Inductive, Abductive and Pragmatic Reasoning

John Bell
Applied Logic Group
Computer Science Department
Queen Mary and Westfield College
University of London
London E1 4NS
jb@dcs.qmw.ac.uk
Abstract induced on the basis of a set of particulars.
Much of our everyday commonsense reasoning
This paper gives a modern version of
is inductive; for example, we reason that the
Pierce’s distinction between induction
sun will rise tomorrow because it has always
and abduction, according to which
done so in the past and there is no reason to
they are both forms of pragmatic (or
think otherwise. Inductive reasoning also plays
context-dependent) reasoning. A de-
a central part in more formal scientific reas-
tailed example of induction and ab-
oning; for example, in much of the reasoning
duction in the formalisation and im-
which is intended to justify particular natural
plementation of predictive and ex-
laws and in the predictions which are based on
planatory commonsense causal reas-
them. A correct inductive argument, according
oning is then given.
to Swinburne, is one in which the “premisses
. . . bear a favourable evidential relation to the
1 Induction, Abduction, and conclusion”, the “premisses make it reasonable
Pragmatic Reasoning for us to accept the conclusion in the sense,
Pierce distinguished between deductive, induct- very roughly, that they report evidence favour-
ive and abductive forms of reasoning in the ing the truth of the conclusion, making it ra-
context of Aritstotle’s syllogistic. Roughly, de- tional to believe the conclusion, making it likely
duction consists in inferring a valid conclusion or adding to the likelihood of its truth”, p. 3.
given the premisses, induction consists in infer- He sharpens this to “an argument . . . [is] a cor-
ring a reasonable (though not valid) conclusion rect inductive argument if its premisses make
given the premisses, and abduction consists in its conclusion more likely to be true that any
inferring an appropriate premise such that the equally detailed rival”, p.4. Swinburne also
conclusion is a valid consequence of it and the notes that ‘the argument will only yield prob-
given premise. This definition has the merit able knowledge if we know nothing else which
of making clear what is given and what is in- affects the probability of the conclusion, apart
ferred in each case. It also draws the distinc- from what is stated by the premisses . . . the ad-
tion between deductive reasoning and the non- dition of new knowledge to the old knowledge
deductive reasoning involved in induction and may mean that I cannot make the same induct-
abduction. However the definition is restricted ive inferences from the old knowledge. Hence
to arguments of syllogistic form. I can only use an inductive argument to give
It is common in Philosophy, to use the term me probable knowledge or reasonable belief, if
‘inductive argument’ more liberally. For ex- the premisses contain all of my relevant know-
ample, Swinburne [15] proposes the following ledge”, pp. 6-7. For futher discussion see [5;
definition: “An inductive argument is an ar- 15].
gument which is not deductively valid but one The term ‘abduction’ will also be interpreted
in which, it is claimed, the premisses ‘make it in a wide sense. Following Harman [9], ab-
reasonable’ for us to accept the conclusion.” duction can be characterised as inference to
[15], p. 2. A special case of this is referred the best explanation. The givens are a set of
to as enumerative induction in philosophy and premisses and a conclusion which is not a de-
rule induction in AI; where a general rule in ductive consequence of the premisses, and the
additional premisses such that the conclusion
is a deductive consequence of the augmented (Identity) Θ |≈ Θ.
premise set. Thus, for consistent Θ Θ |≈ φ if Θ |= φ or
It is possible to use the idea of pragmatic en- there is some consistent extension Θ ∪ Θ′ of
tailment [2; 3] to give a general model-theoretic Θ such that Θ ∪ Θ′ |= φ. What counts as a
account of induction and abduction. Let L reasonable extension can be further constrained
be a language and I be the class of all mod- by imposing additional conditions on the prag-
els of L. A semantic meaning function is a matic function, with corresponding rationality
function [[]] from L to ℘I. Intuitively, for sen- postulates on the inductive consequence rela-
tence φ ∈ L, [[φ]] represents the semantic, or tion. Postulates of this kind are considered by
literal, or context-independent Tmeaning of φ. Flach [7]. In many cases, the postulates of the
Then for each Θ ⊆ L, [[Θ]] = {[[φ]]|φ ∈ Θ}. system P of Kraus, Lehmann and Magidor [12]
The semantic consequence relation induced by together with Consistency Presevation are ap-
the (semantic) meaning function [[]] can then be propriate; as for example is the case for the
defined in the usual way. For any Θ ⊆ L and prediction theories discussed below. In other
φ ∈ L, Θ |= φ (“Θ (semantically) entails φ”) cases, weaker postulates may be appropriate;
iff [[Θ]] ⊆ [[φ]]. So semantic consequence is just for example when the inductive arguments are
the standard Tarkian notion. The analysis is expressed in Default Logic [13]; see [2] for a dis-
then extended to pragmatic reasoning. A prag- cussion. Deciding on an appropriate set of pos-
matic meaning function is a function [] from tulates represents the logical problem of induc-
℘L to ℘I; if φ is a sentence of L then [{φ}] is tion. Picking a particular consequence relation
conveniently abbreviated to [φ]. Intuitively for which satisfies the postulates—picking a par-
each Θ ⊆ L, [Θ] represents the pragmatic, or ticular pragmatic function which satisfies the
intended, or context-dependent meaning of Θ. corresponding conditions—represents the prag-
Then [] and [[]] induce a pragmatic consequence matic problem of induction.
relation: for any Θ ⊆ L and φ ∈ L, Θ |≈ φ (“Θ In the case of abduction, we wish to capture
pragmatically entails φ”) iff [Θ] ⊆ [[φ]]. Thus the idea that the premisses Θ represent a par-
Θ pragmatically entails φ iff φ follows from the tial epistemic context on the basis of which we
intended interpretation of Θ. Assuming that are required to explain φ, and that abduction
L contains the symbols ⊤, ⊥ and ∧, and that is the process of extending Θ in an appropriate
[[⊤]] = I, [[⊥]] = ∅ and [[φ ∧ ψ]] = [[φ]] ∩ [[ψ]], the way in order to do so. Abduction can thus be
properties of pragmatic entailment are follows: thought of as the process demonstrating that Θ
induces φ by finding a particular Θ′ such that
(Normality) Θ |≈ ⊤,
Θ ∪ Θ′ is consistent and Θ ∪ Θ′ |= φ; where,
(Right Weakening)
intuitively, Θ′ represents the explanation (or
If Θ |≈ φ and φ |= ψ then Θ |≈ ψ,
explanations) for φ given Θ. As with induc-
(And)
tion, the postulates of the system P together
If Θ |≈ φ and Θ |≈ ψ then Θ |≈ φ ∧ ψ,
with Consistency Presevation are appropriate
(Consistency Preservation)
for many forms of abduction.
Θ |≈ ⊥ only if Θ |= ⊥.
Thus both induction and abduction can
In the case of induction, we wish to cap- be seen as forms of pragmatic, or context-
ture the idea that the premisses Θ represent dependent, reasoning. What distinguishes
a partial epistemic context and that induction them is what is given and what has to be in-
is the process of extending Θ in an appropri- ferred.
ate way and then reasoning semantically. Thus
Θ induces φ if there is a reasonable expansion
2 Prediction is Induction, but
Θ ∪ Θ′ of Θ which (semantically) entails φ. Explanation is Abduction
Intuitively Θ′ contains premisses to the effect In this section we illustrate the above discussion
that Θ contains all of the relevant information, by outlining the approach to formalising com-
and premisses reflecting the expectations which monsense causal reasoning developed in [6].
arise given Θ. In order to formalise this, we re- The Temporal Calculus, T C, is based on
quire that the pragmatic function is veridical Kleene’s strong three-valued logic [11] and the
(for each Θ ⊆ L, [Θ] ⊆ [[Θ]]) giving the addi- temporal nonmonotonic extension of it in [4].
defined in order to be able to describe com- poral index is added to all atomic formulas of
putations which do not, in principle, termin- the underlying language (to all Kleene atoms);
ate. Instead we adopt an epistemic resource- for example On(B, L)(4) states that block B is
bounded interpretation. We are resource- on location L at time point 4. The earlier-than
bounded epistemically-limited reasoners, who operator, <, allows comparison on time points,
can only know a tiny fragment of what there is and quantification over time points and inter-
to know. However, what matters in our every- vals is introduced.
day commonsense reasoning is that we know The difference in aspect between facts and
those facts which are relevant to us. Con- events is signified by means of the distinguished
sequently given our limitations we concentrate predicate Occurs. Thus a sentence of the form
our efforts on attempting to establish the truth Occurs(e)(t) states that a token of a primitive
or falsity of those facts that we consider relevant event of type e occurs at time t. Note that
and ignore (treat as undefined) the truth values more than one (distinct) event (token) may oc-
of the rest. However, given our limitiations, we cur simultaneously. Compound events can then
do not always succeed in our attempts. On this be introduced by definition.
interpretation then a sentence φ can be true (es- In order to gain comparable expressivity with
tablished by the reasoner(s) as being true), false the Situation Calculus, Kleene atoms and their
(established by the reasoner(s) as being false) or negations, Kleene literals, are encoded as terms.
undefined (ignored by the reasoner(s) as irrel- For simplicity, autonomous naming is used, as
evant, or unestablishable within the reasoner(s) there is no danger of confusion. In addition
resource limitations). the distinguished predicates True, Qual, and
The syntax of Kleene’s logic is identical Aff are introduced. The truth predicate, True,
with that of classical Predicate Calculus. The functions similarly to the Holds relation of the
truth conditions of the logical constants yield a Situation Calculus. Thus a sentence of the form
Boolean truth value wherever possible. Thus, True(λ)(t) states that the Kleene literal (named
for example, a sentence of the form φ∨ψ is true by) λ is true at time t. As might be expected
if at least one of the disjuncts is true (even if of a truth predicate, we have True(λ)(t) ≡ λ(t)
the other is undefined). for any Kleene literal λ. A sentence of the form
As it stands, Kleene’s logic is appropriate for Qual(e)(t) states that the event e is qualified
representing monotonic reasoning with partial at time t; that is, that there is reason to doubt
information. In order to be able to represent that it will succeed, given that it occurs at t. A
nonmonotonic reasoning, an undefined operator sentence of the form Aff (λ)(t) states that the
is added. Informally ?φ states that the truth Kleene literal λ is affected at t; that is, that
value of φ is undefined; that is, that neither the there is reason to doubt that its truth value
truth nor the falsity of φ is established. Then persists beyond t.
weak affirmation, weak negation, and defined Prediction Theories (PTs) are expressed in
operators are introduced: T C. PTs may contain boundary conditions, pre-
diction rules, domain rules and default rules.
def
◦φ = ?φ ∨ φ “φ is not false” The boundary conditions are unconditional
•φ
def
= ?φ ∨ ¬φ “φ is not true” sentences which state facts and the occurrence
def
of events.
!φ = φ ∨ ¬φ “φ is not undefined” Prediction rules are temporally forwards-
directed conditionals (conditionals in which, in
Meta-level distinctions such as “φ is not false” any instance, the latest time point referred to
and “φ is not true”, and consequently non- in the antecedent temporally preceeds that re-
monotonic inference rules, can be expressed in ferred to in the consequent) whose antecedents
the object language; see, for example, the rules are typically hypothetical. These are used to
(1) and (2) given below. define change. For example:
In keeping with the resource-bounded inter-
pretation, a new conditional operator is intro- ∀e, t(Pre(e)(t) ∧ Occurs(e)(t) ∧ •Qual(e)(t) →
duced. A sentence of the form φ → ψ should
Post(e)(t + 1)) (1)
be thought of as expressing a constraint which
must be met if φ is true, but which can other- ∀λ, t(True(λ)(t) ∧ •Aff (λ)(t) →
wise be ignored. T rue(λ)(t + 1)) (2)
states that if event e occurs at t and the ne- literals, and at least one more Qual literal
cessary preconditions of e are true at t and e is established in M ′ at t, or
is not qualified at t, then infer that the post- • M and M ′ agree on all ordinary, Occurs,
conditions of e are true at t + 1; that is, that and Qual literals, and at least one more
e succeeds at t. Rule (2) is a general rule of Aff literal is established in M ′ at t.
inertia, which states that if the Kleene literal
(named by the term) λ is true at t and (the Definition 2.1 A model M is a prioritised
Kleene literal named by) λ is not affected at t chronologically least defined model (p.c.l.d.
then infer that (the Kleene literal named by) model) of φ if M |= φ and there is no other
λ is true at t + 1. Given that, for any Kleene model M ′ such that M ′ |= φ and M ′ ≺pcld M .
literal λ, True(λ)(t) is equivalent to λ(t), (2) Likewise, M is a p.c.l.d. model of a set of sen-
is equivalent to the following higher-order sen- tences Θ if M |= Θ and there is no other model
tence: M ′ such that M ′ |= Θ and M ′ ≺pcld M .
Definition 2.2 A PT Θ predicts a sentence φ,
∀λ, t(λ(t) ∧ •Aff (λ)(t) → λ(t + 1)) (3) written Θ |≈pcld φ, if φ is true in all p.c.l.d.
In practice it is convenient to use (3) as an models of Θ.
abbreviation for (2). In order to illustrate inductive and abductive
Domain rules are categorical temporally un- reasoning in this framework we discuss predic-
directed (bi-)conditionals which are used for a tion and explanation and give a simple example
variety of purposes, including the definition of of each. In the sequel we adopt the conven-
domain constraints and ramifications. tion that free variables are implicitly univer-
Finally, default rules are temporally undirec- sally quantified.
ted conditionals with hypothetical antecedents The frame (or inertia, or persistence) prob-
which are used to do default reasoning at a time lem is the problem of efficiently representing
point(or within an interval. The intention is to what does not change as result of an event. In
allow a limited amount of default reasoning, not other words, we want to represent the common-
to incorporate Default Logic; it is thus assumed sense inference that facts persist, other things
that these rules can be stratified. being equal; where other things are equal if no
The pragmatics of PTs are given by priorit- event occurs which affects those facts. In order
ised chronological minimisation. This captures to do so, we use a general inertia rule which is
three necessary properties of predictive causal analogous to those used in Situation Calculus-
reasoning: it is rule-governed (albeit by defeas- based approaches. For any Kleene literal λ:
ible rules), it is temporally forwards-directed,
and potential change takes precedence over in- λ(t) ∧ •Aff (λ)(t) → λ(t + 1) (4)
ertia; that is, if an action can succeed, it does
In simple cases it is sufficient to use the inertia
succeed. Prioritised chronological minimisation
rule to conclude “by default” that facts persist.
is realised by exploiting the partiality of the
Particular applications of the rule are implicitly
Kleene basis and chronologically minimising in-
blocked by “back propagation”.
formation (what is defined/established). In or-
der to give preference to potential change, qual- Example 2.3 The Yale Shooting Problem [8].
ifications (reasons why an event may not occur At time 1 a gun is loaded and pointed at Fred.
successfully) are minimised in preference to af- Nothing relevant happens at time 2. At time 3
fectations (reasons why a fact should not per- the gun is fired. Is Fred dead as a result? In-
sist) at each time point. Informally, two models tuitively we think so because we infer that the
M and M ′ are said to be comparable if they gun remains loaded until it is fired. The prob-
differ at most on the interpretation of ordin- lem is to represent this commonsense inference
ary literals or on the interpretation of Occurs, formally.
Qual or Aff literals. Then M ≺pcld M ′ (M is Let Θ be the PT {(4), (5), (6)}, where:
preferrred to M ′ ) if M and M ′ are comparable
and there is a time point t such that M and M ′ Alive(1) ∧ Loaded(1) ∧
agree up to t, and: Occurs(Shoot)(3) (5)
• at least one more ordinary or Occurs literal Loaded(t) ∧ Occurs(Shoot)(t) →
is established in M ′ at t, or ¬Alive(t + 1) (6)
Alive(3) ∧ ¬Alive(4). Let M be a p.c.l.d. time 1. The other, less likely but still plausible,
model of Θ. Then M |= ?Aff (L)(1) and is that the gun was loaded at time 1 and fired
M |= ?Aff (A)(1), so by (4), M |= Loaded(2) at time 2.
and M |= Alive(2). Similarly M |= Loaded(3) Let Θ be the PT {(4), (7), (8), (9), (10)};
and M |= Alive(3). Rule (6) now gives M |= where:
¬Alive(4). Rule (6) also gives M |= Aff (A)(3)
thus blocking the application of rule (4) at time Alive(1) ∧ Occurs(Shoot)(1) ∧
3 in which λ is A. 2 ¬Alive(3) (7)
As suggested by the example, the (finite re- 1 ≤ t ≤ 2 → !Aff (Alive)(t) (8)
strictions of) the intended model(s) of appro- Occurs(Load)(t) → Loaded(t + 1) (9)
priate PTs can be generated by starting with Loaded(t) ∧ Occurs(Shoot)(t) →
an initial model, M/0, and then using the rules
¬Alive(t + 1) (10)
as detailed instructions for extending each M/t
to M/t + 1. Initial work on this model-building The abductive task is to find the minimal Θi
implementation is discussed in [3; 4]. such that Θ ∪ Θi |≈pcld ¬Alive(3). As sugges-
PTs can also be used as a basis for the rep- ted, this can be done by reasoning forwards in
resentation of explanatory commonsense causal time from the initial time point, and assuming
reasoning. Standardly [10; 14], given a theory only what is necessary to establish the conclu-
Θ such that Θ 6|= φ, abduction is presented sion at each time point. Thus at time 1 we
as the process of finding appropriate minimal have Θ1 = {Loaded(1)}, and at time 2 we have
sets Θi such that each Θ ∪ Θi is consistent and Θ2 = {Occurs(Load)(1), Occurs(Shoot)(2)} as
Θ∪Θi |= φ. The various Θi are then the explan- possible explanations. The gun was loaded at
ations for φ given Θ. We change this definition time 1 and so, as a result of the shoot event,
slightly and seek appropriate minimal Θi such Fred was dead at time 2. Alternatively, Fred
that Θ ∪ Θi |≈pcld φ. Using |≈pcld rather than was alive at time 2 and died as a result of a
|= as the consequence relation should result in load event at time 1 and a shoot event at time
fewer, more relevant, explanations (as the ob- 2.
vious nondeductive conclusions are already in- In Baker’s original formulation of the prob-
cluded in the |≈pcld relation). Furthermore, the lem, a wait event occurs at time 2. If we add a
present approach suggests that each Θi can be rule to this effect:
generated by starting at the initial time point
and then reasoning forwards in time assuming ¬∃eOccurs(e)(2), (11)
only what is necessary to predict the conclu- then we are left with Θ1 as the only explana-
sion at each interim time point (and then check- tion. 2
ing that the conclusion is true at the final time
point). In order to count as reasonable explan- Acknowledgements
ations, the assumptions, or abducibles, should
I would like to thank the participants of the
be relevant and non-trivial. For present pur-
ECAI-96 Workshop on Abductive and Induct-
poses, relevant assumptions can be taken to be
ive Reasoning for helpful discussions and com-
instantiations of literals occurring in the ante-
ments.
cedents of rules of the theory Θ. While non-
triviality can be taken to be the requirement
References
that Θi 6|= φ. Explanation is thus viewed as
augmented prediction. [1] A. Baker (1991) Nonmonotonic Reason-
ing in the Framework of the Situation
Example 2.4 The Stanford Murder Mystery
Calculus. Artificial Intelligence 49 pp. 5-
[1]. In this variant of the Yale Shooting Prob-
23.
lem, Fred is alive at time 1. As before, if the
gun is loaded when it is fired, then Fred is dead [2] J. Bell. Pragmatic Logics. Proceedings of
as a result. We are also told that a shoot event KR ’91, Morgan Kaufmann, 1991, pp. 50-
occurs at time 1 and that Fred is dead at time 60. Revised and extended in 1992.
3. When and why did Fred die? It seems that [3] J. Bell. Pragmatic Reasoning; a Model-
there are two possible explanations. One is that Based Theory. In M. Masuch and L. Po-
the gun was already loaded at time 1 and Fred los (eds.), Applied Logic, How, What and
 Applied Logic Conference, Amsterdam,
December 1992. Kluwer Academic Pub-
lishers, Amsterdam, 1995, pp. 1-28.
[4] J. Bell. A Model-Based Approach to Pre-
dictive Causal Reasoning. In: P. Doherty
(ed.), Partiality, Modality, and Non-
Monotonicity. FoLLI Studies in Logic
Language and Information series, CSLI
Publications, Stanford, 1996, pp. 169-
195.
[5] J. Bell. Induction in Philosophy and AI.
Working Notes of the ECAI-96 Workshop
on Abductive and Inductive Reasoning,
M. Denecker, L. De Raedt, P. Flach and
T. Kakas (eds.), pp. 7-9.
[6] J. Bell. Prediction Theories. Manuscript,
1997.
[7] P. Flach. Rationality Postulates for In-
duction. Proceedings of the Sixth Con-
ference on Rationality and Knowledge,
TARK’96, Y. Shoham (ed.), pp. 267-281.
[8] S. Hanks and D. McDermott (1986) De-
fault Reasoning, Nonmonotonic Logics,
and the Frame Problem. Proceedings of
AAAI’86, pp. 328-333.
[9] G. Harman. Inference to the best explan-
ation. Philosophical Review 74, pp. 88-95.
[10] A.C. Kakas, R.C. Kowalski and F. Toni.
Abductive Logic Programming. Journal
of Computational Logic 2(6), 1992, pp.
719-770.
[11] S.C. Kleene. Introduction to
Metamathematics. North-Holland, Ams-
terdam, 1952.
[12] S. Kraus, D. Lehmann and M. Ma-
gidor. Nonmonotonic Reasoning, prefer-
ential models and cumulative logics. Ar-
tificial Intelligence 44 1990, pp. 167-207.
[13] R. Reiter. A logic for default reasoning.
Artificial Intelligence 13 1980, pp. 81-132.
[14] M. Shanahan. Explanation in the Situ-
ation Calculus. Proceedings of IJCAI’93,
pp. 160-165.
[15] R. Swinburne (ed.) The Justification of
Induction. Oxford Readings in Philo-
sophy series, Oxford University Press,
Oxford, 1974.