Pathfinding in Prolog

Question

I'm trying to teach myself Prolog. Below, I've written some code that I think should return all paths between nodes in an undirected graph... but it doesn't. I'm trying to understand why this particular code doesn't work (which I think differentiates this question from similar Prolog pathfinding posts). I'm running this in SWI-Prolog. Any clues?

% Define a directed graph (nodes may or may not be "room"s; edges are encoded by "leads_to" predicates).
room(kitchen).
room(living_room).
room(den).
room(stairs).
room(hall).
room(bathroom).
room(bedroom1).
room(bedroom2).
room(bedroom3).
room(studio).
leads_to(kitchen, living_room).
leads_to(living_room, stairs).
leads_to(living_room, den).
leads_to(stairs, hall).
leads_to(hall, bedroom1).
leads_to(hall, bedroom2).
leads_to(hall, bedroom3).
leads_to(hall, studio).
leads_to(living_room, outside).  % Note "outside" is the only node that is not a "room"
leads_to(kitchen, outside).

% Define the indirection of the graph.  This is what we'll work with.
neighbor(A,B) :- leads_to(A, B).
neighbor(A,B) :- leads_to(B, A).

Iff A --> B --> C --> D is a loop-free path, then

path(A, D, [B, C])

should be true. I.e., the third argument contains the intermediate nodes.

% Base Rule (R0)
path(X,Y,[]) :- neighbor(X,Y).

% Inductive Rule (R1)
path(X,Y,[Z|P]) :- not(X == Y), neighbor(X,Z), not(member(Z, P)), path(Z,Y,P).

Yet,

?- path(bedroom1, stairs, P).

is false. Why? Shouldn't we get a match to R1 with

X = bedroom1
Y = stairs
Z = hall
P = []

since,

?- neighbor(bedroom1, hall).
true.

?- not(member(hall, [])).
true.

?- path(hall, stairs, []).
true .

?

In fact, if I evaluate

?- path(A, B, P).

I get only the length-1 solutions.

Answer 1

Welcome to Prolog! The problem, essentially, is that when you get to not(member(Z, P)) in R1, P is still a pure variable, because the evaluation hasn't gotten to path(Z, Y, P) to define it yet. One of the surprising yet inspiring things about Prolog is that member(Ground, Var) will generate lists that contain Ground and unify them with Var:

?- member(a, X).
X = [a|_G890] ;
X = [_G889, a|_G893] ;
X = [_G889, _G892, a|_G896] .

This has the confusing side-effect that checking for a value in an uninstantiated list will always succeed, which is why not(member(Z, P)) will always fail, causing R1 to always fail. The fact that you get all the R0 solutions and none of the R1 solutions is a clue that something in R1 is causing it to always fail. After all, we know R0 works.

If you swap these two goals, you'll get the first result you want:

path(X,Y,[Z|P]) :- not(X == Y), neighbor(X,Z), path(Z,Y,P), not(member(Z, P)).

?- path(bedroom1, stairs, P).
P = [hall]

If you ask for another solution, you'll get a stack overflow. This is because after the change we're happily generating solutions with cycles as quickly as possible with path(Z,Y,P), only to discard them post-facto with not(member(Z, P)). (Incidentally, for a slight efficiency gain we can switch to memberchk/2 instead of member/2. Of course doing the wrong thing faster isn't much help. :)

I'd be inclined to convert this to a breadth-first search, which in Prolog would imply adding an "open set" argument to contain solutions you haven't tried yet, and at each node first trying something in the open set and then adding that node's possibilities to the end of the open set. When the open set is extinguished, you've tried every node you could get to. For some path finding problems it's a better solution than depth first search anyway. Another thing you could try is separating the path into a visited and future component, and only checking the visited component. As long as you aren't generating a cycle in the current step, you can be assured you aren't generating one at all, there's no need to worry about future steps.

The way you worded the question leads me to believe you don't want a complete solution, just a hint, so I think this is all you need. Let me know if that's not right.