sum is Sum1 + I.
b) iteration (with accumulator)sumlist(List, Sum) :- sumlist(List, 0, Sum).sumlist([], Sum, Sum).sumlist([I|Is], Sum1, Sum) :-
Sum2 is Sum1 + I,sumlist(Is, Sum2, Sum).
Using Abstract Data Types
Separation of data definitions (types) and the program's logic.
Specify a set of objectsSpecify set of operations (relations, functions)
on the objects Allow access to objects only through defined
operations
Using Abstract Data Types
Assume the "data type predicates" specifying operations on a given representation of lists:
cons(H,T,[H|T]). nil([]). equal(X,X).
Using this method allows change in representation without change in the code of the algorithm.
Abstract form of append/3
append(A,B,C) :- nil(A), equal(B,C).
append(A,B,C) :-cons(H,T,A),cons(H,R,C),append(T,B,R).
Changing Representation
Note that the representation of lists can be changed without changing the algorithmic code defining append/3
by replacing these "datatype predicates": cons(H,T,foo(T,H)). nil(bar).
Difference-lists
syntax D1-D2, where D1 is a list, which ends with D2
D-D - empty list
diff-lists representing [a,b,c][a, b, c| R] – R[a, b, c, d, e] - [d,e][a, b, c]-[]
adding two lists by unification only [a, b, c] use D1=[a,b,c|R1]-R1=D0-R1 [d] use D2=[d|R2]-R2
[a,b,c,d]use D=D0-R2, assuming that R1=[d|R2]
Difference-lists
concatenationappend_dl(D0-D1, D1-D2, D0-D2).
reversereverse(X, Y) :- reverse_dl(X, Y-[]).
reverse_dl([], Xs-Xs).reverse_dl([X|Xs], Ys-Zs) :-
reverse_dl(Xs, Ys-[X|Zs]).
Difference-lists
sortingquicksort(Xs, Ys) :- quicksort_dl(Xs, Ys-[]).quicksort_dl([X|Xs], Ys-Zs) :-
partition(X, Xs, Littles, Bigs),quicksort_dl(Littles, Ys-[X|Z1]),quicksort_dl(Bigs, Z1-Zs).
quicksort_dl([], Xs-Xs).
partition(X,[Y|Ys],[Y|Ls],Bs) :- X > Y, partition(X,Ys,Ls,Bs).
partition(X,[Y|Ys],Ls,[Y|Bs]) :- X =< Y, partition(X,Ys,Ls,Bs).
partition(_,[],[],[]).
Queues
A queue may be implemented as a difference listenqueue and dequeue
- enqueue(Element, OldQueue, NewQueue) enqueue(X, Qh-[X|Qt], Qh-Qt).
- dequeue(Element, OldQueue, NewQueue) dequeue(X,[X|Qh]-Qt, Qh-Qt).
Queues (cont)
S=[in(5), in(9), in(10), out(X1), out(X2), in(4)]
- an input list for queueing
queue(S) :- queue(S, Q-Q).queue([], Q).queue([in(X)|Xs], Q) :- enqueue(X, Q, Q1), format("In ~d ~n", X),
queue(Xs, Q1). queue([out(X)|Xs], Q) :- dequeue(X, Q, Q1), format("Out ~d ~n", X),
queue(Xs, Q1).
F3: Programming with search
Sterling and Shapiro ch. 6,7,11,14,20,21Nilsson and Maluszynski ch.4,5,11,12, A.3
Outline - Search and control
Proof treesCutNegation, SLDNF
Outline
Cutexecution of a program with cutsinsertion of cuts in your own program:
"green" and "red" cuts
implementation of if-then-else
Negationbasic conceptsexecution of programs with negationimplementation of negation
Controlling search
Sometimes when a solution to a subproblem has been found, no other solutions to it or to earlier proved subgoals of the current goal need to be considered.
By using the non-logical primitive predicate !, named 'cut', you remove alternative branches to subgoals and to the clause that is currently being proved. The alternatives on a 'higher' level, that is to the clause which the current goal is a part of are, though, kept. This can decrease the amount of unnecessary computation.
Cut
syntax !, can be placed in the body of a clause or a goal as one of its atoms to cut branches of an SLD-tree p(X) :- q(X), !, r(X).
effects divides the body into two parts: when "!" is reached, it is evaluated to
true and all backtracking of the left-side of the body is disallowed. The execution of the right-side of the clause body continues as usual.
new matches of the head of the clause are disallowed e.g. backtracking is stopped one level up in the SLD-tree
Cut performs two operations
P :- Q, !, R.P :- ...
removes alternatives to Q that haven't been tried when passing the cut
removes alternatives to P that haven't been tried when passing the cut
Cut
example: execution of a program with cutsConsider the following program:
top(X,Y):- p(X,Y).top(X,X) :- s(X).p(X,Y) :- true(1), q(X), true(2), r(Y).p(X, Y) :- s(X), r(Y).q(a). q(b). r(c). r(d). s(e).true(X).
?- top(X,Y).¤ in the given program (seven answers)¤ when true(1) is replaced by ! (five answers)¤ when true(2) is replaced by ! (three answers)
Cut
inserting cuts in your own program in order to increase efficiency"green" cut:
does not change the semantics of a program (cuts away only failing branches in an SLD-tree)
"red" cut: changes the semantics of a program (cuts also away
success branches in an SLD-tree)in general, the red cuts are considered harmful
Cut
"Green cut": an exampleGiven two sorted integer-lists Xs and Ys, construct a sorted integer-list Zs, containg elements from Xs and Ys.
merge([], Ys, Ys).merge(Xs, [], Xs).merge([X|Xs], [Y|Ys], [X|Zs]) :-
X < Y, merge(Xs, [Y|Ys], Zs).merge([X|Xs], [Y|Ys], [X, Y|Zs]) :-
X = Y, merge(Xs, Ys, Zs). merge([X|Xs], [Y|Ys], [Y|Zs]) :-
X > Y, merge([X|Xs], Ys, Zs).
Cut
"Green cut": an example (cont.)
merge([], Ys, Ys):- !.merge(Xs, [], Xs):- !.merge([X|Xs], [Y|Ys], [X|Zs]) :-
X < Y, !, merge(Xs, [Y|Ys], Zs).merge([X|Xs], [Y|Ys], [X, Y|Zs]) :-
X = Y, !, merge(Xs, Ys, Zs). merge([X|Xs], [Y|Ys], [Y|Zs]) :-
X > Y, !, merge([X|Xs], Ys, Zs).
Cut
"Red cut": an exampleFind the minimum of two integers.
Try:minimum(X, Y, X) :- X =< Y, !.minimum(X, Y, Y).
?- minimum(4,5,Z). - Yes, Z = 4.?- minimum(5,4,Z). - Yes, Z = 4.?- minimum(4,5,5). - Yes.
Correction:minimum(X, Y, Z) :- X =< Y, !, Z=X.minimum(X, Y, Y).
Cut
"Red cut": an exampleChecking membership in a list.
member(X, [X|Xs]) :- !.member(X, [Y|Ys]) :- member(X, Ys).
?- member(a, [b,c,a,d]). - checks membership, OK
?- member(X, [b,c,a,d]). - takes elements from a list, takes only the first element
?- member(b, Z). - generates lists containing bgenerates only one list
OBS: Check the example lookup/3 from the previous lecture!
Cut
implementation of if-then-else
P :- Condition, !, TruePart.P :- ElsePart.
or
(Condition -> TruePart; ElsePart)
For example,
minimum(X, Y, Z) :- (X =< Y -> Z = X; Z = Y).
Negation: how to use and prove negative information?
A negated query 'not_p(x)' should succeed if the proof of the statement 'p(x)' fails and it should fail if the proof of the statement 'p(x)' succeeds.
%not_p(++) (++ stands for a ground term)not_p(X) :- p(X), !, false. not_p(_).
Closed world assumption
That which is not stated explicitly is false
animal(cow).not_animal(X) :- animal(X), !, false.not_animal(_).
?- not_animal(X), X=house.
A house is not an animal so the query should succeedUnfortunately it fails. Why?
Closed world assumption
animal(cow).
not_animal(X) :- animal(X),!, false.not_animal(_).
not_not_animal(X) :- not_animal(X),!, false.not_not_animal(_).
?- not_not_animal(X), X=house.
Unfortunately this succeeds. Why?
Negationhow to use and prove negative information?
to apply closed world assumption (cwa):
the statement \+ A is derivable if A is a formula which cannot be derived by SLD-resolution.- Problem with infinite SLD-trees
implementation: negation as failure (naf):
the statement \+ A is derivable if the goal A has a finitely failed SLD-tree.The problem when A has variables remains.unmarried_student(X) :- \+ married(X), student(X).student(erik).married(jonas).
Negation
SLDNF-resolutionthe combination of SLD-resolution to resolve positive
literals and negation as failure to resolve negative literals
foundation(X) :- on(Y, X), on_ground(X).on_ground(X) :- \+ off_ground(X).off_ground(X) :- on(X, Y).above(X, Y) :- on(X, Y).above(X, Y) :- on(X, Z), above(Z, Y).on(c, b).on(b, a).
Negation
Four kinds of SLDNF-derivations:refutations (that end with success branches);infinite derivations;(finitely) failed derivations;stuck derivations (if none of the previous apply).
NB: check examples in N&M, pp. 71-73.
Negation
implementation of \+
\+ Goal :- call(Goal), !, fail.\+ Goal.
examplep(a).
?- p(X). Yes, X = a.?- \+ \+ p(X). true, X is not instantiated
Search based programming
Sterling and Shapiro ch. 6,7,11,14,20,21Nilsson and Maluszynski ch.4,5,11,12, A.3
Outline- Search based programming
State space programminggenerate-and-testsearching in a state-space
Graph theoretical examples Euler paths, Hamilton paths
Puzzle-solving, game-playing
Sterling and Shapiro ch. 14,20Nilsson and Maluszynski ch. 11
Generate-and-test
A technique in algorithm design, which defines two processes the first generates the set of candidate solutionsthe second tests the candidates
In PROLOG: find(X):- generate(X), test(X).
Important optimisation: to "push" the tester inside the generator as "deep" as possible
Generate-and-test
important optimisation: to "push" the tester inside the generator as "deep" as possible
find(X):- generate1(X), test1(X), generate2(X), test2(X),generate3(X), test3(X), generate4(X), test4(X).
Generate-and-testExample 1
Finding parts of speech in a sentence:
verb(Sentence,Word) :
member(Word,Sentence),verb(Word).
noun(Sentence,Word) :
member(Word,Sentence),noun(Word).
article(Sentence,Word):
member(Word,Sentence),article(Word).
noun(man). noun(woman). article(a). verb(loves).
?- noun([a, man, loves, a woman], N).
N=man; N=woman
NB. member/2 should not contain cuts. Why?
Generate-and-test Example 2
Place N queens on a NxN chess-board in such a way that any two queens are not attacking each other.
a) Naive generate and test places N queens and then test whether they are not attacking each other. The answer is a list of queens' positions, for example [3, 1, 4, 2].
queens(N, Qs) :- range(1, N, Ns), % Ns is the list of integers in 1..N permutation(Ns, Qs), % Qs is a permutation of Nssafe(Qs). % true, if the placement Qs is safe
range(M, N, [M|Ns]) :- M < N, M1 is M +1, range(M1, N, Ns).
range(N, N, [N]).
Generate-and-test
Example 2 (cont.)permutation(Xs, [Z|Zs]) :-
select(Z, Xs, Ys), permutation(Ys, Zs).
permutation([], []).
select(X, [X|Xs], Xs).select(X, [Y|Ys], [Y|Zs]) :- select(X, Ys, Zs).
safe([Q|Qs]) :- safe(Qs), \+ attack(Q, Qs).safe([]).
attack(X, Xs) :- attack(X, 1, Xs).attack(X, N, [Y|Ys]) :- X is Y+N; X is Y-N.attack(X, N, [Y|Ys]) :- N1 is N+1, attack(X, N1, Ys).
Generate-and-test
Example 2 (cont.)b) When generating a position of a queen, test whether it is permitted
queens(N, Qs) :- range(1, N, Ns),queens(Ns, [], Qs).
queens(UnplacedQs, SafeQs, Qs) :-select(Q, UnplacedQs, UnplacedQs1),\+ attack(Q, SafeQs),queens(UnplacedQs1, [Q|SafeQs], Qs).
queens([], Qs, Qs).
select/3, attack/2 are the same as in a).
Searching in a State-space
- loop-avoidance in searching for a path
- efficiency issues
- different search strategies
Searching in a State-space
Many problems in computer science can be formulated as follows:
Given some start-state S0 and a set of goal-states determine whether there exists a sequenceS0 ---> S1, S1 ---> S2, …, Sn-1 ---> Sn,such that Sn belongs to a set of goals.
States can be seen as nodes in a graph whose edges represent the pairs in the transition-relation, then the problem reduces to that of finding a path from the start-state to one of the goal-states.
Searching in a state-space
Finding a pathpath(X,X).path(X,Z) :- edge(X,Y), path(Y,Z).
edge(X,Y) :- % define construction/finding of the next node
For exampleedge(a,b). edge(b,c). edge(c,d).
?- path(a,d). - Yes.?- path(a,X). - Yes. X=a; X=b; X=c; X=d.?- path(X,d). - Yes. X=d; X=a; X=b; X=c.
Searching in a state-space
Loop detectionpath(X, Y) :-
path(X, Y, [X]).
path(X, X, Visited).path(X, Z, Visited):-
edge(X, Y),\+ member(Y, Visited),path(Y, Z, [Y|Visited]).
member(X, [X|Y]) :- !.member(X, [Y|Z]) :- member(X, Z).
Searching in a state-space
Returning the path as an answer to the goal path(X, Y, Path) :-
path(X, Y, [X], Path).
path(X, X, Visited, Visited).path(X, Z, Visited, Path):-
edge(X, Y),\+ member(Y, Visited),path(Y, Z, [Y|Visited], Path).
member(X, [X|Y]) :- !.member(X, [Y|Z]) :- member(X, Z).
Searching in a state-space
Puzzle: Missionaries and cannibalsThree missionaries and three cannibals must cross a river, but the
only available boat will hold only two people at a time. There is no bridge, the river cannot be swum, and the boat cannot cross the river without someone in it. The cannibals will eat any missionaries they outnumber on either bank of the river.
The problem is to get everyone across the river with all the missionaries uneaten.
Searching in a state-space
puzzle( Moves ) :-path( state(3, 3, left), state(3, 3, right), Moves).
path(InitNode, FinalNode, Path) :-path(InitNode, FinalNode, [InitNode], Path).
path(InitNode, FinalNode, _, []) :- InitNode = FinalNode, !.path(Node0, FinalNode, VisitedNodes, [Arc|Path]):-
edge(Node0, Arc, Node1),\+ member(Node1, VisitedNodes),path(Node1, FinalNode, [Node1|VisitedNodes], Path).
Searching in a state-space
Example (cont.) edge( state(M0, C0, L0), move( M, C, D), state (M1, C1, L1) ):-
member(M, [0, 1, 2]),member(C, [0, 1, 2]),M + C >= 1,M + C =< 2,M0 >= M,C0 >= C,M1 is 3 - (M0 - M),C1 is 3 - (C0 - C), ( M1 =:= 0 ; M1 =:= 3; M1 = := C1), (L0 = left -> ( D = leftRight, L1 = right); ( D = rightLeft, L1 = left) ).
Searching in a state-space
Better representation – simpler algorithmStore the whole graph as one fact
graph([edge(a,b),edge(c,d),edge(b,c)]).
path(X, Y, Path) :- graph(G),path(X, Y, Path, G, []).
path(X, X, [X], G, G).path(X, Z, Path, G0, G1):-
deleteedge(X, Y, G0, Gt),path(Y, Z, [Y|Path], Gt,G1).
deleteedge(X, Y, [edge(X,Y)|T], T) :- !.deleteedge(X, Y, [A|T], [A|R]) :-
deleteedge(X, Y, T, R).
Searching in a state-space
Basic search methodsdepth-first - interprets current nodes as a stackbreadth-first - interprets current nodes as a queuebounded-depth - "controls" the depth of the searchiterative-deepening -"controls" the depth of the searchheuristic methods – using domain specific knowledge
Searching in a state-space
Heuristic searchto solve larger problem, some domain-specific
knowledge must be added to improve search efficiency
the term heuristic is used for any advice that is often effective, but isn't guaranteed to work in every case
a heuristic evaluation function estimates the cost of a shortest path between a pair of states
Many search strategiesBottom-up - inductively generate facts from known
factsTop-down - recursively find supporting rules for query
Serial - alternatives one at a time, backtrack for moreOr- parallel - to prove A or B try A and B in parallelcollect all solutions or choose one (first found, best
etc.)Concurrent - to prove A & B try A and B concurrentlyAND-parallel - to prove A & B where A and B have nothing
in common do A and B in parallel, then combine results and go on
Iterative deepening – search all solutions down to a maximum depth, then increase max
Tabled execution – keep ongoing calls in a table to avoid redundant work
Parallelism and Concurrency
Concurrent logic programming modelsParallel execution on multiprocessors
Sterling and Shapiro ch. 14.2 Nilsson and Maluszynski ch.12,14, A.3
Concurrent Logic Programming
essense goals in the body of a clause can execute as interleaving processes p(X) :- q(X), r(X).
Synchronisation on variable unification "X? = Y" means "X must not be bound by unification"
effects
divides the body into two parts and switches execution between the goals.
deadlock is possible
Parallel logic programming
Computation can be distributed over several computers possibly sharing memory.
OR-parallelismcopying of backtrack stack sharing of binding environment
AND-parallelismindependent AND (CIAO-Prolog)stream-and ("Penny", parallel AKL)
Game playingGame trees
a game tree is an explicit representation of all possible plays of the game. The root node is the initial position of the game, its successors are the positions which the first layer can reach in one move, their successors are the positions resulting from the second player's replies and so on.
the trees representing the games contain two types of node: max at even levels from the root, and min nodes at odd levels of the root.
a search procedure combines an evaluation function, a depth-first search and the minimax backing-up procedure.
Game playing
4
4 2
4 8 2 7 5
4 8 5 2 7 53
Maximise
Maximise
Minimise
F4: Logic and grammars
Sterling and Shapiro ch. 19 (except 19.2),24Nilsson and Maluszynski ch.10 SICStus Prolog Manual.
Outline
grammarscontext free grammarscontext dependent grammars
Context Free Grammars
A context free grammar is a 4-tuple < N, T, P, S >
where N and T are finite, disjoint sets of nonterminal and terminal symbols respectively, (N U T)* denotes the set of all strings (sequences) of terminals and non-terminals P is a finite subset of N x (N U T)* ,S is a nonterminal symbol called the start symbol. Empty string is denoted by and elements of P are usually written in the form of production rules:
A ::= B1, …, Bn (n > 0)A ::= (n = 0)
Grammars
Example 1 <sentence> ::= <noun-phrase><verb-phrase><noun-phrase> ::= the <noun><verb-phrase> ::= runs<noun> ::= engine<noun> ::= rabbit
<sentence> derives the strings the rabbit runs, the engine runs. How?
Grammars Example 2
prod_rule(sentence, [noun_phrase, verb_phrase]).prod_rule(noun_phrase, [the, noun]).prod_rule(verb_phrase, [runs]).prod_rule(noun, [rabbit]).prod_rule(noun, [engine]).
An interpreter for production rules:derives_directly(X, Y) :- derives(X, X).
append(Left, [Lhs|Right], X), derives(X, Z) :-prod_rule(Lhs, Rhs), derives_directly(X,
Y),append(Left, Rhs, Temp), derives(Y, Z).append(Temp, Right, Y) .
?- derives([sentence], X). Yes. What is X unified to?
How can you force all productions to be considered before terminating?
Grammars
Example 3sentence(Z) :- append(X, Y, Z), noun_phrase(X), verb_phrase(Y).noun_phrase([the|X]) :- noun(X).verb_phrase([runs]).noun([rabbit]).noun([engine]).
append([], X, X).append([X|Xs], Y, [X|Zs]) :- append(Xs, Y, Zs).
?- sentence([the, rabbit, runs]). Yes.?- sentence([the, X, runs]). X = rabbit, X = engine?- sentence(X).
Grammars
Example 4Usage of difference-lists:
sentence(X0-X2) :- noun_phrase(X0-X1), verb_phrase(X1-X2).
noun_phrase([the|X]-X2) :- noun(X-X2).verb_phrase([runs|X]-X).noun([rabbit|X]-X).noun([engine|X]-X).
?- sentence([the, rabbit|X] - X).?- sentence([the, rabbit, runs] - []).?- sentence([the, rabbit, runs, quickly] - [quickly]).?- sentence(X).
Grammars
Example 5"Collecting" the result:
sentence(s(N, V), X0-X2) :- noun_phrase(N, X0-X1), verb_phrase(V, X1-X2).
noun_phrase(np(the, N), [the|X]-X2) :- noun(N, X-X2).verb_phrase(verb(runs), [runs|X]-X).noun(noun(rabbit), [rabbit|X]-X).noun(noun(engine), [engine|X]-X).
?- sentence(S, [the, rabbit, runs] - []).Yes. S=s(np(the, noun(rabbit)), verb(runs)).
Grammars
Context dependent grammars <sentence> ::= <noun-phrase>(X) <verb>(X)<noun-phrase>(X) ::= <pronoun>(X)<noun-phrase>(X) ::= the <noun>(X)<verb>(singular) ::= runs<verb>(plural) ::= run<noun>(singular) ::= rabbit<noun>(plural) ::= rabbits<pronoun>(singular) ::= it<pronoun>(plural) ::= they
Grammars
Example 6
sentence(X0-X2) :- noun_phrase(Y, X0-X1), verb(Y, X1-X2).
noun_phrase(Y, X - X2) :- pronoun(Y, X-X2).noun_phrase(Y, [the|X]-X2) :- noun(Y, X-X2).
verb(singular, [runs|X]-X).verb(plural, [run|X]-X).
noun(singular, [rabbit|X]-X).noun(plural, [rabbits|X]-X).
pronoun(singular, [it|X]-X).pronoun(plural, [they|X]-X).
Recognizers for languages: Lexers and parsers
A program that recognizes a string in a formal language is often divided into two distinct parts:
Lexer: translation from lists of character codes to lists of 'tokens'
Parser: the translation from lists of 'tokens' to parse trees
Concrete syntax (describes a string in a language, a list of tokens)
<A> ::= <B>< C> [foo,bar]
Abstract syntax (describes a syntax tree, a term)
<A> :: <B>< C> ‘A’(B, C)
Logical Input
char_infile(FileName,Offset,List) :-open(FileName,read,S),skipchars(S,Offset),readchars(S,List),close(S),!.
skipchars(_S,0) :- !.skipchars(S,I) :- I>0, get0(S,_), J is I-1,
skipchars(S,J).
readchars(S,L) :- get0(S,C), readchars0(S,L,C).
readchars0(_,L,-1) :- !, L=[].readchars0(S,L,C) :- L=[C|R], readchars(S,R).
A recognizer for s-expressions
An s-expression is a general representation form for data used in LISP and in some Prolog dialects, especially those embedded in a LISP or SCHEME environment.
We will show how a program identifying s-expressions looks in Prolog.
Backus-Naur grammar for a legal s-expression
<s-expr> ::= <s-atom> | '(' <s-exprs> ['.' <s-expr>] ')'<s-exprs> ::= <s-expr> [<s-exprs>]<s-atom> ::= [<blanks>] <non_blanks> [<blanks>]<non_blanks> ::= <non_blank> [<non_blanks>]<blanks> ::= <blank> [<blanks>]
A non-blank is a character with a character code greater than 32 (decimal). A blank is a character with a character code less than or equal to 32.
The input to the parser is represented as a list of characters (or character codes).
A recognizer for the s-expression grammar
The predicate sExpr(L1,L2,S) succeeds if the beginning of the list L1 is a list of characters representing an s-expression. As a result of the computation the structure S contains a structure (a list) representing the s-expression. The rest of the input text is stored in the list L2.
p :- sExpr("(A.B)",Out,S), write(Out), write(S).
sExpr(In,Out,S) :-(blanks(In,I0); In=I0),(sAtom(I0,Out,S);lpar(I0,I1),sExprs(I1,I2,S,Last),(dot(I2,I3),sExpr(I3,I4,Last);I4=I2,Last=[]),rpar(I4,Out)).
sExprs(In,Out,[H|T],Last) :-sExpr(In,I1,H),(sExprs(I1,Out,T,Last);Out=I1, T=Last).
sAtom(In,Out,A) :- nonBlanks(In,Out,A).
nonBlanks(In,Out,[H|T]) :-nonBlank(In,I1,H),(nonBlanks(I1,Out,T);I1=Out,T=[]).
blanks(In,Out) :- blank(In,I1), (blanks(I1,Out) ; I1=Out).blank(In,Out) :- In=[C|Out], C<=32.
nonBlank(In,Out,C) :- In=[C|Out], C>=48.
lpar(In,Out) :- In=[#(|Out].rpar(In,Out) :- (blanks(In,I1); In=I1), I1=[#)|Out].dot(In,Out) :- (blanks(In,I1); In=I1), I1=[#.|Out].
A Grammar For Horn Clauses
<clause> ::= <head> [ ':-' <body> ]<head> ::= <goal><body> ::= <goal> [ ',' <body> ]<goal> ::= <name> '(' <termlist> ')' | <cut> | <false><termlist> ::= <term> [ ',' <termlist> ]<term> ::= <variable>
| <dataconstructor> [ '(' termlist ')' ]| <constant>| <list>| <integer>| <expression>
<list> ::= '[' <term> [ <moreterms> ] [ '|' <term> ] ']'<moreterms> ::= ',' <term> [ <moreterms> ]<expression> ::= <term> <operator> <term> | <operator> <term><operator> ::= '=' | '<' | '=<' | '>=' | '>' | '+' | '-' | '*' | '!'<cut> ::= '!'<false> ::= 'false'
Outline - Definite Clause Grammars
Definite clause grammarsParsing Translating grammars into logic programs
Sterling and Shapiro ch. 19 (except 19.2)Nilsson and Maluszynski ch 10.4,10.5SICStus Prolog Manual
Definite Clause Grammars
(DCGs) special syntax for language specifications. The system automatically compiles DCG into a Prolog
clause. DCGs are a generalisation of CFGs.
Definite Clause Grammars
Formalism<N, T, P> N - a possibly infinite set of atoms (non-terminals) T - a possibly infinite set or terms (terminals) P is in N x (N U T)* - a finite set of production rules N and T are disjoint
Definite Clause GrammarsSyntax
terminals are enclosed by list-brackets;nonterminals are written as ordinary compound terms or
constants;',' separates terminals and nonterminals in the right-hand
side;'-->' separates nonterminal to the left from terminals and
nonterminals in the right-hand side;extra conditions, in the form of Prolog procedure calls,
may be included in the right-hand side of a grammar rule. Such procedure calls are written enclosed in '{ }' brackets;
the empty string is denoted by the empty list []
Definite Clause Grammars
Examplesentence --> noun_phrase, verb_phrase.noun_phrase --> [the], noun.verb_phrase --> [runs].noun --> [rabbit].noun --> [engine].
?- sentence(X,A).
Yes
X=[the, rabbit, runs], A=[]
X=[the, engine, runs], A=[]WHY? see the next page.
Definite Clause Grammars
Compilation of DCG's into Prologp(A1, …, An) --> T1, .., Tm
is translated into the clause:p(A1, …, An, X0, Xm) :- T1', …, Tm',
where each Ti' is of the form:q(S1,…Sn, Xi-1, Xi) if Ti is of the form q(S1, …, Sn)'C'(Xi-1, X, Xi) if Ti is of the form [X]T, Xi-1 = Xi if Ti is of the form {T}Xi-1 = Xi if Ti is of the form []
and X1,…Xm are distinct variables
Definite Clause Grammars
Example translated to:
expr(X) --> expr(X, X0, X4) :- term(Y), term(Y, X0, X1),[+], 'C'(X1, +, X2),expr(Z), expr(Z, X2, X3),{X is Y + Z}. X is Y+Z, X3 = X4.
NB:sentence --> noun, verb. sentence(B,C) :-
noun(B,A),verb(A,C).
Definite Clause Grammars
Example grammar<expr> ::= <term> + <expr><expr> ::= <term> - <expr><expr> ::= <term><term> ::= <factor> * <term><term> ::= <factor> / <term><term> ::= <factor><factor> ::= 0|1|2|….
Definite Clause Grammars
Example grammar as DCG (cont.)expr(X) --> term(Y), [+], expr(Z), { X is Y + Z}.expr(X) --> term(Y), [-], expr(Z), { X is Y-Z}.expr(X) --> term(X).term(X) --> factor(Y), [*], term(Z), { X is Y*Z}.term(X) --> factor(Y), [/], term(Z), { X is Y/Z}.term(X) --> factor(X).factor(X) --> [X], {integer(X)}.
?- expr(X, [2, *, 2, +, 4, *, 4], []). X = 20.
NB: avoid rules, which lead to "left hand recursion"expr --> expr, [+], expr.
Definite Clause GrammarsExample
Write a DCG which accepts strings in the language an bm cn (n, m >= 0).
a) if n and m are fixed: n=1, m=2.
abc --> a, b, c.a --> [a].b --> [bb].c --> [c].
test(String) :- abc(String, []).?- test(X). Yes. X=[a,bb,c] ?- test([a,bb,c]). Yes.
Definite Clause Grammars
Example (cont.)Write a DCG which accepts strings in an bm cn dm (n, m >= 0).
b)general case:
abcd(N,M) --> lit(a,N), lit(b,M), lit(c,N), lit(d,M).lit(L, 0) --> [].lit(L, I) --> [L], lit(L, I1), {I is I1+1}.
test(String) :- abcd(N, M, String, []).
?- test([a, b, b, c, d, d]). Yes.
Interpreter as transition relation
An interpreter can be formulated as a transition relation p/2 from states to new states
S0 ---> S1 ---> S2 ---> ... ---> Sn-1 ---> Sn
p(S0,S1), p(S1,S2), … p(Sn-1, Sn)
Assignment as predicate
For instance
X:=E would be modeled as a predicate
assign(X,E,StateIn,StateOut) :- ...
imperative program
state transition relation in a procedure in an imperative language
proc {Prog Z := X X := Y Y := Z
}
Program as transition relation
prog(In,Out) :- assign(Z,X,In,T1), assign(X,Y,T1,T2), assign(Y,Z,T2,Out).
Using grammar notation
The state transformation could be expressed using the grammar notation
prog --> assign(z,x), assign(x,y), assign(y,z).
Transition relations declaratively
State transition predicates are expressed in a perfectly declarative way
They still look very similar to the usage of assignment in an imperative language
It expresses the change of a thread of state objects logically
Ö4: Case study: A compiler for three model computers
Clocksin, ch. 9
F7: Program Transformation
Sterling and Shapiro ch. 13,16,18
Outline
Transformations of programsfold/unfold - partial evaluationHigher order programming
defining and using apply, map
Transformation rules for programs
Since logic programs are defined as axioms, it is often possible to define logically based transformation rules, usable to improve programs.
Here are some examples of such rules:
Equality reordering
P :- … e1…q…e2…-------------------------
-P :- e1,e2…q …
Motivation:
In a (pure) logic program the search will be made more efficient if information about equalities is known as soon as possible to the search procedure.
Equality removal
P :- …,e1,…,e1,…,e1,…-------------------------
-------P :- …,e1,…
Motivation:
In a (pure) logic program statements about equalities need not be repeated. Unification is idempotent, that is the bindings of variables achived by applying a unification Term1 = Term2 will not be extended by a second application of the same unification.
Clause level transformation
A clause (j) is a specialized version of a clause (i)
if (Pi :- Q1,…,Qm) = Pj :- Q1',…,Qm' in a program:..(i) Pi :- Q1,…,Qm
.
.(j) Pj :- Q1',…,Qm'
.
.When clause (j) is a specialized version of clause (i) it is redundant (in a pure logic program).
But: specialization can increase efficiency and also make an otherwise looping program terminate. That is it might be advantageous to generate different versions for different forms of the arguments and then apply other transformations to the resulting clauses.
Removal of failing clauses
Suppose that
(Pi :- Q1,…,Qm) = Pj :- falseThis specialized clause may be removed, since it can never contribute to a solution.
Moreover, clauses Ck containing a goal that only matches the head Pj can be converted into
Ck :- false
and the process can be repeated.
Removal of repeated goals
P :- …,Qn,…,Qn,…-------------------------
P :- …,Qn,…Motivation:
The reasoning concerning equalities above holds for goals in general in a pure logic program. Note that we consider programs equivalent with respect to the Herbrand model. The set of solutions found in proofs may not coincide exactly after such a transformation, but the set of true consequences of a program indicated by the model do coincide given an appropriate interpretation of the free variables in the proofs.
Reordering of goals
P :- …,Qi,…,Qk,…-------------------------
-P :- …,Qk,…,Qi,…
Motivation:
This rule can for instance be used to push recursive calls towards the end of a clause.
Fold/Unfold
P :- Q. | P :- (B1 ; B2).Q :- B1. | Q :- B2. |
-----------> (unfolding)<---------- (folding)
Applying the rules above in combination with rules like those given earlier can create more efficient programs (note that variables must be renamed and the introduction of explicit equalities might be necessary to keep the semantics of P).
Partial deduction/evaluation
Partial deduction - logical semantics is kept
Partial evaluation - operational semantics is kept (same number of solutions, presented in the same order)
Partial evaluation is much harder than partial deduction since it should work even in the presence of non-logical goals such as cut (!), I/O etc.
Stepwise enhancement
identify program skeletons that indicate the control flow.
enhance by adding operations (S&S chap 13)
list([X|Xs]) :- list(Xs).list([]).
sumlist([X|Xs],S) :- sumlist(Xs,S0), S is S0+X.
sumlist([],0).
Stepwise enhancement (cont)
length([_|Xs],L) :- length(Xs,L0), L is L0+1.
length([],0).
sum_length_list([X|Xs],S,L) :- sum_length_list(Xs,S0,L0), S is S0+X, L is L0+1.
sum_length_list([],0,0).
Functions Functions are deterministic relations.There is one unique value in the output
domain for each input tuple.A function f: Term* -> Term can for
instance be encoded as a definition in a logic program as f(X1,…,Xn,Y) with a unique output Y for each input tuple X1,…,Xn.
The relation f/n corresponding to a function with n-1 arguments is deterministic, that is, when all arguments (except possibly the one corresponding to the output value of the function) are fully instantiated.
Higher order
Higher order functions are not directly expressible since functions are not objects in the first-order logical model of a program. We will see later how such programming techniques can be encoded using metaprogramming techniques.
Higher order programming:apply
general formapply(foo,X1...Xn) :- foo(X1...Xn).
mapping a predicate Predname(In,Out) to each element of a list
map_list([X|Xs],Predname,[Y|Ys]) :- apply(Predname,X,Y), map_list(Xs,Predname,Ys).
map_list([],_,[]).
Higher order programming: map
All-solutions predicates
It might be useful to collect several solutions in a list. Prolog gives support for this through some "higher-order" predicates.
father(sven,olle). father(sven,lisa). father(bengt,lisa). father(bengt,sven).
children(X,Kids) :- findall(Kid, father(X,Kid),Kids).
The query
?- children(bengt,Kids).gives
Kids=[lisa,sven]
All-solutions predicates (cont)
father(sven,olle). father(sven,lisa). father(bengt,lisa). father(bengt,sven).
the query
?- findall(F, father(F,Kid),Fathers).
gives
Fathers=[sven,sven,bengt,bengt]
All-solutions predicates (cont)
Instead of a single solution collecting all fathers to some child we might want a separate solution for each child. There is another set precidate for this.
father(sven,olle). father(sven,lisa). father(bengt,lisa). father(bengt,sven).
?- bagof(F, father(F,Kid),Fathers).Kid=lisa Fathers=[sven,bengt]Kid=sven Fathers=[bengt]Kid=olle Fathers=[sven]
All-solutions predicates (cont)
It is often sensible to present sorted lists of unique solutions. This is achieved by setof.
father(sven,olle). father(sven,lisa). father(bengt,lisa). father(bengt,sven).
?- setof(F, father(F,Kid),Fathers).Kid=lisa Fathers=[bengt,sven]Kid=sven Fathers=[bengt]Kid=olle Fathers=[sven]
Ö5: Metaprogramming, Expert Systems
Sterling and Shapiro ch. 10,17 (not 17.5),19.2,22 Nilsson and Maluszynski ch. 8,9
Outline - Metaprogramming
What is a metastatement?Metalogic predicates (built in)solve, augmenting solveIterative deepeningMixing object and metalevel programmingSupport for dynamically changing knowledge bases
What is a metastatement?
A metastatement is a statement about statements
Stockholm is a nine-letter word.
'X+1-Y' has the size five.
'X+1-Y' contains two variables.
This statement is true.
This statement is false.
'P :- Q1...Qn.' is a clause.
This is a metastatement.
Meta-logic
Metalogic refers to reasoning about a formalization of some (other) logical system.
If the metalogic deals with itself it is called circular or metacircular.
In logic programming metalogic has two meanings:
1.using logical inference rules expressed as axioms with a meta-interpreter.
2.expressing properties of the proof procedure.
For the latter case the term "meta-logical predicate" is used, for instance in the SICStus manual.
Ground representation of facts and rules
The logical approach to metaprogramming requires a clear division between object level and meta-level.
Formulas are represented as ground facts, where in particular variables on the object level are represented as constants on the meta-level
Each constant of the object language is represented by a unique constant of the meta-language
Each variable of the object language is represented by a unique constant of the meta-language
Each n-ary functor of the object language is represented by a unique n-ary functor of the meta-language
Each n-ary predicate symbol of the object language is represented by a unique n-ary functor of the meta-language
Each connective of the object language is represented by a unique functor of the meta-language (with corresponding arity)
Ground representation
clause(if(list(x),equals(x,[]))).clause(if(list(x),and(equals(x,cons(x,h,t)),list(t)))).
clause(if(p(x),true)).clause(if(p(x),and(q(x,a),p(b)))).
SLD-resolution rule
The SLD-resolution rule (p. 43 Nilsson & Maluszynski)
<- A1,..,A(i-1),Ai,A(i+1),...,Am B0 <- B1,...,Bn
----------------------------------------------------------
<- (A1,...,A(i-1),B1,...,Bn,A(i+1),...,Am)
The SLD-rule encoded as a relation
step(Goal,NewGoal) :-select(Goal,Left,Selected,Right),clause(C),rename(C,Goal,Head,Body),unify(Head,Selected,Mgu),combine(Left,Body,Right,TmpGoal),apply(Mgu,TmpGoal,NewGoal).
The SLD-rule encoded as a relation (cont)
select/4 describes the relation between a goal and the selected subgoals
clause/1 describes the property of being a clause in the object language
rename/4 describes the relation between four formulas such that two are uniquely renamed variants of the other two
unify/3 describes the relation between two atoms and their mgu
combine/4 describes the relation between a goal and three conjunctions
apply/3 describes the relation between a substitution and two goals
The SLD-rule encoded as a relation (cont)
derivation(G,G).derivation(G0,G2) :-
step(G0,G1), derivation(G1,G2).
Why self-interpreters?
flexibility alternative search strategies
debugging
programs that change during their execution
collecting the actual proof of a satisfied goal (explanation)
non-standard logics: fuzzy logic, non-monotonic logic, modal logic
program transformation, program verification, program synthesis
Non-ground representation
Efficiency of the previous representation is low-> Use object language variables for meta-level also
Seems straightforward, but mixing object level and meta-level has important semantic consequences.
Consider this representation.(with 'if' and 'and' as infix operators):
for facts: ax(Fact if true).for rules: ax(Head if Body).
Metainterpreter for pure Prolog
using the above representation
solve(true).solve(P) :- ax(P if Q), solve(Q).solve(P and Q) :- solve(P), solve(Q).
In order to extend the meta-interpreter to handle also non-logical features of Prolog a different interpreter must be written. For instance ! (cut) is hard to handle.
Metainterpreter generating a proof
Augmented metainterpreter
solve(true,true).solve(P,if(P,ax(P if Q),Qt)) :- ax(P if Q), solve(Q,Qt).solve(P and Q,and(Pt,Qt)) :- solve(P,Pt), solve(Q,Qt).
This metainterpreter generates a proof .
Depth-bounded pure Prolog
An augmented meta-interpreter for depth-bounded search pure Prolog
solve(true,N) :- N>=0.solve(P,N) :- N>0, ax(P if Q), N1 is N-1, solve(Q,N1).solve(P and Q,N) :- solve(P,N), solve(Q,N).
This meta-interpreter finds a proof with a search tree depth of at most N levels.
depth(0).depth(N) :- depth(N1), N is N1+1.
solve(G) :- depth(N), solve(G,N).
solve/1 tries at gradually deeper levels of the tree, iterative deepening.
Meta-logic
built-in predicates that perform operations that require reasoning about: the current instantiation of terms;decomposing terms into their constituents.
Instantiation checking:var(X) - checks that X is uninstantiated variable (not a
structure)nonvar(X) - opposite to var/1.ground(X) - checks that X is completely instantiated
Meta-logic
Example.?- var(X), X=1 -Yes, X=1.?- X=1, var(X). -No.?- nonvar(father(X,Y)). -Yes.
Define the predicate plus/3 which uses built-in arithmetics and performs plus and minus.
plus(X, Y, Z):- nonvar(X), nonvar(Y), Z is X+Y.
plus(X, Y, Z):-nonvar(X), nonvar(Z), Y is Z-X.
plus(X, Y, Z):-nonvar(Y), nonvar(Z), X is Z-Y.
Meta-logic
Type checking:integer(X) - X is instantiated to an integer
float(X) - X is instantiated to a floatnumber(X) - X is instantiated to a numberatom(X) - X is instantiated to an atom (non-variable term of arity 0, other than a number)atomic(X) - X is instantiated atom or numbersimple(X) - X is uninstantiated or instantiated to an
atom or number compound(X) - X is instantiated to a term of arity > 0, i.e.
a list or a structure
Meta-logic, decomposing terms
functor(Term, FunctorName, Arity)Term has functor FunctorName and arity Arity
?- functor(father(erik, jonas), father, 2). - Yes
?- functor(father(erik, jonas), F, A). - F=father, A=2
?- functor(Term, father, 2). - Term=father(_A, _B)
?- functor(Term, father, N).- instantiation error
Meta-logic, decomposing termsarg(N, Term, Argument)
the Nth argument of a compound term Term is Argument
?- arg(1, father(erik, jonas), Arg). - Arg=erik.
?- arg(2, father(erik, X), jonas). - X = jonas.
?- arg(N, father(erik, jonas), jonas).- instantiation error.
?- arg(2, Y, jonas).- instantiation error.
?- arg(1, father(X, Y), Z). - Z = X.
?- arg(3, father(X, Y), Z). - No.
Meta-logic, decomposing terms
Term =.. List (=.. is called univ)List is a list whose head is the atom corresponding to the principal functor of Term, and whose tail is a list of the arguments of Term.
?- father(person(erik,A,b),person(jonas,X,Y)) =.. List.
- List=[father, person(erik,A,B), person(jonas, X,Y)].
?- Term =.. [father, erik, X]. - Term=father(erik,X).
?- father(erik, jonas) =.. [father, erik, jonas].- Yes.
Meta-logic
Define the predicate subterm(Sub,Term), for checking if Sub is subterm of Term
subterm(T, T).subterm(S, T) :-
compound(T),functor(T, F, N),subterm(N, S, T).
subterm(N, S, T) :-N > 1, N1 is N-1,subterm(N1, S, T).
subterm(N, S, T) :-N > 0, arg(N, T, Arg),subterm(S, Arg).
Meta-logicDefine the predicate subterm(Sub,Term) for checking if Sub is
subterm of Term
subterm(T, T).subterm(S, T) :-
compound(T), T =.. [F|Args],
subtermList(S, Args).
subtermList(S, [Arg|Args]) :-
subterm(S, Arg).
subtermList(S, [Arg|Args]) :-
subtermList(S, Args).
Meta-logic
define =.. using functor/3 and arg/3a) Term T is given
T =.. [F|Args] :- functor(T, F, N),args(T, Args, 0, N).
args(_, [], N, N).args(T, [Arg|Args], I, N) :-
I < N,I1 is I+1,arg(I1, T, Arg),args(T, Args, I1, N).
Meta-logicdefine =.. using functor/3 and arg/3
b) List [F|Args] is given T =.. [F|Args] :-
length(Args, N),functor(T, F, N),args(Args, T, 1).
args([], _, _).args([Arg|Args], T, N) :-
arg(N, T, Arg),N1 is N+1,args(Args, T, N1).
Support for dynamic knowledge bases
assert(Clause) - Clause is added to the programclause(Head, Body) - clause with head Head and body Bodyretract(Clause) - Clause is erased from the program
For example,
member(X, [X|Ys]).member(X, [Y|Ys]) :- member(X,Ys).
?- clause(member(H1, H2), B).
H1=X, H2=[X|Ys], B= true;H1=X, H2=[Y|Ys], B=member(X,Ys).
Expert systems
Expert System = Knowledge-base (KB) + Inference engine (IE) + User Interface (UI)
Expert System Shell = (IE) + (UI)
Expert systems: Production rules
knowledge base often consist of a set of production rules of the form
IF A1 AND(OR) A2 ….THEN C1 AND C2 …
Expert systems: forward/backward chaining
forward-chaining: start from assumptions (axioms) and find conclusions
backward-chaining: start from conclusions (hypothesis) and look for supporting assumptions
Expert systems in Prolog
in PROLOG knowledge base = set of clauses inference engine = SLD-resolution
SLD-resolution is a backward chaining proof procedure
Uncertainty
Expert systems might contain rules and/or facts which hold with some degree of uncertainty
IF it is summer and there are no clouds THEN the temperature is above 20 degrees CCERTAINTY 80%
Expert systems, forward/backward
ExampleDevelop an expert system, which finds a name of a fruit when some fruit characteristics such as shape, diameter, surface , colour and the number of seeds are given. Gonzalez: "The engineering of knowledge-based systems." (p91).
Rule 1:IF Shape=long and Colour=green or yellowTHEN Fruit=bananafruit(Name, Shape, Diameter, Surface, Colour, FruitClass, SeedCount, SeedClass):-
Shape == long, (Colour == green; Colour == yellow), Name = banana.
Rule 2:IF Shape=round or oblong and Diameter > 4 inchesTHEN FruitClass=vinefruit(Name, Shape, Diameter, Surface, Colour, FruitClass, SeedCount, SeedClass):-
var(FruitClass), (Shape == round; Shape ==oblong), integer(Diameter), Diameter>4, FruitClass= vine, fruit(Name, Shape, Diameter, Surface, Colour, FruitClass, SeedCount, SeedClass).
Expert systems, forward/backwardRule 4:IF SeedCount=1THEN SeedClass=stonefruitfruit(Name, Shape, Diameter, Surface, Colour, FruitClass, SeedCount, SeedClass)):-
var(SeedClass), integer(SeedCount), SeedCount =:= 1, SeedClass = stonefruit, fruit(Name, Shape, Diameter, Surface, Colour, FruitClass, SeedCount, SeedClass).
Rule 5:IF Seedcount>1THEN SeedClass=multiplefruit(Name, Shape, Diameter, Surface, Colour, FruitClass, SeedCount, SeedClass):-
var(SeedClass), integer(SeedCount), SeedCount > 1, SeedClass = multiple, fruit(Name, Shape, Diameter, Surface, Colour, FruitClass, SeedCount, SeedClass).
Rule 11:IF FruitClass=tree and Colour=red and SeedClass=stonefruitTHEN Fruit=cherryfruit(Name, Shape, Diameter, Surface, Colour, FruitClass, SeedCount, SeedClass):-
FruitClass == tree, Colour == red, SeedClass == stonefruit, Name = cherry.
?- fruit(Name, round, 3, Surface, red, FruitClass, 1, SeedClass). Yes. Name = cherry.
User interface - dialogue
An expert system is often not automatic=> user interaction guides the search
graphical user interfacedialogue
Explanation facilities
Expert systems should be able to explain its conclusions to different people (experts, programmers, users)
- How did you come to your conclusion?
- Why does A follow from B ?
Knowledge acquisition
An expert system should allow incremental updates of the knowledge base and rule base
knowledge can be aquired through dialogue with an expert, or through analysis of the system
Rule base as Prolog clauses
IF A1 AND (A2 OR A3) THEN C1 AND C2
C1 :- A1, (A2 ; A3).C2 :- A1, (A2 ; A3).
Taxonomy of a car-engine
car
fuel system ignition electric system
fuel pump starting motor
spark plugs
fuel battery
fuse
needs(car,fuel_system). ....needs(electric_system,fuse).
Expert System: diagnosis
IF Y is a necessary component for X and Y is malfunctioning THEN X is also malfunctioning
IF X exhibits a fault-symptom Z THEN either X is malfunctioning or there exists another malfunctioning component which is necessary for X
Expert system: diagnosis
X has an indirect fault if there exists a component which is necessary for X and which malfunctions
malfunctions(X) :- needs(X,Y), malfunctions(Y).
malfunctions(X) :- symptom(Y,X), not
indirect(X).
indirect(X) :- needs(X,Y), malfunctions(Y).
Expert systems: abduction
The knowledge base is usually incomplete
symptoms are not known, but need to be established by asking questions in a dialogue
KB + cause |- symptom
Finding cause is named abduction
self-interpreter generating a proof
solve(true,true).solve(P,proof(P,Qt)) :-
kb(P if Q), solve(Q,Qt).solve(P and Q,and(Pt,Qt)) :-
solve(P,Pt), solve(Q,Qt).
This meta-interpreter generates a proof .
query-the-user
solve(true).solve(P and Q) :- solve(P), solve(Q).solve(symptom(X,Y)) :- confirm(X,Y).solve(P) :- kb(P if Q), solve(Q).
confirm(X,Y) :-write('Is the '),write(Y), tab(1),write(X), write('? '),read(yes).
knowledge base
kb(malfunctions(X) if possible_fault(Y,X) and symptom(Y,X)).
kb(possible_fault(flat,tyre) if true).
:- solve(malfunctions(X)).
Is the tyre flat?>yes
X=tyre
(anything but 'yes' as answer makes the query fail)
F8: Case study: Support for reasoning about electronic circuits
Clocksin, ch. 7,8
Ö6: Constraint Logic Programming
SICStus Prolog ManualNilsson and Maluszynski ch.14Course compendium
This lecture is given by Christian Schulte. The written material is based on slides from Per Kreuger SICS and Christian Schulte KTH/ICT/ECS
Significance
Constraint programming identified as a strategic direction in computer science research
[ACM Computing Surveys, December 1996]
This Talk…
…concerned with constraints for
solving combinatorial problems
designed as basic tutorial
Application Areas
Timetabling Scheduling Crew rostering Resource allocation Workflow planning and optimization Gate allocation at airports Sports-event scheduling Railroad: track allocation, train allocation, schedules Automatic composition of music Genome sequencing Frequency allocation …
Techniques
Artificial intelligence
Operations research
Algorithms
Programming languages
Overview
Basics modeling: variables and constraints solving: propagation, distribution, search
Solving realistic problems: scheduling modeling solving
Constraint programming research systems
Case studies instruction scheduling for compiler bus scheduling for real-time system
Conclusion
Constraint Programming (CP)
based on the idea of an abstract space of statements or conditions, a
constraint space
x{4,5} y{4,5}
xy y>3
Constraint Programming (CP)
Some such statements can be regarded as completely determined:
E.g. “ A given train trip will leave from Avesta
15.05 on Thursday, October 10th 2002.”
Constraint Programming (CP)
Other statements are less exact with respect to e.g. the resources some task will need:
E.g.“OVAKO Steel needs to transport between
320 and 280 kilotons of steel from Hofors and Hellefors to Malmö next year.”
Constraint Programming (CP)
Such statements can be represented as constraints in a constraint programming system.
These constitute conditions on the values that variables may take (i.e. a form of type).
Constraint Programming (CP)
It is nontrivial to determine e.g. how the two above mentioned propositions would be related in some model of a planning problem.
Sometimes it is possible to determine some form of “consistency” of such propositions.
Constraint Programming (CP)
EnumerationIt is also possible (in principle) to compute
one or more witnesses (a consistent assignment of values to all the variables) of such a system of conditions.
This is called to enumerate the constraint space and generally involves search.
Constraint Programming (CP)
If more than one witness is computed they can be compared with respect to various cost measures.
To determine the (in some sense) best assignment is modeled as an optimization problem in the constraint system.
Constraint propagation
Much of the search in the enumeration of a constraint space for a given problem can generally be eliminated by a technique called constraint propagation.
x in {2,…,5} & y in {3,…,9}2x+3=y x in {2,3} & y in {7,…,9}2x+3=y & y<9 x<3 x=2 y=7
Constraint propagation
Much of the search in the enumeration of a constraint space for a given problem can generally be eliminated by a technique called constraint propagation.
Each (non primitive constraint) can be regarded as a temporarily suspended computation (of the values of the involved variables) that can be made to interact with other such suspended/reinvoked computations.
Constraint propagation
Computations are suspended when information needed to determine a value is missing but is rescheduled as soon as that information becomes available.
One can view the constraint programming system as a pool of concurrent processes that communicate, interact and synchronize through shared variables.
Constraint propagation
Propagation is normally not sufficient to completely determine the values of the involved variables.
To achieve this we need to combine propagation with enumeration in such a way that values chosen during search trigger further propagation which in turn guides the continued search, etc.
Constraint Modelling (CP)
The computational complexity of the task to find a solution to a given problem depends to a large extent on the expressive power of the language used to formalize the problem.
To formulate a mathematical model of some real process is generally difficult. It requires a thorough understanding of both the problem domain and the methods employed to solve the problem.
Constraint Programming (CP)
To some extent this is still more of a craft than a science.
A large body of typical problems with standard models have been identified.
Early attempts to develop a methodology has started to give results.
Finite Domains
The constraint programming systems that have been most actively developed the last ten years are those that built on finite domains.
In such a system each variable can take on values from a finite set of discrete values.
This type of variable is natural to use to model discrete entities such as the number of engines or staff that have been allocated to a given task.
Continuous Domains
Finite domain constraints are, however, unnecessarily restrictive when the modeling concerns values that can be assumed to vary over continuous domains (with an infinite number of possible values), for instance time.
Comparison with Operations Research ---OR
Techniques from Operations Research, e.g. linear programming (LP) and integer programming (IP) efficiently handles models where:
1. Most of the variables are continuous.2. The model can be relatively directly expressed
as a set of simple (linear) equations and inequalities.
3. A simple (linear) and well defined cost function captures well the “goodness” of different solutions to a given problem.
Comparison with Operations Research (2)
The techniques that have been developed in constraint programming, using finite domains, work well also when:
1. A majority of the variables model naturally discrete entities.
2. The cost function is hard to determine.3. The model contains complicated (for instance
non-linear) conditions.In this way these two classes of techniques can be
said to complement each other.
Global constraints
The first type of constraint that was studied in constraint programming was constraints that limits one or relate two variables, e.g. constraints like:
< > =< >= = /=
Global constraints
In contrast to these simple binary constraints the focus has in recent years more and more been on complex constraints between an unlimited number of variables.
E.g. constraints: - relating variables with the value of a linear sum - maintaining pairwise disequality of an
arbitrarily large set of variables - implementing various scheduling, matching,
packing and/or placement mechanisms
Global constraints (2)
The expression “global constraints” for this type of constraints was introduced in {BC94} and refers to arguments that can be made over a multitude of variables related with a non-binary condition.
Global constraints (2)
Global constraints can in principle often be encoded in terms of a set of simpler binary constraints which semantically have the same meaning.
This is rarely practical, however, since an efficient solution can seldom be achieved by only considering the variables pairwise.
Global constraints (2)
Global constraints constitute abstractions of more complicated properties of problems and enables computations on a more detailed model
Algorithms as constraint abstractions Often methods from operations analysis,
matching theory or graph algorithms can be integrated into a constraint programming system as global constraints.
This is an active and very promising research area in constraint programming{bel00}.
An example: Getting Started
A toy problem… Modeling
Solving: propagation and search
Send More Money (SMM)
Find distinct digits for letters, such that
SEND+ MORE= MONEY
Constraint Model for SMM
Variables: S,E,N,D,M,O,R,Y {0,…,9}
Constraints:distinct(S,E,N,D,M,O,R,Y)
1000×S+100×E+10×N+D + 1000×M+100×O+10×R+E = 10000×M+1000×O+100×N+10×E+Y
S0 M0
Solution for SMM
Find values for variables such thatall constraints satisfied
Finding a Solution
Enumerate assignments: poor! Constraint programming
compute with possible values prune inconsistent values
constraint propagation search
distribute: define search tree explore: explore for solution
Some Concepts
Constraint store Basic constraint Propagator Non-basic constraint Constraint propagation
Constraint Store
Stores basic constraints map variables to possible values
Domains: finite sets, real intervals, trees, …
x{3,4,5} y{3,4,5}
finite domain constraints
Propagators
Implement non-basic constraints
distinct(x1,…,xn)
x + 2×y = z
Propagators
Amplify store by constraint propagation
x{3,4,5} y{3,4,5}
xy y>3
Propagators
Amplify store by constraint propagation
x{3,4,5} y{4,5}
xy y>3
Propagators
Amplify store by constraint propagation
x{3,4,5} y{4,5}
xy y>3
Propagators
Amplify store by constraint propagation
x{4,5} y{4,5}
xy y>3
Propagators
Amplify store by constraint propagation Disappear when entailed
no more propagation possible
x{4,5} y{4,5}
xy
Constraint Space
Store with connected propagators
x{4,5} y{4,5}
xy y>3
Propagation for SMM
Results in storeS=9 E{4,…,7} N{5,…,8} D{2,…,8}
M=1 O=0 R{2,…,8} Y{2,…,8}
Propagation alone not sufficient! create simpler sub-problems distribution
Distribution
Yields spaces with additional constraints Enables further constraint propagation
x{4,5} y{4,5}
xy y>3
x{4} y{4}
xy y>3
x{5} y{5}
xy y>3
x=4 x4
Distribution Strategy
Pick variable x with at least two values Pick value n from domain of x Distribute with
x=n and xn
Part of model
Search
Iterate propagation and distribution Orthogonal: distribution exploration Nodes:
Distributable Failed Succeeded
SMM: Solution
SEND+ MORE= MONEY
9567+ 1085= 10652
Heuristics for Distribution(CLP jargong: Labelling)
Which variable? least possible values (first-fail) application dependent heuristic
Which value? minimum, median, maximum
x=m or xm split with median m
x<m or xm In general: application specific
SMM: Solution With First-fail
SEND+ MORE= MONEY
9567+ 1085= 10652
Send Most Money (SMM++)
Find distinct digits for letters, such that
and MONEY maximal
SEND+ MOST= MONEY
Best Solution Search
Naïve approach: compute all solutions choose best
Branch-and-bound approach: compute first solution add “betterness” constraint to open nodes next solution will be “better” prunes search space
Also possible: restart strategy
Branch-and-bound Search
Find first solution
Branch-and-bound Search
Explore with additional constraint
Branch-and-bound Search
Explore with additional constraint
Branch-and-bound Search
Guarantees better solutions
Branch-and-bound Search
Guarantees better solutions
Branch-and-bound Search
Last solution best
Branch-and-bound Search
Proof of optimality
Modelling SMM++
Constraints and distribution as before Order among solutions with constraints
so-far-best solution S,E,N,D,M,O,T,Y current node S,E,N,D,M,O,T,Y constraint added10000×M+1000×O+100×N+10×E+Y
<10000×M+1000×O+100×N+10×E+Y
SMM++: Branch-and-bound
SEND+ MOST= MONEY
9782+ 1094= 10876
SMM++: All Solution Search
SEND+ MOST= MONEY
9782+ 1094= 10876
Summary
Modeling variables with domain constraints to state relations distribution strategy solution ordering
Solving constraint propagation constraint distribution search tree exploration
The Art of Modeling
Avoid search, avoid search, avoid… Techniques
increase propagation strength stronger propagators redundant propagators
remove symmetrical solutions good distribution heuristics smart search engines
Distribution Exploration
Distributiondefines shape of search tree
Exploration left-most depth-first interactive, graphical parallel branch-and-bound [prunes tree]
Scheduling
Modeling Propagation
Global constraints
Scheduling
Among the examples of global reasoning that have successfully been introduced into constraint programming systems are a number of fundamental scheduling mechanisms.
Scheduling
A scheduling problem consists of a number of tasks with restrictions on start times, stop times and task duration.
Often the tasks are partially ordered into totally ordered sequences.
Such a totally ordered subset of tasks is often called a job.
Scheduling
Each task uses one or more resources during certain time intervals.
The so called job shop scheduling problem is a classic and well studied case.
Scheduling
Resources can in general model widely different type of entities.
For instance:1. processing equipment in a production
process.2. staff or vehicles in a transport net.3. network resources such as routers and
transport links with limited capacity.
Scheduling
To arrange the tasks so that no limitations in resources are violated is called to schedule the tasks and it is in general a very difficult (NP-complete) computational problem.
Nevertheless the many practical applications for methods in this area make it fairly well studied.
Scheduling: Given
Tasks duration resource
Precedence constraints determine order among two tasks
Resource constraints at most one task per resource
[disjunctive, non-preemptive scheduling]
Scheduling: Bridge Example
Scheduling: Solution
Start time for each task
All constraints satisfied
Earliest completion time minimal make-span
Scheduling: Model
Variable for start-time of task astart(a)
Precedence constraint: a before bstart(a) + dur(a) start(b)
Propagating Precedence
start(a){0,…,7}start(b){0,…,5}
a
b
a before b
Propagating Precedence
start(a){0,…,7}start(b){0,…,5}
a
b
a
b
a before b
start(a){0,…,2}start(b){3,…,5}
Scheduling: Model
Variable for start-time of task astart(a)
Precedence constraint: a before bstart(a) + dur(a) start(b)
Resource constraint:a before b
orb before a
Scheduling: Model
Variable for start-time of task astart(a)
Precedence constraint: a before bstart(a) + dur(a) start(b)
Resource constraint:start(a) + dur(a) start(b)
orb before a
Scheduling: Model
Variable for start-time of task astart(a)
Precedence constraint: a before bstart(a) + dur(a) start(b)
Resource constraint:start(a) + dur(a) start(b)
orstart(b) + dur(b) start(a)
Reified Constraints
Use control variable b{0,1}c b=1
Propagate c entailed propagate b=1 c entailed propagate b=0 b=1 entailed propagate c b=0 entailed propagate c
not easy!
Reification for Disjunction
Reify each precedence[start(a) + dur(a) start(b)] b0=1
and
[start(b) + dur(b) start(a)] b1=1
Model disjunctionb0 + b1 1
Model Is Too Naïve
Local view individual task pairs O(n2) propagators for n tasks
Global view all tasks on resource single propagator smarter algorithms possible
Edge Finding
Find ordering among tasks (“edges”) For each subset of tasks {a}B
assume: a before Bdeduce information for a and B
assume: B before adeduce information for a and B
join computed information can be done in O(n2)
Scheduling Architecture
Constraint programmingnatural modeling
specification modelcompiler
Scheduling Architecture
Constraint programmingexpressive modeling
specification modelcompiler
Summary
Modeling easy but not always efficient constraint combinators (reification) global constraints smart heuristics
More on constraint-based schedulingBaptiste, Le Pape, Nuijten. Constraint-basedScheduling, Kluwer, 2001.
Why Does CP Matter?
Middleware for combining smart algorithmic components
scheduling graphs flows …
plus essential extra constraints
Research in Constraints
Propagation algorithms Search methods Innovative applications Programming and modeling languages
Hybrid methods linear programming local search
Constraints in Sweden
Swedish constraint network
SweConsNet [founded May 2002]
http://www.dis.uu.se/~pierref/astra/SweConsNet/
CP Systems
Commercial ILOG Solver C++ OPL Studio Modeling SICStus Prolog Prolog-based Eclipse Prolog-based
Free GNU Prolog Prolog-based Mozart Oz CHOCO Claire
Oz and Mozart
Constraint-based programming system concurrent and distributed programming combinatorial problem solving and combinations: intelligent agents, …
Mozart implements Oz concurrent constraint programming language with: objects, functions, threads, …
Developed by Mozart ConsortiumSaarland University, GermanySICS/KTH, SwedenUniversité catholique de Louvain, Belgium
Mozart Fact Sheet
Freely available at www.mozart-oz.org
Many platforms supportedUnix Windows Mac OS
Active user community Comes with extensive documentation Many applications
Constraints in Mozart
Rich set of constraintsfinite domains schedulingfinite sets records
New propagators via C++ API Search and combinators [unique]
programmable concurrency-compatible fully compositional
Book: Schulte, Programming Constraint Services. LNAI, Springer 2002.g
Some Research Issues
Search methods Architecture and implementation Automatic selection of good propagator
domain versus bound
Challenging applications
Case Study
Instruction scheduling Bus scheduling
Instruction Scheduling
Optimized object code by compiler Minimum length instruction schedule
precedence latency resources
per basic block
Best paper CP 2001, Peter van Beek and Kent Wilken, Fast Optimal Scheduling for Single-issue Processors with Arbitrary Latencies, 2001.
Model
All issue times must be distinct use single distinct constraint (as in SMM) is resource constraint or unit duration
Latency constraints precedence constraints (as before) difference: duration latency
Making It Work
Only propagate bounds information relevant for distinct
Add redundant constraints regions: additional structure in DAG successor and predecessor constraints
[special case of edge-finding]
Results
Tested with gcc SPEC95 FP Large basic blocks
up to 1000 instructions
Optimally solved less than 0.6% compile time increase limited static improvement (< 0.1%) better dynamic impact (loops)
Far better than ILP approach
Off-Line Scheduling of Real-Time System
System with global clock time-triggered real-time processes message exchange over shared bus
Infinite, periodic schedule map to single fixed time window repeat
Klaus Schild, Jörg Würtz. Scheduling of Time-Triggered Real-Time Systems, Constraints 5(4), 2000.
Model
Single resource: data bus Maximal latencies:
messages valid for at most n time units
Infinite schedule repeat finite schedule of given length repetition does not violate constraints
Performance
Problem size 6,000,000 time units 3500 processes and messages
Model size up to 10 million constraints
Run time from 10 min to 2 hrs (200 MHz PPro)
Summary
Useful for small components in software systems large offline optimization
Widely applicable in your area? hardware design?
Conclusion
Constraint programming useful easy modeling open to new techniques
Constraints for Concurrency
Constraints describe data structures used for control [as opposed to Prolog]
Logic variables as dataflow variables unconstrained suspension constrained resumption synchronization is automatic
Well established idearesumption condition is logical entailment
[Maher,87], ccp [Saraswat,90]
Problem Solving
Constraint domains tree constraints (records, feature) finite domains finite sets
Programmable search and combinators based on computation spaces makes search compatible with concurrency book:
Christian Schulte, Programming Constraint Services
LNAI 2302, Springer-Verlag, 2002