Pfc -- a package for
forward chaining in Prolog

Tim Finin


Prolog, like most logic programming languages, offers backward chaining as the only reasoning scheme. It is well known that sound and complete reasoning systems can be built using either exclusive backward chaining or exclusive forward chaining. Thus, this is not a theoretical problem. It is also well understood how to ``implement'' forward reasoning using an exclusively backward chaining system and vice versa. Thus, this need not be a practical problem. In fact, many of the logic-based languages developed for AI applications allow one to build systems with both forward and backward chaining rules.

There are, however, some interesting and important issues which need to be addresses in order to provide the Prolog programmer with a practical, efficient, and well integrated facility for forward chaining. This notet describes such a facility, Pfc , which we have implemented in standard Prolog.

A somewhat more complete users manual is available, as is some information on how to get the software and how to use it at UMBC . If you want to cite it, you can use

Finin, Tim, Rich Fritzson and Dave Matuszek, ``Adding Forward Chaining and Truth Maintenance to Prolog'', IEEE Conference on Artificial Intelligence Applications, (pp. 123 - 130), Miami, March 1989.

The basic idea

The Pfc system is a package that provides a forward reasoning capability to be used together with conventional Prolog programs. The Pfc inference rules are Prolog terms which are asserted as facts into the regular Prolog database. For example, here is a file of Pfc rules and facts which are appropriate for the ubiquitous kinship domain.
spouse(X,Y) <=> spouse(Y,X).
spouse(X,Y),gender(X,G1), {otherGender(G1,G2)}
     => gender(Y,G2).
gender(P,male) <=> male(P).
gender(P,female) <=> female(P).
parent(X,Y),female(X) <=> mother(X,Y).
parent(X,Y),parent(Y,Z) => grandparent(X,Z).
grandparent(X,Y),male(X) <=> grandfather(X,Y).
grandparent(X,Y),female(X) <=> grandmother(X,Y).
grandparent(X,Y),female(X) <=> grandmother(X,Y).
parent(X,Y),male(X) <=> father(X,Y).

The Pfc package allows one to define forward chaining rules and to add ordinary Prolog assertions into the database in such a way as to trigger any of the Pfc rules that are satisfied. An example of a simple Pfc rule is:

gender(P,male) => male(P)
This rule states that whenever the fact unifying with gender(P,male) is added to the database, then the fact male(P) is true. If this fact is not already in the database, it will be added. In any case, a record will be made that the validity of the fact male(P) depends, in part, on the validity of this forward chaining rule and the fact which triggered it. To make the example concrete, if we add gender(john,male) , then the fact male(john) will be added to the database unless it was already there.

In order to make this work, it is necessary to use the predicate add/1 rather than assert/1 in order to assert Pfc rules and any facts which might appear in the lhs of a Pfc rule.

Backward-Chaining Pfc Rules

Pfc includes a special kind of backward chaining rule which is used to generate all possible solutions to a goal that is sought in the process of forward chaining. Suppose we wished to define the ancestor relationship as a Pfc rule. This could be done as:
parent(P1,P2) => ancestor(P1,P2).
parent(P1,P2), ancestor(P2,P3) => ancestor(P1,P3).
However, adding these rules will generate a large number of assertions, most of which will never be needed. An alternative is to define the ancestor relationship by way of backward chaining rules which are invoked whenever a particular ancestor relationship is needed. In Pfc, this need arises whenever facts matching the relationship are sought while trying a forward chaining rule.
ancestor(P1,P2) <= {\+var(P1)}, parent(P1,X), ancestor(X,P2).
ancestor(P1,P2) <= {var(P1),\+var(P2)}, parent(X,P2), ancestor(P2,X).


We sometimes want to draw an inference from the absence of some knowledge. For example, we might wish to encode the default rule that a person is assumed to be male unless we have evidence to the contrary:
person(P), ~female(P) => male(P).
A lhs term preceded by a \sim is satisfied only if no fact in the database unifies with it. Again, the Pfc system records a justification for the conclusion which, in this case, states that it depends on the absence of the contradictory evidence. The behavior of this rule is demonstrated in the following dialogue:
?- add(person(P), ~female(P) => male(P)).
?- add(person(alex)).
?- male(alex).
?- add(female(alex)).
?- male(alex)

Procedural Interpretation

Note that the procedural interpretation of a Pfc rule is that the conditions in the lhs are checked from left to right. One advantage to this is that the programmer can chose an order to the conditions in a rule to minimize the number of partial instantiations.

Another advantage is that it allows us to write rules like the following:

   => {remove(at(Obj,Loc1))}.
Although the declarative reading of this rule can be questioned, its procedural interpretation is clear and useful:
If an object is known to be at location Loc1 and an assertion is added that it is at some location Loc2 , distinct from Loc1 , then the assertion that it is at Loc1 should be removed.

The Right Hand Side

The examples seen so far have shown a rules rhs as a single proposition to be ``added'' to the database. The rhs of a Pfc rule has some richness as well. The rhs of a rule is a conjunction of facts to be ``added'' to the database and terms enclosed in brackets which represent conditions/actions which are executed. As a simple example, consider the conclusions we might draw upon learning that one person is the mother of another:
mother(X,Y) =>
As another example, consider a rule which detects bigamists and sends an appropriate warning to the proper authorities:
spouse(X,Y), spouse(X,Z), {Y\==Z} => 
   {format("~N~w is a bigamist, married
      to both ~w and ~w~n",[X,Y,Z])}.
Each element in the rhs of a rule is processed from left to right --- assertions being added to the database with appropriate support and conditions being satisfied. If a condition can not be satisfied, the rest of the rhs is not processed. We would like to allow rules to be expressed as bi-conditional in so far a possible. Thus, an element in the lhs of a rule should have an appropriate meaning on the rhs as well. What meaning should be assigned to the conditional fact construction (e.g. P/Q ) which can occur in a rules lhs? Such a term in the rhs of a rule is interpreted as a conditioned assertion. Thus the assertion P/Q will match a condition P\prime in the lhs of a rule only if P and P\prime unify and the condition Q is satisfiable. For example, consider the rules that says that an object being located at one place is reason to believe that it is not at any other place:
at(X,L1) => not(at(X,L2))/L2\==L1
Note that a conditioned assertion is essentially a Horn clause. We would express this fact in Prolog as the backward chaining rule:
not(at(X,L2)) :- at(X,L1),L1\==L2.
The difference is, of course, that the addition of such a conditioned assertion will trigger forward chaining whereas the assertion of a new backward chaining rule will not.

The Truth Maintenance System

As discussed in the previous section, a forward reasoning system has special needs for some kind of truth maintenance system. The Pfc system has a rather straightforward TMS system which records justifications for each fact deduced by a Pfc rule. Whenever a fact is removed from the database, any justifications in which it plays a part are also removed. The facts that are justified by a removed justification are checked to see if they are still supported by some other justifications. If they are not, then those facts are also removed. Such a TMS system can be relatively expensive to use and is not needed for many applications. Consequently, its use and nature are optional in Pfc and are controlled by the predicate pfcTmsMode/1 . The possible cases are three: A fact is considered to be supported by God if it is found in the database with no visible means of support. That is, if Pfc discovers an assertion in the database that can take part in a forward reasoning step, and that assertion is not supported by either the user or a forward deduction, then a note is added that the assertion is supported by God . This adds additional flexibility in interfacing systems employing Pfc to other Prolog applications.

For some applications, it is useful to be able to justify actions performed in the rhs of a rule. To allow this, Pfc supports the idea of declaring certain actions to be undoable and provides the user with a way of specifying methods to undo those actions. Whenever an action is executed in the rhs of a rule and that action is undoable, then a record is made of the justification for that action. If that justification is later invalidated (e.g. through the retraction of one of its justificees) then the support is checked for the action in the same way as it would be for an assertion. If the action does not have support, then Pfc trys each of the methods it knows to undo the action until one of them succeeds.

In fact, in Pfc , one declares an action as undoable just by defining a method to accomplish the undoing. This is done via the predicate pfcUndo/2 . The predicate pfcUndo(A1,A2) is true if executing A2 is a possible way to undo the execution of A1 . For example, we might want to couple an assertional representation of a set of graph nodes with a graphical display of them through the use of Pfc rules:

at(N,XY) => {displayNode(N,XY)}.
arc(N1,N2) => {displayArc(N1,N2}.


For more information

Contact Tim finin,