Artificial Intelligence 53 (1992) 255-272 255

Elsevier

Research Note

Abduction versus closure in

causal theories
Kurt Konolige
Artificial Intelligence Center, SRI International, 333 Ravenswood Avenue, Menlo Park,
CA 94025, USA

Received September 1990

Revised April 1991

Abstract
Konolige, K., Abduction versus closure in causal theories (Research Note), Artificial
Intelligence 53 (1992) 255-272.
There are two distinct formalizations for reasoning from observations to explanations, as in
diagnostic tasks. The consistency based approach treats the task as a deductive one, in
which the explanation is deduced from a background theory and a minimal set of
abnormalities. The abductive method, on the other hand, treats explanations as sentences
that, when added to the background theory, derive the observations. We show that there is
a close connection between these two formalizations in the context of simple causal theories:
domain theories in which a set of sentences are singled out as the explanatorily relevant
causes of observations. There are two main results, which show that (with certain caveats)
the consistency based approach can emulate abductive reasoning by adding closure axioms
to a causal theory; and that abductive techniques can be used in place of the consistency
based method in the domain of logic based diagnosis. It is especially interesting that in the
latter case, the abductive techniques generate only relevant explanations, while diagnoses
may have irrelevant elements,

1. Introduction

Reasoning to t h e b e s t e x p l a n a t i o n is a c o m m o n t a s k in m a n y areas of
a r t i f i c i a l i n t e l l i g e n c e . O n e o f t h e c l e a r e s t e x a m p l e s is d i a g n o s i s , in w h i c h o n e
r e a s o n s f r o m o b s e r v a t i o n s s u c h as p a t i e n t s y m p t o m s to t h e i r u n d e r l y i n g c a u s e s ,
a disease or physiological malfunction. In the literature, there are two fun-
damentally d i f f e r e n t f o r m a l i z a t i o n s o f this t a s k [9, 11]. I n o n e , t h e p r o c e s s o f
f i n d i n g a c a u s e is t r e a t e d as a s t r a i g h t f o r w a r d a b d u c t i v e t a s k . R e p r e s e n t a t i v e o f

0004-3702/92/$05.00 ~) 1 9 9 2 - Elsevier Science Publishers B.V. All rights reserved

256 K. Konolige

this approach is the set-covering model of diagnosis [12], which assumes two
disjoint sets, d a set of disorders, and m a set of manifestations. Disorders are
assumed to "cause" manifestations, represented by a relation d x m. The
problem of diagnosis is recast as the problem of finding a minimal cover for
observed manifestations m' C m, that is, a minimal subset of d that causes m'.
The competing formalization, the consistency based approach, is best repre-
sented by Reiter's logic based theory of diagnosis [14]. In this theory, the
functionality of a system containing a finite number of components is character-
ized by a set of first-order sentences, the domain theory. The special predicate
ab(c) is used to state that the component c is abnormal or not functioning
correctly. The observed behavior of the system is represented by a set of
sentences. A diagnosis of the behavior is a minimal set of abnormality
assumptions that is consistent with the observations and the domain theory.
These two formalizations seem fundamentally different. The abductive ap-
proach looks for a set of causes that will imply the observations; the con-
sistency based approach looks for a set of abnormality assumptions that are
consistent with the observations. Nevertheless there is a connection between
the two: Reiter showed how to express the set-covering model within his
framework. Recently, Console [2] and Poole [9] have shown that either
formalization can be used in restricted settings to compute the same explana-
tions for diagnostic tasks. In the abductive framework, the domain theory has
axioms that relate causes and their effects, e.g., c i D e would be used to say
that the effect e is a result of cause q . A corresponding consistency based
theory is created by adding closure axioms stating that the only way to achieve
an effect is by the set of causes given (e ~ c~ v c 2 v • • "). The closure axioms
are local in that they are easily derived by looking at all the implications that
have a c o m m o n head atom. The explanations computed by the two methods
are the same, as long as the domain theory contains just Horn clause
implications from causes to effects, and is acyclic.
This result applies to diagnostic tasks that require explanations, that is, the
unexpected observations must be predicted or explained from the assumed
malfunctions. In the literature, explanatory diagnosis is usually signalled by the
presence of fault models [4, 15]. Reiter's framework may also be used for a
weaker form of diagnosis, which could be called excusing diagnosis: identify
components that, if malfunctioning, would cancel or excuse predicted normal
behavior of the system that conflicts with the observations. Here we look only
at the case of explanatory diagnosis (and causal explanation in general), since
excusing diagnosis has no analog in the abductive framework.
The restrictions on the domain theory for the Console/Poole result are very
tight; in particular, there can be no correlation information (e.g., that two
causes are mutually exclusive, or that one effect is the negation of another) or
uncertainty (e.g., a cause implying a disjunction of effects). In this paper we
will examine the connection between abduction and closure in the setting of
 Abduction versus closure in causal theories 257

explanation in general causal models, allowing correlations, uncertainty, and

acyclicity in the causal structure. We answer the following questions:
• Is there a notion of explanatory closure that is appropriate for the more
general domain theory? Is there an equivalent local closure?
• Is consistent explanatory closure of a general domain theory possible?
• When consistent closure is possible, does minimization of causes in the
closed theory compute the same explanations as does abduction in the
original theory?

There are both positive and negative results. With an appropriate notion of
explanatory closure, given certain technical conditions, the consistency based
approach will compute the same explanations as the abductive approach.
However, the utility of the former method is open to question, since local
closure will no longer suffice for explanatory closure; there seems to be no way
to close the domain theory other than by computing all explanations. Further,
the consistency based method is strictly stronger than the abductive one in
explanatory diagnostic tasks, and the answers it produces may have elements
that are not relevant to a causal explanation.
A second area that we address is whether abductive methods may be used in
the setting of logic based diagnosis with fault models. This area is closely
related to the previous one, except that we assume that there is already a
closed theory, and that the causes take on a specific form, namely normality
and abnormality assumptions about components. Our main result here is that
the abductive method produces kernel diagnoses, ~ but without any of the
irrelevant causes that may be present in the latter.
The next three sections of this note describe simple causal theories, and
define abductive and consistency based methods in this context. Section 5
develops the concept of explanatory closures, and Section 6 gives the main
results on emulating abduction with the consistency based method. Section 7
describes how abductive methods can perform logic based diagnosis.

2. Simple causal theories

We are interested in domains in which there is a concept of cause and effect.

Much of our commonsense view of the world can be cast into this form.
Typical here is reasoning about actions or events and their results, usually
formalized in the situation calculus or some variant [7]. Other domains include
medical diagnosis with diseases as causes, symptoms as effects; mechanical or
electrical systems with components and inputs as causes, outputs as effects; and
planning domains with plans as causes, actions as effects.

i This term is from [3], and is defined in Section 7.

258 K. Konolige

While there is a great deal of complexity and controversy in defining

causation, for this p a p e r most of these problems can be bypassed because we
are interested in a formal representation of the simplest aspects of causal
consequence, given by the following definition.

Definition 1. Let 5f be a first-order language. A simple causal theory is a tuple

( C, E, Z > where
• C, a set of atomic sentences of ~ , are the causes;
• E, a set of sentences of oT, are the effects;
• X, a set of sentences of ~q, is the domain theory.

T h e set C contains those atomic propositions which represent the possible

causative agents of the domain. If we are looking for an answer to the question
of " w h a t caused e ? " , then an acceptable answer is the conjunction of some
subset of C. 2
Effects E are those aspects of the domain that we might observe, and for
which we want to know the cause. Note that E and C need not be disjoint; an
observed cause may require no further explanation.
T h e domain theory 2 contains information about the relation between causes
and effects. For example, in the situation calculus we might take C to be
occurrences of actions, E to be properties of the final state, and 2f to hold
information about the initial state and the way in which actions affect prop-
erties of situations.
H e r e is a simple causal theory that will be used as an example in the rest of
the paper; a graphical presentation appears in Fig. 1. The intended meaning of
the predicates should be obvious from their names.

Causes: rain, sun, warm, sprinkler ;

Effects: wet-lawn, wet-road ;
D o m a i n theory: rain D wet-road, rain D wet-lawn, sun =-Train
sprinkler D wet-lawn, sun/x warm D sprinkler.

wet-lawn wet-road

wa,rIl] Slln -'- - / - -,- rail]

Fig. 1. A sample causal theory.

2 Allowing only atoms simplifies the analysis, but is not restrictive, since we can include
equivalences such as c --- ~, where ~b is a complex sentence.
 Abduction versus closure in causal theories 259

A notational convention: a finite set of sentences will often be taken as a

conjunction, e.g., if A and B are such sets, we write

AvB for (alAa2A''')V(blAb2A''") ,

7A for -~(a I A a 2 / x . . . ) .

3. The abductive approach

Given a simple causal theory, the problem of reasoning from observations to

causes can be expressed formally using abduction. The account of logical
abduction we give here draws on ideas already present in the literature (e.g.,
[9]).

Definition 2. Let (C, E, ~ ) be a simple causal theory. An abductive explana-

tion (or ABE) of a set of observations O C_ E is a finite set A _C C such that
• A is consistent with ~ ;
• .~ U A~-O;
• A is subset-minimal over sets satisfying the first two conditions.
If O has a nonzero finite number of ABE's then the cautious explanation is
their disjunction: Vi Ai.

Remarks. A must be a minimal set of members of C; by minimal is meant that

there is no other set of causes for O consistent with the domain theory that is a
proper subset. This is a relevancy condition, since it excludes from the
explanation 3 elements of C that are not relevant to deriving O. Other than this
we say nothing about preferences among multiple explanations. It is obvious
that often such preferences will be required for reasoning, e.g., we may want
the most specific explanation, or the most normal (where we partition causes
into ones that normally occur and ones that do not), or the X-est, where X is
some measure on explanations. The preference could be expressed mathemati-
cally by a partial order on the subsets of C. Since such an order will be closely
related to the domain of application, and we have no way of making any
general statements about the order, we omit it from further consideration here.
In a given problem domain, we may be interested in the best explanation, or
the cautious explanation, or even any (satisficing) explanation. For example, if
we want to predict the possible states of the world after a sequence of events,
then the cautious explanation might be most appropriate, while tasks like plan
recognition usually require the best explanation. And for some problems there
is no ordering of solutions, and any one would be acceptable.
3A note on terminology: we will use the simple term explanation to refer to abductive
explanations when no ambiguity is possible, and abductive explanation or ABE when we want to
distinguish them from explanations derived by the consistency based approach.
260 K. K o n o l i g e

Finally, it is possible that one explanation A will imply another A' in a

simple causal theory. For example, sun and warm implies sprinkler in the
sample theory. More generally, let A~ v A~ v . . . v A,, be any disjunction of
explanations for O; A j is independent for 0 in the theory Z if Z U A ~~ A: v
A 3 V "'" V A n.
Using the example causal theory of the previous section, there are three
explanations of O --- {wet-lawn}, namely {rain}, {sprinkler}, and {sun, warm}.
The cautious explanation is rain v sprinkler v (sun ^ warm), which simplifies
in the domain theory to rain v sprinkler. The explanation sun/x warm is not
independent, since it implies sprinkler. The observation set O = {wet-
road, wet-lawn} has the single explanation {rain}, which is also its cautious
explanation.

4. The consistency based approach

The consistency based approach has been most clearly developed in the
domain of diagnosis, especially in [3, 14]. In this section we will modify the
terminology slightly to apply to the more general causal theories of Section 2,
and to make comparison to the abductive approach easier. The particulars of
the diagnostic task are discussed later, in Section 7.

Definition 3. Let (C, E, ~ ) be a simple causal theory, and O (the observa-

tions) a subset of E. A denial set for 0 is a subset D C C such that

0 0 0 {-Td [ d E D} is consistent.

When a denial set is maximal (that is, there is no other denial set that
contains it), no more negative causes can be consistently added to it, and thus it
is possible to deduce a set of (positive) causes from the maximal denial set, the
domain theory, and the observations [14]:

,~ O 0 0 { - T d ] d ~ D } ] - C - D. (1)

In this case, we call C - D a consistency based explanation for O, or CBE for

short. CBE's are called diagnoses in Reiter's original paper, and they were
used mainly for producing excusing diagnoses in the domain of electronic
circuits. However, Reiter recognized that the consistency based approach was
also capable of producing explanatory diagnoses, as long as the domain theory
contained implications from effects to causes. For example, in reconstructing
the set-covering model of (explanatory) diagnosis, Reiter used axioms of the
form:
O B S E R V E D ( m ) 3 PRESENT(d 1) v . . . v PRESENT(dn) ,

where m is the observed symptom and d i are diseases that cause the symptom.
 Abduction versus closure in causal theories 261

These axioms give necessary conditions for the observation, namely, that one
of a set of diseases be present. Later papers refer to implications of this sort as
fault models [4, 15].
The difference between the consistency based approach and abduction is
twofold. First, the form of inference is distinct: rather than abducing causes
that imply the observations O given the domain theory 27, the consistency
approach tries to minimize the extent of the causation set C by denying as
many of its elements as possible. Second, these methods encode knowledge of
the domain differently: in the abductive framework, there are implications
from the causes to the effects, while in the consistency based systems, if we
want to derive explanatory rather than excusing diagnoses, the most important
information seems to be the implication from observations to possible causes.
Despite these differences, it is known that, under certain conditions on the
domain theory, abductive and consistency based explanations coincide.

Theorem 4 (Console, Poole). 4 Let (C, E, 2f) be a simple causal theory over a
propositional language, with ,~ a set of nonatomic definite clauses whose directed
graph is acyclic. Let C be a set of atoms that do not appear in the head of any
clause of 2f, and E any set of atoms. Let 17 be the Clark completion [1] of ~,.
Then the CBEs of ( C, E, 17 ) are exactly the ABEs of ( C, E, 27 ).

The simple causal theory of Fig. 1 does not satisfy the conditions, because it
contains the equivalence sun -~ rain, and sprinkler is a cause that appears as the
head of a clause. If we eliminate these anomalies, then the Clark completion of
the domain theory is:
wet-road =--rain,
wet-lawn - rain v sprinkler, (2)
sprinkler-~ warm ^ sun.
The CBEs of wet-lawn are {rain} and {sun, warm}; the ABE {sprinkler} is
missing.
For more complicated domain theories, Clark completion does not give the
required closure over abductive explanations. If the theory has cycles, for
example {a ~ b, a' ~ b, b ~ a}, then the completion will only pick out a subset
of the abductive explanations (in this case, b ~ a). If there is disjunction in the
head of a clause, the completion is undefined.
In the next sections we will extend the scope of Theorem 4 by considering a
more general notion of completion for a simple causal theory, that of explana-
tory closures.
4 Neither of these authors states the t h e o r e m in this form, although Poole ll0] is close. It is clear
that the t h e o r e m follows from their results. Poole's t h e o r e m as stated s e e m s to have a broader
application, but personal correspondence with him disclosed that the conditions of application are
as given here.
262 K. Konolige

5. Explanatory closures

Let ( C, E, ~ ) be a simple causal theory, and suppose g E E has a cautious

explanation V i A i. Now consider the statement

g D A I v A 2 V " " " V A. , (3)

where we understand each Ai to be the conjunction of its elements. This
expression says that whenever g is present, it must have been caused by one of
the A~; we call this expression the explanatory closure of g with respect to the
simple causal theory (C, E, ~ ) ; it is abbreviated T ( g ) . If the explanatory
closures of all effects E exist, then the theory ( C, E, H ) formed by adding the
closures to ~ is called the closure of (C, E, ~ ).
By forming the closure of a causal theory we can deduce the cautious
explanation from any given effect. One immediate question is whether we
should add something stronger or weaker to close the theory. If we add a
stronger closure, then we have excluded some original abductive explanation
from consideration; e.g., if the explanatory closure is g D a~ v a 2, and we use
g D a~ instead, then a~ will be the only CBE for g. On the other hand, suppose
instead we add something weaker, e.g., g D (a 1 v a 2 v 6) for some arbitrary
sentence ~. If we try to derive CBEs by minimizing causes, then since
~ ( a l v a2) is consistent with the closure, we could assume it, and derive 6 as
the "explanation" for g, which is certainly not intended.
A n o t h e r question is whether explanatory closures are always consistent with
the original causal theory, and if so, whether the original abductive explana-
tions remain unchanged. Unfortunately, the answer to both parts of this
question is " n o " .

Example 5. Let ( { a l , a 2, a3}, {g~, g2, g3}, ~ ) be a simple causal theory, with
2: equal to the conjunction of

a~ A a 2 ~ g l , a~ A a 2 ~ g 2 v g 3 , -qa~ v -~a 2 v ~a 3 ,

a z A a 3 D g2, a 2 /x a 3 D gl v g3 , g~ v g2 v g3 ,

a 3 m a I D g3 , a3 ^ al D gl v g2 •

The closures of this theory are

gl D a I A a 2 ,

g2 ~ a2 A a 3 ,

g3 ~ a3 A a 1 .
 Abduction versus closure in causal theories 263

It is easy to show that the c o n j u n c t i o n o f these closures is inconsistent with ~.5

T h e technical conditions for inconsistency are s o m e w h a t complicated, and it
takes s o m e w o r k to create a causal t h e o r y that will have inconsistent closures;
e.g., the e x a m p l e of Fig. 1 can be consistently closed, but it was not originally
d e s i g n e d with this p r o p e r t y in mind. T h e necessary conditions involve interact-
ing effects and causes such that in the causal t h e o r y at least one of the effects is
true, and one o f the causes false. T h e following proposition states this m o r e
precisely.

Proposition 6. Let {'y(gi) I 0 < i ~ n) be a set o f closures f o r ( C, E, 2f >. For

each i <~ n, let Pi be either gi or -7 A i, where A ~ is any abductive explanation f o r
g~. Each sentence

Pl v P 2 V "'" vpn

m u s t be a theorem o f ,~ f o r these closures to be inconsistent with ,Y.

Proof. F o r the closures to be inconsistent, V i ~ ( 7 ( g ~ ) ) must be a t h e o r e m of

~. We have:

V, V,(g, ^
=- V i ( g i A 7A~ ^ -TA~...)
where Eg i is the cautious explanation for gi, and the A is are all abductive
explanations for it. T h e p r o p o s i t i o n follows by tautological c o n s e q u e n c e . []

W e n o w turn to the question o f h o w adding closures can m o d i f y abductive

explanations.

E x a m p l e 7. Let ( { a l , a3, a4} , {gl, g2, g3}, ~ > be a simple causal t h e o r y , with
equal to the c o n j u n c t i o n o f

a I 3 gl , 7a 1 v -la 3 ,

al D g2 , aa D gl v g 2 ,
aa ~ g3 ,

a4 ~ g3 •
T h e closures of this t h e o r y are

5 It was suggested by a reviewer that the sentence gl D (a 1 ^ a2) V (a 2 ^ a 3 ^ "lg3) V (al ^ a 3 A

"7g2) and similar ones for the other effects be used; these closures are consistent with the domain
theory. However, as noted above, this would generate an anomalous explanation: by asserting
-7a~, we derive a2 ^ a 3 A--Ig3, which, although it derives g~ in the domain theory, is not an
abductive explanation for g~.
264 K. Konolige

gl~al ,

g2 Da~ ,
g3 ~ a3 v a 4 .
If the first two closures are added to ~,, a I becomes true, a 3 becomes false, and
the only explanation for g3 is a 4.

This example shows that some causes may b e c o m e true or false, thus
modifying the available abductive explanations. However, no truly " n e w "
explanations are created by the addition of closures, since every explanation
must be the subset of one of the original ones.

Proposition 8. Let ( C , E, X ) be a simple causal theory, and {y(gi)} a set of

explanatory closures with respect to it. Suppose I1 = ~ U { y(gi) } is a consistent
set. For an arbitrary effect g, every abductive explanation of g w.r.t. 11 is a
subset of some abductive explanation of g w.r.t. ~.

Proof. A s s u m e A is an explanation for g w.r.t. H, but there is no A' _D A such

that A' is an explanation for g w.r.t. Y~. Using a technique similar to that of
Proposition 6, the following must be theorems of ~, where each p~ is either gi
or -1 A~ for any explanation A~ of g~:

Pl v p 2 v "'" vp,, v-aAvg.

Choosing each p~ to be -nA~, this is a sentence which contradicts the original

assumption. []

5.1. Augmented domain theories

R a t h e r than trying to determine if a causal theory has a consistent closure,

we might find it useful to modify the theory so that it does. The simplest way to
do this is to add an escape cause for each effect: a new cause r~ is included in C
for each gi, and the sentence r~ D g i is added to y,.6 The new causes are
sufficiently isolated from the original domain theory so that inconsistency
cannot result. In effect, the closure conditions no longer force one of the
original abductive explanations for g~ to be true, since ri is an alternative.
Further, a u g m e n t e d theories do not change their original abductive explana-
tions at all when closures are added.

Proposition 9. Let (C', E, ~ ' ) be a simple causal theory formed from

( C,E,~ ) by adding r i to C and q D g~ to X for each g ~ E; call this an
a u g m e n t e d causal theory. Suppose that { y(gi) } is a set of explanatory closures

Escape causes are the same idea as the unknown faults of [4, 15].
 Abduction versus closure in causal theories 265

with respect to the augmented theory, and let H = ~ ' U (T(gi)}. Then 171 is
consistent, and for an arbitrary effect g, a subset A C D is an abductive
explanation o f g w.r.t. H if and only if it is an abductive explanation o f g w. r. t.
2f.

Proof. By Proposition 6, if the closure of ,~' is to be inconsistent, the following

must be theorems of ,~ ', where each Pi is either gi or 7 ri:

plvpEv'"vp,.
Because the only expressions containing r~ are of the form r~ D gi, the above
sentences are theorems of ~ ' only if there are corresponding theorems of
with each 7r~ replaced by -Tg~. This is impossible, since such a set is
unsatisfiable.
Assume A C_ C is an abductive explanation for g w.r.t. F/, but not w.r.t. ~.
By reasoning similar to that in the proof of Proposition 8, the following must
be theorems of ~ ' , where each p~ is either g~ or -Tg~:

Pl VPz v . . . vpn v-TA v g .

By tautological consequence, these sentences imply A D g, contradicting the
initial assumption. []

5.2. Local closure

Cautious explanations for a proposition g are defined by reference to the

entire contents of the causal theory ,~. Is there a way of deriving these
explanations in a local manner, that is, by looking only at the sentences of ,~ in
which g occurs? From Theorem 4, Clark completion works for a restricted
language. But if arbitrary correlations are allowed in ,~, then adding cautious
explanations by a local closure operation is not possible. The simplest example
showing this contains loops in the implication structure; e.g., let ~ be

aDg, a'Db, bDg,

gDc, cDb. (4)

Let a and a' be the causes. Adding the local closure g D a v b is insufficient,
because it is subsumed by g D c D b, so that a as a cause of g will never be
inferred. Any local closure for g cannot find the connection between c and b,
and thus has the chance of being incorrect.
Loops in the implication structure also cause problems for other global
closure methods such as circumscription, which is equivalent to Clark comple-
tion for the restricted language [13]. In the case of the above example,
minimizing g while holding the causes fixed yields g 3 b, which is again
stronger than the explanatory closure.
266 K. Konolige

6. Closure + minimization implies abduction

T h e closure of a causal theory contains the explanatory closure

gDA 1 vA, v'"vA,,

of each effect g. Suppose the closed theory is consistent, and we observe g.
T h e n A 1 v A 2 v • • • v A,, is true in all models of g and the closed theory. If we
now try to minimize causes, that is, to assert -~A i for as m a n y abductive
explanations as possible, we will eliminate possible explanations from the
disjunction, until we are left with a single one. Thus we can p e r f o r m abductive
reasoning in the consistency based approach.
T h e r e is one caveat to this reasoning: 7 if an abductive explanation A1 is not
independent, then it will not be found by closure and minimization. Suppose
there is another A 2 that is implied by A~ and the domain theory; then A~ will
be shadowed f r o m the minimization by A2: we cannot assert - q A 2 without
concluding ~ A ~ . Thus using closure and minimization will only produce the
independent abductive explanations.
This discussion is m a d e more precise with the following theorem.

Theorem 10. Let (C, E, ,~ ) be a simple causal theory, and suppose that
( C, E, 1I), its closure, is consistent and does not entail an effect g. Let A be an
A B E for g in ~, and suppose that A is consistent with 11 and independent in I1.
Then A is a subset (not necessarily proper) of some CBE for g in II.
Conversely, every CBE for g in 11 is a superset (not necessarily proper) of
some A B E for g in Z.

Proof. Suppose A is an A B E of g (in X ) , and let X be the disjunction of the

rest of the A B E s of g: A 2 v A 3 v • • • v A n. -nX is consistent with H, or else
H ~ X and so H ~ g, contradicting the assumptions. Also, by assumption,
H U A is consistent, and since A is independent in H, / / U A U - n X is
consistent, and hence so is H U {g} U - n X . Let m be a model of H U {g} U
~ X , and let D = {a2, a 3. . . . . a,} be a set of elements, one from each A B E of
X, that are false in m. Let - - D = (-ha2, -ha 3. . . . . -nan). Now H U {g} U - D is
consistent, and because of the presence of the closure of g, A is a consequence
of it. D can be extended to some maximal set D ' that is a denial set for g, and
its c o m p l e m e n t w.r.t. C contains A.
For the converse part, let H U {g} U ~ D be consistent for some maximal
denial set D C_ C. Suppose the associated C B E H is not a superset of any A B E
of g in Z. T h e n for any A B E A i of g, ~ A i is consistent with H U {g} U - D ,
and so D by maximality must contain some element of each of A i. Thus

7 I am indebted to Eunok Paek for pointing out this problem.

 Abduction versus closure in causal theories 267

D D -7 Eg, where Eg is the cautious explanation for g. This is a contradiction,

since g 3 Eg is a sentence of H. []

Remarks. This theorem shows the general correspondence between abductive

and consistency methods. If an inverse causal theory is formed by closing a
causal theory, then, with several restrictions, consistency based and abductive
explanations are isomorphic to one another. The restrictions have to do with
the problems encountered in adding closures to a causal theory; given the
results of the last section, it may not be possible to do so consistently, as
abductive explanations may change, and so forth.
This not to say that the two approaches are equivalent, however. The
consistency based method in general entails more than the abductive one, as a
consequence of adding the closures.

Corollary 11. A s s u m e the same conditions as in Theorem 10 above. For some

maximal denial set D, every consequence o f 2f and A is a consequence o f H and
D. On the other hand, some consequence o f D and 1I may not be a consequence
o f ~ and A .

In Example 7, --7a3 is a consequence of H, but not of any abductive

explanation for g3.

6.1. Representational issues

The designer of a domain theory who wishes to employ the consistency

based approach can use the results of the previous sections to help determine
how to formalize the domain. If the closure conditions are given directly as
part of the available knowledge of the domain, then the consistency based
approach may be used to generate explanations. For example, in the system of
Morgenstern on temporal projection [8], the axiomatization gives a closed
causal theory, since it specifies exactly what events must occur given a
sequence of states, and vice versa.
On the other hand, often the designer only has information about causal
effects, together with some noncausal correlations (e.g., forbidden states). In
order to employ the consistency based approach, the explanatory closures must
be generated and added. Here the form of the causation axioms can be
exploited. If they are Horn, definite and acyclic, then local closure (Clark
completion) can be used. For more complicated theories, a technique such as
circumscription may be appropriate. A good example from the domain of
temporal projection is the work of Lifschitz [6]; in effect, this theory is similar
to that of Morgenstern above, with the following differences. First, the
sequence of actions is fully specified by the result function, but exceptions to
the actions are allowed, in the form of miracles: these are the assumable
268 K. Konolige

atoms. Second, there is a theory of causation for action types, which is used to
generate the closure conditions by circumscription. That is, the causation
axioms state what must follow if the preconditions of an action hold and the
action takes place; circumscription then generates the closure axioms. Minimi-
zation over miracles gives the desired explanations.
Another good example of the derivation of closure axioms by circumscriptive
techniques is Kautz's theory of plan recognition [5]. The domain theory is a
hierarchical set of actions; the causes are the goals at the highest level of the
hierarchy, the so-called E N D events. Relations between actions at different
levels in the hierarchy are given by a first-order domain theory. Circumscrip-
tion is used to close off the axioms, producing the explanatory closure axioms.
Given a set of observed actions, minimizing over the E N D event produces an
explanation of the observations. ~
The motivation behind the multiple circumscription in systems such as [5, 6]
has often been obscure. Given the results of this paper, it should be clear that
the circumscriptions are performing abduction by using closure and minimiza-
tion. Whether the circumscription corresponds to an appropriate closure can be
tested by checking whether it produces the explanatory closure axioms, and
whether adding these axioms changes the set of abductive explanations. The
examples and propositions of Section 5 should be helpful in this regard; for
example, by adding escape causes it is always possible to retain the original
causal structure. In general, if there are cycles in the implication structure of
the causal domain theory, then neither circumscription nor local closure will
work correctly in generating explanatory closures.

7. Logic based diagnosis

In the previous sections we considered how it was possible to derive

explanations in a causal theory using consistency based methods. Here we
consider the converse question in a more particular setting: can abductive
methods be used to perform logic based diagnosis? In logic based diagnosis,
the domain theory takes on a restricted form, with a distinguished set of
abnormality predicates a b i used to describe the expected behavior of a system.
For example, consider the double inverter of Fig. 2. The proposition in i means

Fig. 2. A double inverter.

T h e accounts of these systems are of necessity somewhat simplified, but the basic structure is
correct.
 Abduction versus closure in causal theories 269

that the input of device i is a logical 1, and ~ i n i that it is a 0 (similarly for out i
and output). The domain axioms are

-lab i D (in i =- --]outi) ,

ab i D (in i =- outi), (5)

out a =- in b
for i = a, b. Each inverter can either have normal behavior, or have a short
circuit so that input and output are the same. In this example, both normal and
abnormal behaviors of the system are fully specified, that is, there is an
exhaustive fault model. It has been recognized that exhaustive fault models are
required for explanatory diagnosis; in our terms, the domain theory contains
the explanatory closures for all possible input/output behaviors.
In the most recent formulation of logic based diagnosis [3], the presence of
exhaustive fault models has made it necessary to modify Reiter's original
definition of diagnoses. Let {C, E, 2 ) be a simple causal theory, with C
containing a set of abnormality predicates and their negations. 9 In the example
above, C = {ab a, abb, ~ a b a, -labb}.
A partial diagnosis G for an observation set O C_ E is a subset of C that is
consistent with ,~ and O, such that every noncontradictory way of extending G
with elements of C is also consistent with 2~ and O. For example, for
O = { i n a , OUtb}, there are two partial diagnoses, {-qaba,-qabb} and
{aba, abb}. Only the first of these would be considered as a diagnosis in the
original theory.
A kernel diagnosis is a subset-minimal partial diagnosis. For O = {inb, O U t b },
there are three partial diagnoses, {abb}, {abb, aba}, and {abb, 7abe}; the first
of these is subset-minimal, and so is a kernel diagnosis. Note that kernel
diagnoses eliminate some of the irrelevancy present in partial diagnosis, since
the state of inverter a is not relevant to the observed behavior.
We would like to know if we can produce the same kernel diagnoses with the
abductive approach. The following theorem shows that this is possible, except
(as in the case of Theorem 10) that the ABEs are, in general, more compact
than kernel diagnoses.

Theorem 12. Let (C, E, 2 } be a simple causal theory, with C a set of

abnormality predicates and their negations, and ~ closed and consistent. Let A
be an independent A B E for g in ~. Then A is a subset (not necessarily proper)
of some kernel diagnosis for g.
Conversely, every kernel diagnosis for g is a superset (not necessarily proper)
of some A B E for g.

9We could also change the vocabulary and add ok i =-~ab i as a new set of causes. But we are
trying to stay as close as possible to the logic based diagnosis terminology.
270 K, Konolige

,' 1

Fig. 3. An unconnected circuit.

Proof. Let A be an independent A B E for g. Using an argument similar to the

first part of the proof of T h e o r e m 10, construct the set D, such that ~ U { g} U
D ~- A. Now - - D U A is a subset of some partial diagnosis, and therefore A
must be a subset of some kernel diagnosis, since X U - D U -7 A is inconsistent.
For the converse part, let K be a kernel diagnosis for g. Assume no A B E A i
is contained in K. Then K can be consistently extended by adding one negated
element from each A i, thus contradicting g D Eg, which must be in X. []

In general, although kernel diagnoses get rid of some irrelevancies with

respect to their corresponding ABEs, they still assume more than is really
needed for an explanation: hence the necessity of subset/superset relations in
the correspondence theorem. As an example, consider the three unconnected
inverters of Fig. 3. Suppose the basic axioms governing the inverters are the
same as before; there are no connections between the inverters, and the faults
are coupled by the axiom ab~ D (ab b v abc). If we observe {in,, out,}, there
are two kernel diagnoses, {aba, abb} and {abe, abe}. There is only one A B E ,
namely {aba}. The kernel diagnoses must reflect the constraint that one of b or
c is abnormal, which is not a relevant factor in explaining the observations. The
abductive explanation is just the set of causes that account for the observed
behavior. The abductive approach distinguishes between direct causes of the
observations and irrelevant causes while the consistency based approach does
not. On the other hand, the information that either b or c is abnormal is still
derivable in the abductive system, as a consequence of the domain theory and
the A B E ; but it is not a part of the explanation itself,

8. Conclusion

We have shown how to extend the correspondence between abductive and

consistency based methods to the case of causal theories that have arbitrary
first-order relations between causes and effects. The correspondence requires
that a domain theory expressing how causes produce effects be closed, that is,
 Abduction versus closure in causal theories 271

contain statements that the only causes are the known ones. The appropriate
closure axioms are identified in this paper as explanatory closures. The main
result of the paper is that minimization of causes in the closed theory produces
almost the same explanations as abduction in the original causal theory. The
caveat is that the abductive explanations are generally weaker than their
consistency based counterparts. There are two reasons for this: adding closures
may change the available abductive explanations; and the consistency based
method can conclude causes that are intuitively irrelevant to the observed
behavior.
If one is interested in the representation of domain knowledge, then the
abductive approach offers several advantages. It does not require the assump-
tion of complete knowledge of causation, and it is not necessary to assert the
explanatory closures. Adding the closures can lead to inconsistency and change
the available abductive explanations (although it will not add new ones). The
computational aspect of adding closures is also discouraging, since there is no
general local method that accomplishes the addition. Stronger global methods
such as circumscription will work only in special circumstances.
In logic based diagnosis, using an abductive method may be appropriate if it
is important to distinguish causes relevant to producing the observations from
those that are mere side-effects. But one must be careful here in giving too
much weight to the term "causally relevant", since the simple causal theories
we have introduced give only a very narrow interpretation of the complex
concept of causation.

Acknowledgement

I would like to thank David Poole, Eunok Paek, Oskar Dressier, and
Nicolas Helft for many helpful discussions on the evolving draft. The two
anonymous referees contributed many useful comments that I have used in
revising the paper. The research reported here was supported partially by the
NTT Corporation, and partially by the Office of Naval Research under
Contract No. N00014-89-C-0095.

References

[1] K. Clark, Negation as failure, in: Logic and Data Bases (Plenum, New York, 1978) 293-322.
[2] L. Console, D.T. Dupre, and P. Torasso, Abductive reasoning through direct deduction from
completed domain models, in: Z.W. Ras, ed., Methodologies for Intelligent Systems 4
(North-Holland, New York, 1988) 175-182.
[3] J. de Kleer, A. Mackworth, and R. Reiter, Characterizing diagnoses, in: Proceedings
AAAI-90, Boston, MA (1990).
[4] J. de Kleer and B.C. Williams, Diagnosis with behavioral modes, in: Proceedings IJCAI-89,
Detroit, MI (1989).
272 K. Konolige

[5] H. Kautz, A formal theory for plan recognition, Tech. Report TR-215, University of
Rochester (1987).
[6] V. Lifschitz and A. Rabinov, Miracles in formal theories of action, ArtiJ~ lntell. 38 (2) (1989)
225 -237.
[7] J. McCarty, First order theories of individual concepts and propositions, in: B. Meltzer and
D. Michie, eds., Machine Intelligence 9 (Edinburgh University Press, Edinburgh, 1979)
120-147.
[8] L. Morgenstern and L.A. Stein, Why things go wrong: A formal theory of causal reasoning,
in: Proceedings AAAI-88, St. Paul, MN (1988) 518-523.
[9] D. Poole, Representing knowledge for logic-based diagnosis, in: Proceedings of the Interna-
tional Conference on Fifth Generation Computing Systems, Tokyo (1988) 1282-1290.
If0] D. Poole, A methodology for using a default and abductive reasoning system, Tech. Report,
Department of Computer Science, University of Waterloo, Waterloo, Ontario (1988).
[11] D. Poole, Normality and faults in logic-based diagnosis, in: Proceedings IJCA1-89, Detroit,
MI (1989).
[12] J.A. Reggia, D.S. Nau, and Y. Wang, A formal model of diagnostic inference I. Problem
formulation and decomposition, Inf. Sci. 37 (1985).
[13] R. Reiter, Circumscription implies predicate completion (sometimes), in: Proceedings AAAI-
82, Pittsburg, PA (1982) 418-420.
[14] R. Reiter, A theory of diagnosis from first principles, Artif. lntell. 32 (1987) 57-95.
[15] P. Struss and O. Dressier, Physical negation-integrating fault models into the general
diagnostic engine, in: Proceedings 1JCAI-89, Detroit, MI (1989) 1318-1323.