程序代写代做代考 algorithm chain database C flex graph Programming Languages

Programming Languages
Prolog
CSCI-GA.2110-003 Fall 2020

Prolog overview
■ Stands for Programming in Logic.
■ Invented in approximately 1972.
■ Belongs to the logical & declarative paradigms.
■ Based on first order predicate calculus.
■ Used for artificial intelligence, theorem proving, expert systems, and
natural language processing.
■ Used as a standalone language or complements traditional languages.
■ Radically different than most other languages.
■ Each program consists of 2 components:
◆ database (program): contains facts and rules
◆ query : ask questions about relations
2 / 31

Propositional Logic
■ Sentences have truth values.
■ Sentences fragments are represented by Boolean variables, P, Q, etc.
■ Logical connectives are used: P ∧ Q, P ∨ Q, ¬P , etc.
■ Inference rules (e.g. P → Q) expressed using logical connectives.
■ Logical formulas over the variables and connectives yield truth values.
3 / 31

First Order Predicate Logic
■ Used by Prolog
■ Establishes facts about and relationships between items in a universe.
■ Relationships are expressed using predicates.
■ Predicate happy(john) means “john is happy.”
■ Predicate likes(john,mary) means “john likes mary.”
■ Quantification allows us to reason more generally: ∀, ∃
Examples:
∀x : loves(x, mary) means “everyone loves mary.”
∃x : loves(x, mary) means “someone loves mary.”
∀x∃y :loves(x,y) means “everyone loves someone.” ∃y∀x : loves(x, y) means “someone is loved by everyone.” ∃x∃y :loves(x,y) means “someone loves someone.”
4 / 31

Stating Facts
Two ways to state facts:
?- [user].
sunny.
% user://1 compiled 0.00 sec, 408 bytes true.
consult user state the fact
(same as ?- consult(user).) Or:
?-assert(sunny). statethefact true.
5 / 31

Stating Facts 2
What facts can we describe?
1. Items ?- assert(sunny).
2. Relationships between atoms:
?- assert(likes(john,mary)).
Query the database:
?- likes(john,mary). true.
?- likes(mary,john). false.
?- likes(john,sue). false.
6 / 31

Prolog Terminology
■ Functors : an atom (defined below) with arguments.
■ Arguments can be legal Prolog terms : integer, atom, variable, structure.
■ Atoms: lowercase characters, digits, underscore (not first), graphic
characters (e.g. #,&,@) or anything in quotes.
◆ Legal: hello, hi123, two words, @pl, “G 1)!#)@blah”
◆ Illegal: Hello, 123hi, hello, two-words
■ Variables: Any word beginning with a capital letter.
■ Structures: Functors with a list of arguments.
Structures are also known as relations, compound terms, and predicates.
Like functional languages, variables bind to values (not memory locations). Unlike functional languages, there is no clear notion of input and output.
?- likes(john,Who). Who = mary
Prolog will display one instantiation at a time. Type a semicolon for more.
7 / 31

More Relations
All satisfying likes relations: ?- likes(Who1,Who2). Who1 = john; Who2 = mary
Constrain queries using variables:
?- likes(Who,Who). false.
(People who like themselves.)
Use wild card to determine if some instantiation exists:
?- likes(john,_).
true.
(That is, john likes someone—we don’t care who.)
Wild cards can be used in conjunction with variables:
?- likes(Who,_). Who = john
8 / 31

Rules
Rules express conditional statements about our world. Consider the assertion: “All men are mortal.”
Expressible as modus ponens: human → mortal (“human implies mortal.”) mortal is a goal (or head ), and human is a subgoal (or body ).
In Prolog, we write it in the following form:
mortal ← human. Or more generally,
goal ← subgoal.
There can be multiple subgoals. Example:
goal ← subgoal1, . . . , subgoaln. This form is called a Horn clause.
9 / 31

Rules Example
?- assert(mortal(X) :- human(X)). true.
?- assert(human(socrates)). true.
Now we query:
?- mortal(socrates). true.
You can also ask who is mortal:
?- mortal(X). X = socrates
10 / 31

Closed World Assumption
Prolog relies on everything it is told being true: both facts and rules. e.g., if you tell Prolog the sky is green, it won’t argue with you.
?- assert(sky_color(green)). true.
This is called a closed world assumption.
11 / 31

Conjunction and Disjunction
Conjunction is expressed using commas:
?- fun(X) :- red(X), car(X). A red car.
Disjunction is expressed with semicolons or separate clauses:
?- fun(X) :- red(X); car(X). Something red or a car.
…is the same as
?- fun(X) :- red(X). ?-fun(X):-car(X). Orderofrulesmatters!
Subgoal red(X) will be attempted first, then car(X).
12 / 31

Multi-Variable Rules
daughter(X,Y) :- parent(Y,X), female(X).
grandfather(X,Y) :- male(X), parent(X,Z), parent(Z,Y).
Quantification:
■ Variables appearing in the goal are universally quantified.
■ Variables appearing only in the subgoal are existentially quantified.
The grandfather goal reads as:
∀X,Y∃Z :grandfather(X,Y) ←male(X),parent(X,Z),parent(Z,Y).
13 / 31

Unification
Prolog variables take on values by means of unification. Unification is a
binding between any of the following:
■ A variable and a value.
■ A variable and another variable.
■ A value and a value.
Examples of unification:
■ (Prolog) The rule happy(X) :- go_walking(X) and query happy(joe) (X = joe).
Result: happy(joe) :- go_walking(joe)
■ (ML) Function signature ’a -> ’b -> int -> ’a and input pattern int -> real -> ’c -> ’c.
Result: int -> real -> int -> int
Unification generally happens “behind-the-scenes,” although one may explicitly unify variables in Prolog also.
14 / 31

Explicit Unification
?- a=a.
true.
?- a=b.
false.
?- foo(a,b) = foo(a,b). true.
?- foo(a,X) = foo(a,b). X=b.
?- X=a.
X=a.
?- A=B.
A=B.
?- A=B, A=a, B=Y. A=a; B=a; Y=a.
15 / 31

Unification Algorithm
1. Constants: any constant unifies with itself.
2. Structures: same functor, same arity, arguments unify recursively.
3. Variables: unify with anything.
(a) Value: variable takes on the value.
(b) Another Variable: unify by reference.
Some examples: Operand 1
21
8
Operand 2 Result
X
X
love(X,me)
love(X,Y)
love(X,Y)
c(X,c(Y,c(Z,n)))
love(X,Y)
X=5
X=Y
X=you,Y=me X=you,Y=Y
error
X=he, Y=she, Z=it X=you,Y=??
21 21
15
5
Y
love(you,Y)
love(you,Y)
foobar(you,Y)
c(he, c(she, c(it,n)))
love(you,f(Y))
error
16 / 31

Occurs Check
Consider:
equal(Y, f(Y)).
Let’s try unifying Y=f(Y). We have:
equal(Y, f(Y))
equal(f(Y), f(f(Y)))
equal(f(f(Y)), f(f(f(Y)))) equal(f(f(f(Y))),f(f(f(f(Y))))) nomatch Infinite recursion!
When two expressions are unified and a variable X in one expression “occurs” in the other expression, unsound logic may result.
■ Prolog will not report any warning or error by default.
■ This situation can be caught with an occurs check.
■ The occurs check is not run by default, for performance reasons.
no match no match no match
17 / 31

More on Occurs Check
When attempting to unify variable v and structure s, an occurs check determines whether v is contained within s. If so, unification fails.
■ Prevents infinite loops or unsoundness.
■ Inefficient to implement (linear in the size of the largest term).
■ Most implementations of Prolog (like SWI Prolog) omit it.
Therefore, in SWI Prolog:
?- equal(Y, f(Y)). Y = f(Y).
If you insist on the occurs check, you can force it in SWI:
?- unify_with_occurs_check(X,f(X)). false.
18 / 31

Execution Order
There are two ways to answer a query:
1. Forward chaining: start with given facts and work forward to find a goal.
2. Backward chaining: start with goal and work backward toward the facts.
(Used by Prolog).
If the body of a rule unifies with the heads of other rules in some particular order, it can be expressed as a tree.
■ Forward chaining: most suitable for: many rules, few facts.
◆ Few facts reduces the possible fact combinations, yielding fewer rules
to test
■ Backward chaining: most suitable for: few rules, many facts.
◆ Fewer rules mean fewer possible trees, yielding fewer facts to test
19 / 31

Execution Tree
Consider:
fun(X) :- red(X). red(jeep). fun(X) :- car(X). car(jeep).
red(honda).
?- fun(X).
This produces an AND/OR tree:
red(honda)
red(jeep)
car(jeep)
red(X)
car(X)
fun(X)
20 / 31

Execution Order
Consider:
precip(seattle).
precip(rochester).
cold(seattle).
snowy(X) :- precip(X), cold(X).
?- snowy(X).
precip(seattle)
precip(rochester)
cold(seattle)
precip(X)
cold(X)
snowy(X)
21 / 31

Backtracking
■ Prolog maintains a list of goals to be satisfied.
■ When a goal is queried, all subgoals of the goal are added to the list.
◆ goal(X,Y):-subgoal1(X),subgoal2(Y).
■ Prolog will try to satisfy all subgoals.
■ If a subgoal cannot be satisfied, Prolog will try another way.
■ This is called backtracking.
22 / 31

Backtracking Example
Consider:
precip(seattle).
precip(rochester).
cold(rochester).
snowy(X) :- precip(X), cold(X).
?- snowy(X).
precip(seattle)
precip(rochester)
cold(rochester)
precip(X)
cold(X)
snowy(X)
23 / 31

Backtracking in Prolog
?- precip(seattle). ?- precip(rochester).
?- cold(rochester). ?- snowy(X) :- precip(X), cold(X).
Print the backtrace by invoking trace., then snowy(X).
Call: (6) snowy(_G466) ? creep
Call: (7) precip(_G466) ? creep
Exit: (7) precip(seattle) ? creep
Call: (7) cold(seattle) ? creep
Fail: (7) cold(seattle) ? creep
Redo: (7) precip(_G466) ? creep
Exit: (7) precip(rochester) ? creep
Call: (7) cold(rochester) ? creep
Exit: (7) cold(rochester) ? creep
Exit: (6) snowy(rochester) ? creep
X = rochester
24 / 31

Reflexive Transitive Closure
More than one “application” of a rule:
connect(Node,Node).
connect(N1,N2) :- edge(N1,Link), connect(Link,N2).
Now add some edges:
?- assert(edge(a,b)). ?- assert(edge(c,d)). ?- assert(edge(a,c)). ?- assert(edge(d,e)). ?- assert(edge(b,d)). ?- assert(edge(f,g)).
?- connect(a,e). true.
?- connect(d,f). false.
connect(a,e) :- edge(a,b), connect(b,e)
connect(b,e) :- edge(b,d), connect(d,e)
connect(d,e) :- edge(d,e), connect(e,e)
25 / 31

Lists
Lists are denoted by [ a, b, c ].
A cons pair is denoted [X|Y] where X is the head and Y is the tail.
Rules for testing list membership:
?- assert(member(X, [X|Xs])).
?- assert(member(X, [Y|Ys]) :- member(X,Ys)).
Testing membership:
?- member(b,[a,b,c]). true.
?- member(b,[a,c]). false.
You can also extract list membership:
?- member(X,[a,b,c]). X = a; X = b; X = c.
26 / 31

Reversing Lists
Consider a list reverse rule:
reverse([],[]).
reverse([X|Xs],Zs) :- reverse(Xs,Ys), append(Ys,[X],Zs).
Reverse-accumulate:
reverse(Xs,Ys) :- reverse(Xs,[],Ys).
reverse([X|Xs],Acc,Ys) :- reverse(Xs,[X|Acc],Ys).
reverse([],Ys,Ys).
Invoking the reverse rule:
?- reverse([a,b,c], X).
X = [c, b, a].
?- reverse([a,b,c], [a,c,b]). false.
27 / 31

Tree for Reverse
The reverse rule at work:
reverse([],[])
append([],[c],[c]) append([],[b],[b])
append([],[a],[a])
reverse([c],[c])
append([c],[b],[c,b])
append([b],[a],[b,a])
reverse([b,c],[c,b])
append([c,b],[a],[c,b,a])
reverse([a,b,c],[c,b,a])
28 / 31

Cut Operator
You can tell Prolog to stop backtracking using the cut operator, !.
■ Used to “commit” all unifications up to the point of the !
■ Prevents backtracking through any subgoal to the left of !. Bindings
remain in place.
■ Works across multiple rules (same goal).
Consider:
max(A,B,B) :- A < B. max(A,B,A). ?- max(3,4,M). M=4;M=3(not correct) Try instead: max(A,B,B) :- A < B,!. max(A,B,A). ?- max(3,4,M). M=4 29 / 31 More on Cut The cut operator can also serve as an if-then-else construct: statement :- condition, !, then_part. statement :- else_part. ■ Cut prevents the condition from being retested. ■ If condition is true, subgoal then part will be attempted. ■ If then part fails, the system will not backtrack into the condition. ■ Second rule will not be called if then part is called and fails. ■ If first goal fails (meaning the condition failed), else part will be called. 30 / 31 Negation Definition of not (also known as \+): not(Goal) :- call(Goal), !, fail. not(Goal). ■ Predicate fail unconditionally fails (known as the cut-fail pattern). ■ Predicate call treats the input term as a goal and attempts to satisfy it. Example: female(Person) :- \+ male(Person). Note: A true \+ expression generally indicates inability to prove—not falsehood. (It could be that a particular Person is male, but no such fact was ever asserted.) For the semantics of the not to mean falsehood, the universe of facts must be complete (everything that is true has been asserted accordingly.) 31 / 31