The following is the classic "textbook" vanilla meta-interpreter for prolog.
% simplest meta-interpreter
solve(true) :- !.
solve((A,B)):- !, solve(A), solve(B).
solve(A) :- clause(A,B), solve(B).
The following is simple program which establishes facts two relations which are "positive" and one relation which makes use of negation by failure \+
.
% fruit
fruit(apple).
fruit(orange).
fruit(banana).
% colour
yellow(banana).
% Mary likes all fruit
likes(mary, X) :- fruit(X).
% James likes all fruit, as long as it is yellow
likes(james, X) :- fruit(X), yellow(X).
% Sally likes all fruit, except yellow fruit
likes(sally, X) :- fruit(X), \+ (yellow(X)).
The meta-interpeter can handle goals related to the first two relations ?-solve(likes(mary,X))
and ?- solve(likes(james,X)_
.
However it fails with a goal related to the third relation ?- solve(likes(sally,X)
. The swi-prolog reports a stack limit being reached before the program crashes.
Question 1: What is causing the meta-interpreter to fail? Can it be easily adjusted to cope with the \+
negation? Is this related to the sometimes discussed issue of built-ins not being executed by the vanilla meta-interpreter?
Question 2: Where can I read about the need for those cuts in the vanilla meta-interpreter?
Tracing suggests the goal is being grown endlessly:
clause(\+call(call(call(call(yellow(apple))))),_5488)
Exit:clause(\+call(call(call(call(yellow(apple))))),\+call(call(call(call(call(yellow(apple)))))))
Call:solve(\+call(call(call(call(call(yellow(apple)))))))
Call:clause(\+call(call(call(call(call(yellow(apple)))))),_5508)
Exit:clause(\+call(call(call(call(call(yellow(apple)))))),\+call(call(call(call(call(call(yellow(apple))))))))
Call:solve(\+call(call(call(call(call(call(yellow(apple))))))))
Change solve(A)
into:
solve(Goal) :-
writeln(Goal),
sleep(1),
clause(Goal, Body),
solve(Body).
... and we see:
?- solve_mi(likes(sally,X)).
likes(sally,_8636)
fruit(_8636)
\+yellow(apple)
\+call(yellow(apple))
\+call(call(yellow(apple)))
\+call(call(call(yellow(apple))))
...
clause/2
determines the body of \+yellow(apple)
to be \+call(yellow(apple))
, which is not a simplification.
Can use instead:
solve_mi(true) :-
!.
solve_mi((Goal1, Goal2)):-
!,
solve_mi(Goal1),
solve_mi(Goal2).
solve_mi(\+ Goal) :-
!,
\+ solve_mi(Goal).
solve_mi(Goal) :-
clause(Goal, Body),
solve_mi(Body).
Result in swi-prolog:
?- solve_mi(likes(sally,X)).
X = apple ;
X = orange ;
false.
I'm using solve_mi
because solve
conflicts with e.g. clpBNR, and I'm not using variable names A
and B
because they convey no meaning.
For understanding the cuts, I'd recommend gtrace, to see the unwanted unification with other goals that would otherwise take place.