ATOCHA ALISEDA
LOGICS IN SCIENTIFIC DISCOVERY
ABSTRACT. In this paper I argue for a place for logic in scientific methodology,
at the same level as that of computational and historical approaches. While it is
well known that a a whole generation of philosophers dismissed Logical
Positivism (not just for the logic though), there are at least two reasons to reconsider logical approaches in the philosophy of science. On the one hand, the present
situation in logical research has gone far beyond the formal developments that
deductive logic reached last century, and new research includes the formalization of several other types of reasoning, like induction and abduction. On the
other hand, we call for a balanced Philosophy of Science, one in which both
methods, the formal and the historical may be complementary, together providing
a pluralistic view of science, in which no method is the predominant one.
KEY WORDS: abduction, heuristics, logic, scientific discovery
1. INTRODUCTION
A central theme in the study of human reasoning is the construction
of explanations that give us an understanding of the world we live
in. Broadly speaking, abduction is a reasoning process invoked to
explain a puzzling observation. If we wake up, and the lawn is wet,
we might explain this observation by assuming that it must have
rained, or by assuming that the sprinklers have been on. This is a
practical setting found in our day-to-day common sense reasoning.
Abduction also occurs in more theoretical scientific contexts. For
instance, it has been claimed that Johannes Kepler’s great discovery
that the orbit of the planets is elliptical rather than circular was a
prime piece of abductive reasoning (Hanson, 1961; Peirce 2.623,
1958). What initially led to this discovery was his observation that
the longitudes of Mars did not fit circular orbits. However, before
even dreaming that the best explanation involved ellipses instead
of circles, he tried several other forms. Moreover, Kepler had to
make several other assumptions about the planetary system, without
which his discovery does not work. His heliocentric view allowed
Foundation of Science 9: 339–363, 2004.
© 2004 Kluwer Academic Publishers. Printed in the Netherlands.
340
ATOCHA ALISEDA
him to think that the sun, so near to the center of the planetary
system, and so large, must somehow cause the planets to move as
they do. In addition to this strong conjecture, he also had to generalize his findings for Mars to all the other planets, by assuming that
the same physical conditions obtained throughout the solar system.
This whole process of explanation took many years.
Abduction is thinking from evidence to explanation, a type of
reasoning characteristic of many different situations with incomplete information. Note that the word explanation – which we treat
as largely synonymous with abduction – is a noun which denotes
either an activity, indicated by its corresponding verb, or the result
of that activity. That is, it may be used both to refer to a finished
product, the explanation of a phenomenon, or to an activity, the
process that led to that explanation. These two uses are closely
related. The process of explanation produces explanations as its
product, but the two are not the same.
One can relate this distinction to more traditional ones. An
example is Reichenbach’s (1938) well-known opposition of ‘context
of discovery’ versus ‘context of justification’, which served as
a basis for logical positivism. In this tradition, a fundamental
assumption is that the methodology of science has logic as its
tool and that it deals only with aspects of justification in scientific
research. Kepler’s explanation-product “the orbits of the planets are
elliptical”, which justifies the observed facts, does not include the
explanation-process of how he came to make this discovery. The
context of discovery is taken to be purely psychological. This situation is not particular to the philosophy of science. The high level
of formal rigour that logical research reached in the 20th century,
is due on the one hand to its divorce from Psychologism and on
the other hand in compliance with the deductive method inherited
from Euclides, in which all (geometrical) truths are derived from
a few basic axioms and rules of inference. Studies of the ways in
which proofs are generated or new axioms invented, has been totally
outside formal analysis. If something is ever mentioned, it is only
part of the history of mathematics.
Our claim, however, is that the study of the context of discovery
allows, to some extent, for a precise formal treatment. Cognitive
psychologists study mental patterns of discovery, learning theorists
LOGICS IN SCIENTIFIC DISCOVERY
341
in artificial intelligence study formal hypothesis formation, and one
can even work with concrete computational algorithms that produce
explanations.
Reinchenbach’s distinction between the contexts of justification
and of discovery has left out of its analysis – especially from a
formal point of view – a very important part of scientific practice,
that which includes issues related to the generation of new theories
and scientific explanations, concept formation as well as aspects of
progress and discovery in science.
Section two deals with the contexts of research in the methodology of science in philosophy as well as in artificial intelligence
and cognitive psychology. Section three argues for the role of logics
in scientific methodology, and section four gives a general account
of the logic of abduction. Finally, section five presents further
challenges and suggests directions for future research.
2. CONTEXTS OF RESEARCH IN THE METHODOLOGY OF
SCIENCE
2.1. Context of Justification
The dominant trend in philosophy has focused on abduction as
product rather than as process, just as it has done for other epistemic
notions. Aristotle, Mill, and in the XXth century, the influential
philosopher of science Carl Hempel, all based their accounts of
explanation on proposing criteria to characterize its products. These
accounts generally may be classified as argumentative and nonargumentative types of explanation (Ruben, 1990; Salmon, 1990).
Of particular importance is the ‘argumentative’ Hempelian tradition.
Its followers aim to model empirical why-questions, whose answers
are scientific explanations in the form of arguments. In these arguments, the ‘explanandum’ (the fact to be explained) is derived
(deductively or inductively) from the ‘explanans’ (that which does
the explaining) supplemented with relevant ‘laws’ (general or statistical) and ‘initial conditions’. For instance, the fact that an explosion
occurred may be explained by my lighting the match, given the
laws of physics, and initial conditions to the effect that oxygen was
present, the match was not wet, etcetera. Therefore, explanations
are public objects of ‘justification’, that can be checked and tested
by independent logical criteria.
342
ATOCHA ALISEDA
In its deductive version, the Hempelian account, found in many
standard texts on the philosophy of science (e.g. Salmon, 1990) is
called deductive-nomological, for obvious reasons. But its engine is
not just standard deduction. Additional restrictions must be imposed
on the relation between explanans and explanandum, as neither
deduction nor induction is a sufficient condition for genuine explanation. To mention a simple example, every formula is derivable
from itself (E ⊢ E), but it seems counterintuitive, or at least very
uninformative, to explain anything by itself.
Other, non-deductive approaches to explanation exist in the literature. For instance, Ruben (1990) points out these two: Salmon
(1977, p. 159) takes them to be: “an assemblage of factors that
are statistically relevant . . .”, while van Frassen (1980, p. 134)
makes them simply: “an answer”. For Salmon, the question is not
how probable the explanans renders the explanandum, but rather
whether the facts adduced make a difference to the probability of the
explanandum. Moreover, this relationship need not be in the form
of an argument. For van Fraassen, a representative of pragmatic
approaches to explanation, the explanandum is a contrastive whyquestion. Thus, rather than asking “why E?”, one asks “why E
rather than G?”. The pragmatic view seems closer to abduction
as a process, and indeed, the focus on questions introduces some
dynamics of explaining. Still, it does not tell us how to produce
explanations.
There are also alternative deductive approaches. Another, and
rather famous deductivist tradition is Popper’s logic of scientific
discovery (1958). Its method of conjectures and refutations proposes
the testing of hypotheses, by attempting to refute them:
The actual procedure of science is to operate with conjectures: to jump to
conclusions – often after one single observation. (Popper, 1963, p. 53)
Thus science starts from problems, and not from observations; though observations may give rise to a problem, specially if they are unexpected; that is to say, if
they clash with our expectations or theories. (Popper, 1963, p. 222)
Popper’s deductive focus is on refutation of falsehoods, rather than
explanation of truths. One might speculate about a similar ‘negative’
counterpart to abduction. Although Popper’s method claims to be
a logic of scientific discovery, he views the actual construction of
explanations as an exclusive matter for psychology – and hence his
LOGICS IN SCIENTIFIC DISCOVERY
343
‘trial and error’ philosophy offers no further clues for abduction
as a logic of discovery. What is common to all these approaches
in the philosophy of science is the conception of scientific practice as a finished product and not as a process in itself. They
characterize notions like explanation through logical, statistical or
pragmatic criteria, but do not describe how is it that explanations are
constructed. However, they do highlight the importance of a hidden
parameter in abduction. Whether with Hempel, Salmon, or Popper,
scientific explanation never takes place in isolation, but always in
the context of some background theory.
2.2. Context of Discovery
It is well-known that great philosophers and mathematicians have
been brilliant exceptions in the study of discovery in science, and
that their non conventional contributions to this field, although great
inspirations, have not set new paradigms in the methodology of
science.
An example is the work of Rescher (1978), which introduces
a direction of thought. Interestingly, this establishes a temporal
distinction between ‘prediction’ and ‘retroduction’ (another term
for abduction), by marking the precedence of the explanandum over
the hypothesis in the latter case. Another author emphasizing explanation as a process of discovery is Hanson (1961), who gave an
account of patterns of discovery, recognizing a central role for retroduction. Also relevant here is the work by Lakatos (1976), a critical
response to Popper’s logic of scientific discovery:
There is no infallibilist logic of scientific discovery leading infallibly to results,
but there is a fallibilistic logic of discovery which is the logic of scientific
progress. But Popper, who has laid the basis for this logic of discovery was not
interested in the meta-question of what is the nature of this investigation, so he
did not realize that it is neither psychology nor logic, but an independent field, the
logic of discovery, heuristics. (Lakatos, 1976, p. 167. My emphasis.)
It is interesting to note that Lakatos was greatly inspired by the
history of mathematics, paying particular attention to processes that
created new concepts – often referring to G. Polya as the founding
father of heuristics in mathematical discovery.
What all these examples reveal is that in science, explanation
involves the invention of new concepts, just as much as the positing
344
ATOCHA ALISEDA
of new statements (in some fixed conceptual framework). So far,
this has not led to extensive formal studies of concept formation,
similar to what is known about deductive logic. (Exceptions that
prove the rule are occasional uses of Beth’s Definability Theorem
in the philosophical literature. A similar lacuna vis-à-vis concept
revision exists in the current theory of belief revision in Artificial
Intelligence (Thagard, 1992; Kuipers, 1999; Aliseda, 2000).)
Several other exceptions in the study of the context of discovery
worthy of mention include the pioneering work of Herbert Simon
and his team (Simon et al., 1981; Langley et al., 1987) who
conceives scientific reasoning as ‘problem solving’ and proposes
to use the machinery of his programs based on heuristic search
(e.g. General Problem Solver) to develop programs that simulate the discovery of quantitative laws (the famous BACON) in
Physics and qualitative laws (GLAUBER) in chemistry. In the same
spirit, Paul Thagard proposes a new field of research, “Computational Philosophy of Science” (Thagard, 1988) and puts forward the
computational program PI (Processes of Induction) to model some
aspects of scientific practice, such as concept formation and theory
building.
Even though these approaches come from apparently distant
disciplines to philosophy of science, namely cognitive psychology
and artificial intelligence, they are proposals which suggest the
inclusion of computational tools in the philosophy of science
research methodology and by so doing claim to reincorporate
aspects from the context of discovery within its agenda. The key
concept in all this is that of heuristics, the guide in scientific
discovery which is neither totally rational nor absolutely blind.
However, despite the fact that computational approaches propose
a dynamic view, in which notions like explanation and theory
building are not regarded as finished products but rather as processes
in themselves, the role of logic seems to be foreign to their
methodology. Their systems are based on concepts like heuristic
search in artificial intelligence, but the proposed computational
systems lack logical foundation (we will argue for a logic with
heuristics in the last section). Researchers in this field resemble
historians rather than (post)positivists with computers. The fact that
they use computational methods seems to place them far from the
LOGICS IN SCIENTIFIC DISCOVERY
345
historical approach and close to the formal one, but a quick view to
their programs shows that their program design is inspired by historical reconstructions of the scientific discoveries they simulate. There
are fundamental differences in the assumptions they take for their
implementations, showing in many cases that their discrepancies are
of historical and not of computational nature. For instance, there
is no agreement in the relevant assumptions that Kepler uses for
his discovery that the orbits of the planets are elliptical rather than
circular. While for Thagard (1992) the main assumption is that the
earth is stationary and the sun moves around it, Simon et al. (1987)
and his team together with Hanson (1961), think that Kepler’s
discovery would have been impossible without an heliocentric view
of the planetary system.
3. LOGIC(S) FOR SCIENTIFIC METHODOLOGY
Logic, classical or otherwise, in philosophy of science is nowadays,
to put it simply, out of fashion. In fact, although classical logic is still
part of the curricula in philosophy of science graduate programs,
students soon learn that a whole generation of philosophers regarded
logical positivism as a failed attempt (though not just for the logic),
claiming that scientific practice does not follow logical patterns of
reasoning, many of which favoured studies of science based on
historical cases. So, why bother about the place of logic in scientific
research?
On the one hand, the present situation in logical research has
gone far beyond the formal developments that deductive logic
reached last century, and new research includes the formalization
of several other types or reasoning, like induction and abduction.
On the other hand, we claim for a balanced philosophy of science,
one in which both methods, the formal and the historical may be
complementary, together providing a pluralistic view of science, in
which no method is the predominant one.
3.1. Logical Research in the XIX Century
As for the first claim, one can place logical research in this century
in a broader setting for general logic (van Benthem, 1996), looking
back at the original program of Bernard Bolzano (1781–1848), in his
346
ATOCHA ALISEDA
“Wissenschaftslehre” (Bolzano, 1973), in which he engaged (among
other things) in the study of different varieties of inference.
Bolzano’s notion of deducibility (Ableitbarkeit) has long
been recognized as a predecessor of Tarski’s notion of logical
consequence (Corcoran, 1975). However, the two differ in several
respects and in our broader view of logic they even appear radically
different. These differences have been studied both from a philosophical (Thompson, 1981) and from a logical point of view (van
Benthem, 1984).
One of Bolzano’s goals was to show why the claims of science
form a theory as opposed to an arbitrary set of propositions. For
this purpose, he defines his notion of deducibility as a logical
relationship extracting conclusions from premises forming compatible propositions, those for which some set of ideas make all
propositions true when uniformly substituted throughout. In addition, compatible propositions must share common ideas. Bolzano’s
use of ‘substitutions’ is of interest by itself, but for our purposes
here, we will identify these (somewhat roughly) with the logicomathematical use of ‘models’. Thompson attributes the difference
between Bolzano’s consequence and Tarski’s to the fact that the
former notion is epistemic while the latter is ontological. These
differences have pronounced technical effects. With Bolzano, the
premises must be consistent (sharing at least one model), with
Tarski, they need not. Therefore, from a contradiction, everything
follows for Tarski, and nothing for Bolzano.
Restated in model-theoretic terms, Bolzano’s notion of deducibility reads as follows (cf. van Benthem, 1984):
T, C ⇒ E if
(1) The conjunction of T and C is consistent.
(2) Every model for T plus C verifies E.
Therefore, Bolzano’s notion may be seen (anachronistically) as
Tarski’s consequence plus the additional condition of consistency.
Bolzano does not stop here. A finer grain to deducibility occurs in
his notion of exact deducibility which imposes greater requirements
of ‘relevance’. A modern version, involving inclusion-minimality
for sets of abducibles, may be transcribed (again, with some historical injustice) as:
T, C ⇒+ E if
LOGICS IN SCIENTIFIC DISCOVERY
347
(1) T, C ⇒ E
(2) There is no proper subset of C, C′ , such that T, C′ ⇒ E.
That is, in addition to consistency with the background theory,
the premise set C must be ‘fully explanatory’ in that no subpart
of it would suffice for the derivation. Notice that this leads to
non-monotonicity.1 Here is an example:
T, a → b, a ⇒+ b
T, a → b, a, b → c + b
Bolzano’s agenda for logic is relevant to the study of abductive
reasoning (and the study of general non-monotonic consequence
relations) for several reasons. It suggests the methodological point
that what we need is not so much proliferation of different logics
as a better grasp of different styles of consequence. Moreover, his
work supports the claim we present in what follows, that truth is not
all there is to understanding explanatory reasoning.
3.2. Logical Research in the XX Century and Beyond
In the first half of the XX century, Tarski’s notion of truth gave
meaning to the notion of logical consequence. A conclusion follows
if it is true in all models where the premises are true. But the contemporary philosophical and computational traditions have shown that
natural notions of inference may need more than truth in the above
sense, or may even hinge on different properties altogether. For
example, among the candidates which revolve around truth, statistical inference requires not total inclusion of premise models in
conclusion models, but only a significant overlap, resulting in a high
degree of certainty. Other approaches introduce new semantic primitives. Notably, Shoham’s notion of causal and default reasoning
(Shoham, 1988) introduces a preference order on models, requiring
only that the most preferred models of the premises be included in
the models of the conclusion.
More radically, dynamic semantics replaces the notion of truth
by that of information change, aiming to model the flow of information. This move leads to a redesign for Tarski semantics, with
e.g. quantifiers becoming actions on assignments (van Benthem and
Cepparello, 1994). This logical paradigm has room for many different inferential notions (Groeneveld, 1995). An example is update-
348
ATOCHA ALISEDA
to-test-consequence: “Process the successive premises, thereby
absorbing their informational content into the initial information
state. At the end, check if the resulting state is rich enough to satisfy
a conclusion”. Informational content rather than truth is also the key
semantic property in situation theory (Perry and Barwise, 1983).
In addition to truth-based and information-based approaches,
there are, of course, also various proof-theoretic variations on
standard consequence. Examples are deafult reasoning: “E is
provable unless and until E is disproved” (Reiter, 1980), and indeed
Hempel’s hypothetico-deductive model of scientific inference itself.
There are new notions of derivability as, for example, the one that
dictates that a conclusion is followed only with a high degree of
certainty (e.g. statistical inference). There are also more dynamic
versions in which the concern is to incorporate a new piece of
information into a database, a scientific theory, or a set of common
sense beliefs (Gärdenfors, 1988). In this approach (a work whose
roots lie in the philosophy of science), three main types of belief
change are operations of ‘expansion’, ‘contraction’, and ‘revision’.
A theory may be expanded by adding new formulas, contracted
by deleting existing formulas, or revised by first being contracted
and then expanded. These operations are defined in such a way as
to ensure that the theory or belief system remains consistent and
suitably ‘closed’ when incorporating new information.
In connection to philosophy of science, there are some proposals
that lay bridges between non-monotonic logics and Hempel’s
models for explanation. In Tan (1992) the inductive statistical model
is constructed based on Reiter’s default logic and in (Aliseda, 1997)
two models of scientific explanation (deductive and statistical) are
presented as cases of abductive logic, thus claiming that these
models no dot follow the canons of classical logic.
All these logics evolved within the last two decades of the
XX century, seeking to model common sense as well as scientific
reasoning. Therefore they need to allow for a richer language
and logical formats than the one developed in classical logic for
mathematical reasoning.
All above alternatives agree with our view of abduction. On
our view, abduction is not a new notion of inference. It is rather
a topic-dependent practice of explanatory reasoning, which can
LOGICS IN SCIENTIFIC DISCOVERY
349
be supported by various members of the above family. In fact, it
is appealing to think of abductive inference in several respects,
as inference involving preservation of both truth and explanatory
power. In fact, appropriately defined, both might turn out to be
equivalent. It has also been argued that since abduction is a form
of reversed deduction, just as deduction is truth-preserving, abduction must be falsity-preserving (Michalski, 1994). However, Flach
(1995) gives convincing arguments against this particular move.
Moreover, as we have already discussed intuitively, abduction is not
just deduction in reverse.
As stated earlier, we argue for a balanced philosophy of science,
one in which we take advantage of a variety of methodologies, such
as logical, computational and even historical, all together giving a
broad view of science. I think present research has finally overcome
the view that a single treatment in scientific methodology is enough
to give us an understanding of scientific practice, a view that was
already fostered by Pat Suppes (1979) in the late sixties.
We know at present that logical models (classical or otherwise)
are insufficient to completely characterize notions like explanation,
confirmation or falsification in philosophy of science, but this fact
does not exclude that some problems in the history of science may
be tackled from a formal point of view. For instance, “some claims
about scientific revolutions, seem to require statistical and quantitative data analysis, if there is some serious pretension to regard
them with the same status as other claims about social or natural
phenomena” (Suppes, 1979, p. 97).
In fact, “Computational Philosophy of Science” may be regarded
as a successful marriage between historical and formal approaches.
It is argued that although several heuristic rules have been derived
from historical reconstructions in Science, they are proposed to be
used for future research (Meheus and Nickles, 1999).
This is not to say however, that historical analysis of scientific
practice could be done in a formal fashion, or that logical treatment should care for some kind of “historic parameter” in its
methodology, but we claim instead that these two views should share
their insights and findings in order to complement each other. After
all, recall that part of Hempel’s inspiration to device a “logic of
350
ATOCHA ALISEDA
explanation” were his studies on the logic for ‘drawing documents
from history’.
4. LOGIC(S) FOR ABDUCTION
So far I have argued for a place for logic in scientific methodology, at
the same level of computational and historical approaches. However,
I have not proposed a particular logical enterprise with a specific
application to philosophy of science. Instead of giving the precise
details of any single existing proposal, I will sketch the general
problems and solutions toward a logic for abductive explanations.
In the next section I briefly summarize several other proposals.
To start with, the construction of an explanation depends on
what we take to be the scientific background theory. But even this
is not the only parameter that plays a role, there are still further
complicating factors. Sometimes, no single obvious explanation is
available, but rather several competing ones – and we have to select.
Sometimes, the explanation involves not just advancing facts or
rules in our current conceptual frame, but rather the creation of new
concepts, that allow for new description of the relevant phenomena.
Evidently, we must draw a line somewhere for a formal study of
abductive reasoning.
Abductive reasoning is not a case of ordinary deductive
reasoning, and this for a number of reasons. In particular, the
explanations produced might be defeated. Maybe the lawn is wet
(cf. Introduction) because children have been playing with water.
Co-occurrence of clouds and the lawn being wet does not necessarily link them in a causal way. What we learn subsequently can
invalidate an earlier abductive conclusion. Moreover, the reasoning
involved in this type of thinking seems to go in the reverse order,
as compared with deduction, as all these cases run from evidence
to hypothesis, and not from premises to conclusion, as is usual
in deductive patterns. Finally, describing the way in which an
explanation is found, does not seem to follow specific rules. Indeed,
the precise nature of Kepler’s ‘discovery’ remains under intensive
debate.
In what follows, we shall describe abduction in artificial intelligence. Research in this field dates back to the seventies (Pople,
LOGICS IN SCIENTIFIC DISCOVERY
351
1973), but it is only fairly recently that it has attracted great interest,
in areas like logic programming, knowledge assimilation (Kakas et
al., 1993), and diagnosis (Poole, 1987), to name just a few. Abduction is also coming up in the context of data bases and knowledge
bases, that is, in mainstream computer science.
In this setting, the product-process distinction has a natural
counterpart, namely, in logic-based vs computational-based
approaches to abduction. While the former focuses on giving a
semantics to the logic of abduction, usually defined as ‘backwards
deduction plus additional conditions’, the latter is concerned with
providing algorithms to produce abductions.
It is impossible to give an overview here of this exploding field.
Therefore, we limit ourselves to a brief description of abduction as
logical inference. There is much more in this field of potential philosophical interest, however. For abduction in Bayesian networks,
connectionism, and many other angles, the reader is advised to
consult Josephson (1994), Konolige (1996) and Paul (1993).
4.1. Abduction as Logical Inference
The general trend in logic based approaches to abduction in AI interprets abduction as backwards deduction plus additional conditions.
This brings it very close to deductive-nomological explanation in
the Hempel style, witness the following format. What follows is the
standard version of abduction as deduction via some consistent additional assumption, satisfying certain extra conditions. It combines
some common requirements from the literature (cf. Konolige, 1990;
Kakaes et al., 1993; Mayer and Pirri, 1993; and Aliseda, 1997 for
further motivation):
Given a theory T (a set of formulae) and a formula E, C
is an explanation if
(1) T ∪ C |= E
(2) C is consistent with T
(3) C is ‘minimal’ (there are several ways to characterize
minimality, cf. Aliseda, 1997)
(4) C has some restricted syntactical form (usually an atomic
formula or a conjunction of them).
An additional condition not always made explicit is that T | = E. This
says that the fact to be explained should not already follow from the
352
ATOCHA ALISEDA
background theory alone. Sometimes, the latter condition figures as
a precondition for an abductive problem.
What can one say in general about the properties of such an
‘enriched’ notion of consequence? A logical trend in AI studies
variations of classical consequence via their ‘structural rules’, which
govern the combination of basic inferences, without referring to
any special logical connectives. (Cf. the analysis of non-monotonic
consequence relations in AI of Gabbay (1994), Kraus et al. (1990),
and the analysis of dynamic styles of inference in linguistics and
cognition in van Benthem (1996).) Perhaps the first example of this
approach in abduction is the work in Flach (1995) – and indeed our
analysis in Aliseda (1997) follows this same pattern.
4.2. A Taxonomy for Abduction
What we have seen so far may be summarized as follows. Abduction is a general process of explanation, whose products are specific
explanations, with a certain inferential structure. We consider these
two aspects to be of equal importance. Moreover, on the process
side, we distinguished between constructing possible explanations
and selecting the best one amongst these.
As for the logical form of abduction, we have found that it may
be viewed as a threefold relation:
T, C ⇒ E
between an observation E, an abduced item C, and a background theory T. (Other parameters are possible here, such as a
preference ranking – but these would rather concern the further
selection process.) Against this background, we propose three
main parameters that determine types of abduction. (i) An ‘inferential parameter’ (⇒) sets some suitable logical relationship
among explananda, background theory, and explanandum. (ii) Next,
‘triggers’ determine what kind of abduction is to be performed:
E may be a novel phenomenon, or it may be in conflict with the
theory T. (iii) Finally, ‘outcomes’ (C) are the various products of an
abductive process: facts, rules, or even new theories.
In the above schema, the notion of explanatory inference ⇒ is
not fixed. It can be classical derivability ⊢ or semantic entailment
|=, but it does not have to be. Instead, we regard it as a parameter
which can be set independently. It ranges over such diverse values
LOGICS IN SCIENTIFIC DISCOVERY
353
as probable inference (T, C ⇒probable E), in which the explanans
renders the explanandum only highly probable, or as the inferential mechanism of logic programming (T, C ⇒prolog E). Further
interpretations include dynamic inference (T, C ⇒dynamic E, cf.
van Benthem, 1996), replacing truth by information change potential along the lines of belief update or revision. Our point here is
that abduction is not one specific non-standard logical inference
mechanism, but rather a way of using any one of these.
According to Peirce, abductive reasoning is triggered by a
surprising phenomenon. The notion of surprise, however, is a
relative one, for a fact E is surprising only with respect to some
backgound theory T providing ‘expectations’. What is surprising to
me (e.g. that the canal bridge floor goes up from time to time) might
not be surprising to a Dutch person. We interpret a surprising fact as
one which needs an explanation. From a logical point of view, this
assumes that the fact is not already explained by the background
theory T: T E.
Moreover, our claim is that one also needs to consider the status
of the negation of E. Does the theory explain the negation of observation instead (T ⇒ ¬E)? Thus, we identify at least two triggers for
abduction: novelty and anomaly:
Abductive Novelty: T E, T ¬E
E is novel. It cannot be explained (T E), but it is
consistent with the theory (T ¬E).
Abductive Anomaly: T E, T ⇒ ¬E
E is anomalous. The theory explains rather its negation
(T ⇒ ¬E).
In the computational literature on abduction, novelty is the condition
for an abductive problem (Kakas et al., 1993). My suggestion is to
incorporate anomaly as a second basic type.
Of course, non-surprising facts (where T ⇒ E) should not be
candidates for explanation. Even so, one might speculate if facts
which are merely probable on the basis of T might still need
explanation of some sort to further cement their status.
Abducibles themselves come in various forms: facts, rules, or
even theories. Sometimes one simple fact suffices to explain a
surprising phenomenon, such as rain explaining why the lawn is wet.
354
ATOCHA ALISEDA
In other cases, a rule establishing a causal connection might serve
as an explanation, as in our case connecting cloud types with rainfall. And many cases of abduction in science provide new theories
to explain surprising facts. These different options may sometimes
exist for the same observation, depending on how seriously we want
to take it.
Moreover, we are aware of the fact that genuine explanations
sometimes introduce new concepts, over and above the given
vocabulary. (For instance, the eventual explanation of planetary
motion was not Kepler’s, but Newton’s, who introduced a new
notion of ‘force’ – and then derived elliptic motion via the Law of
Gravity.) Abduction via new concepts will be outside the scope of
our analysis.
Once the above parameters get set, several kinds of abductive
processes arise. For example, abduction triggered by novelty with
an underlying deductive inference, calls for a process by which the
theory is expanded with an explanation. The fact to be explained is
consistent with the theory, so an explanation added to the theory
accounts deductively for the fact. However, when the underlying
inference is statistical, in a case of novelty, theory expansion might
not be enough. The added statement might lead to a ‘marginally
consistent’ theory with low probability, which would not yield a
strong explanation for the observed fact. In such a case, theory revision is needed (i.e. removing some data from the theory) to account
for the observed fact with high probability.
Our aim is not to classify abductive processes, but rather to point
out that several kinds of them are used for different combinations of
the above parameters (in Aliseda (1997) we explore in detail some
procedures for computing different types of outcomes in a deductive
format).
This taxonomy gives us the big picture of abductive reasoning.
We can now see the patterns in clearer focus. Varying the inferential
parameter, we cover not only cases of deduction but also statistical
inferences. Thus, Hempel’s statistical model of explanation also
becomes a case of abduction. As for triggers, novelty drives the
rain examples and trigger by anomaly occurs in the Kepler example,
since his initial observation of the longitudes of Mars contradicted
the previous rule of circular orbits of the planets. Different forms
LOGICS IN SCIENTIFIC DISCOVERY
355
of outcomes will play a role in different types of procedures for
producing explanations. In computer science jargon, triggers and
outcomes are, respectively, preconditions and outputs of abductive
devices, whether these be computational procedures or inferential
ones.
4.3. Abductive Logic: Inference + Search Strategy
Classical logical systems have two components: a semantics and
a proof theory. The former aims at characterizing what it is for a
formula to be true in a model, and it is based on the notions of
truth and interpretation. The latter characterizes what counts as a
valid proof for a formula, by providing the inference rules of the
system; having for its main notions proof and derivability. These
two formats can be studied independently, but they are closely
connected. At least in classical (first-order) logic, the completeness
theorem tells us that all true formulas have a proof, and the converse
claim made by the Soundness theorem. Many logical systems have
been proposed that follow this pattern: propositional logic, predicate
logic, modal logic, and various typed logics.
From a modern perspective, however, there is much more to
reasoning than this. Computer science has posed a new challenge
to logic; that of providing automatic procedures to operate logical
systems. This requires a further fine-structure of reasoning. In fact,
recent studies in AI give precedence to a control strategy for a
logic over its complete proof theory. In particular, the heart of
logic programming lies in its control strategies, which lead to much
greater sensitivity as to the order in which premises are given, the
avoidance of search loops, or the possibility to cut proof searches
(using the extra-logical operator!) when a solution has been found.
These features are extralogical from a classical perspective, but they
do have a clear formal structure, which can be brought out, and
has independent interest as a formal model for broader patterns of
argumentation (cf. van Benthem, 1996; Kalsbeek, 1995; Kowalski,
1979).
Several contemporary authors stress the importance of control
strategies, and a more finely-structured algorithmic description of
logics. This concern is found both in the logical tradition (Gabbay,
1994; van Benthem, 1996), and in the philosophical tradition
356
ATOCHA ALISEDA
(Gillies, 1996), the latter arguing for a conception of logic as: inference + control. (Note the shift here away from Kowalski’s famous
dictum “Algorithm = Logic + Control”.) In line with this philosophy,
we wish to approach abduction with two things in mind.
First, there is the inference parameter, already discussed, which
may have several interpretations. But given any specific choice,
there is still a significant issue of a suitable search strategy over
this inference, which models some particular abductive practice.
The former parameter may be defined in semantic or proof-theoretic
terms. The search procedure, on the other hand, deals with concrete
mechanisms for producing valid inferences. It is then possible
to control which kinds of outcome are produced with a certain
efficiency. In particular, in abduction, we may want to produce
only ‘useful’ or ‘interesting’ formulas, preferably even just some
‘minimal set’ of these. In this light, the aim of an abductive search
procedure is not necessarily completeness with respect to some
semantics. A procedure that generates all possible explanations
might be of no practical use, and might also miss important features
of human abductive activity.
5. DISCUSSION AND FURTHER CHALLENGES
Our positional claim for this paper has been that the study of the
context of discovery allows for a precise formal treatment. The
dominant trend in contemporary philosophy of science neglected the
study of the processes of theory discovery in science, partly because
other questions were on focus, still we knew nothing about the
canons of justification, and the historical approach to scientific practice was making its way into methodological issues. It was research
on scientific discovery as problem solving by computational means
that really brought back issues of discovery to the philosophy of
science agenda. And the field was ready for it, as philosophy
of science questions had also evolved. Issues of theory building,
concept formation, theory evaluation and scientific progress were
at issue. It is now time to unravel these questions with the hand
of logical analysis. And here again, we get inspiration, or rather
the techniques, from work in logics for common sense reasoning
and scientific reasoning developed for artificial intelligence. The
LOGICS IN SCIENTIFIC DISCOVERY
357
present situation resembles the time of pre-Fregean logic, when
the boundaries between logic and general methodology were still
rather fluid, as witnessed by the work of Bolzano in his several
conceptions of consequence relations, and later by Charles Peirce’s
investigations into styles of reasoning (deduction, induction, abduction) from a logical perspective. In our view, current post-Fregean
logical research is slowly moving back towards this same broader
agenda.
Our claim is that we must bring back logic into scientific methodology, on a par with computational and historical approaches. This
move would complement the existing historical account, and as far
as its relation to computational approaches in scientific discovery
is concerned, it may bring benefits in regard to giving a foundation to the heuristics, to the logic(s) behind the (human) processes
of discovery in science. Let me now answer the following question (raised by Pat Langley at the time of the conference): how to
integrate a logic in the algorithmic design for scientific discovery?
Well, we must deal with a logic which besides (or even aside
from) its semantics and proof theory, gives an account of the
search strategy tied to a discovery system. That is, we must provide
the automatic procedures to operate a logic, its control strategy,
its procedures to acquire new information without disturbing its
coherence and hopefully, achieve some learning in the end.
By putting forward a logic for scientific discovery we claim no
lack of rigour. But what is clear is that standard deductive logic
cannot account for abductive or inductive types of reasoning. And
the field is now ready to use new logics for this purpose. An
impressive body of results has been obtained about deduction – and
a general goal could be the study the wider field of human reasoning
while hanging on to these standards of rigour and clarity.
In order to glance at the logical structure of abductive reasoning,
we have proposed a general taxonomy. Several forms of abduction are obtained by instantiating three parameters: the kind of
reasoning involved (e.g., deductive, statistical), the kind of observation triggering the abduction (novelty or anomaly with respect to
some background theory), and the kind of explanations produced
(facts, rules, or theories). But there is much more to the logic of
abduction than what we have been able to describe in this paper. The
358
ATOCHA ALISEDA
product-process distinction in abduction is given by defining it as a
logical inference on the one hand, and by giving search strategies
to produce explanations on the other. However, there is no single
way to define either its inferential pattern or its search strategy (cf.
Aliseda (1997) for several characterizations of abductive inference
as well as for algorithms to produce abductions). The fact of the
matter is that abduction is a non-standard logical system: it is not
a single phenomenon but rather a general practice of explanatory
reasoning resulting in a family of logics.
There are still many challenges ahead for the formal study
of reasoning in scientific discovery, such as giving an integrated
account of deductive, inductive, abductive and analogical styles of
inference, the use of diagrams by logical means, and in general
the device of logical operations for theory building and change.
Already new logical research (some of it presented at MBR’01)
is moving into these directions. In Burger and Heidema (forthcoming) degrees of ‘abductive boldness’ are proposed as spectrum
for inferential strength, ranging from cases with poor background
information to those with (almost) complete information. Systems
dealing with several notions of derivability all at once have also
been proposed. A formula may be ‘unconditionally derived’ or
‘conditionally derived’, the latter case occurring when a line in a
proof asserts a formula which depends on hypotheses which may be
later falsified, thus pointing to a notion of proof which allows for
addition of lines which are non-deductively derived as well as for
deletion of them when falsifying instances occur. This account is
found in the framework of (ampliative) adaptive logics, a natural
home for abductive inference (cf. Meheus, forthcoming) and for
(enumerative) induction inference (cf. Batens, forthcoming) alike.
The formalization of analogical reasoning is still a new area of
research (I do not know of any rigorous account for it), without
even a clear idea of what exactly an analogy amounts to. Perhaps
research on mathematical analogy (Polya, 1945) or investigations
into analogical argumentation theory recently proposed for abduction (Gabbay and Woods, forthcoming), may serve to guide research
in this direction. Finally, the study of ‘diagrammatic reasoning’ is
a research field in its own right (Barwise and Etchemendy, 1995),
showing that the logical language is not restricted to the two-
LOGICS IN SCIENTIFIC DISCOVERY
359
dimensional left to right syntactic representation, but its agenda still
needs to be expanded on research for non-deductive logics.
As for formal approaches for theory building and change applied
to philosophy of science, my present research in this direction
concerns the extension of a classical logical method to model empirical progress in science, as conceived by Theo Kuipers (1999, 2000).
In particular, the goal is to operationalize the task of instrumentalist
abduction, that is, theory revision aiming at empirical progress. My
account shows that evaluation and improvement of a theory can be
modeled by (an extension of) the framework of Semantic Tableaux
(Aliseda, forthcoming).
All the above suggests that the use of non-standard logics to
model processes in scientific practice, such as confirmation, falsification, explanation building and theory improvement is, after all, a
feasible project. Nevertheless, this claim requires a broad conception of what logic is. Still, many logicians remain doubtful, and
withhold the status of bona fide ‘logical inference’ to the products of
non-standard styles. This situation is somewhat reminiscent of the
emergence of non-Euclidean geometries in the nineteenth century.
Euclidean geometry was thought of as the one and only geometry
until the fifth postulate (the parallel axiom) was rejected, giving rise
to new geometries. Most prominently are the one by Lobachevsky
which admits of more than one parallel, and the one by Riemann
admitting none.2 The legitimacy of these geometries was initially
doubted but their impact gradually emerged. The analogy with logic
can be carried out even further, as these new geometries were
sometimes labeled ‘meta-geometries’ (cf. Aliseda, 1997).
In our view, all these alternative logics might get their empirical
vindication, too – as reflecting different modes of human reasoning.
Whether non-classical modes of reasoning are really logical is like
asking if non-Euclidean geometries are really geometries. The issue
is largely terminological, and we may decide – as Quine (1961) did
on another occasion – to just give conservatives the word ‘logic’ for
the more narrowly described variety, using the word ‘reasoning’ or
some other suitable substitute for the wider brands.
360
ATOCHA ALISEDA
NOTES
1. A consequence ⇒ is non-monotonic whenever T ⇒ b does not ensure T,
a ⇒ b. That is, the addition of new premises (a) is no warantee for validity
preservation.
2. In fact, the first to invent non-Euclidean geometry with no parallels was
Gauss, followed by Bolyai, and then Lobachevsky. Riemann invented a more
general system, including those of Gauss et al., but also allowing geometries
with an infinite number of parallels (D. Atkinson, personal communication).
REFERENCES
Aliseda, A., 1997: Seeking Explanations: Abduction in Logic, Philosophy of
Science and Artificial Intelligence. Dissertation Stanford University, Published
by the Institute for Logic, Language, and Computation (ILLC). Universiteit van
Amsterdam.
Aliseda, A.: 2000, Abduction as Epistemic Change: A Peircean Model in Artificial Intelligence. In P. Flach and A. Kakas (eds.), Abductive and Inductive
Reasoning: Essays on their Relation and Integration. Kluwer Academic Press.
Aliseda, A.: forthcoming, Lacuna, Empirical Progress and Semantic Tableaux. In
Logics of Scientific Cognition: Essays in Debate with Theo Kuipers, Poznan
Studies in the Philosophy of the Sciences and the Humanities.
Barwise, J. and J. Etchemendy: 1995, Hyperproof. Center for the Study of
Language and Information. Lecture Notes Series 42.
Batens, D.: forthcoming, A Logic of Induction. In Logics of Scientific Cognition:
Essays in Debate with Theo Kuipers, Poznan Studies in the Philosophy of the
Sciences and the Humanities.
van Benthem, J.: 1984, The Variety of Consequence, According to Bolzano,
Studia Logica 44: 389–403.
van Benthem, J.: 1991, Logic as Programming, Fundamenta Informaticae 17(4):
285–318.
van Benthem, J. and G. Cepparello: 1994, Tarskian Variations: Dynamic Parameters in Classical Semantics, Technical Report CS–R9419. Amsterdam:
Centre for Mathematics and Computer Science (CWI).
van Benthem, J.: 1996, Exploring Logical Dynamics. CSLI and Cambridge
University Press. Studies in Logic, Language and Information Series.
Beth, E.W.: 1959, The Foundations of Mathematics. Amsterdam: North-Holland
Publishing Co.
Bolzano, B.: 1973, Wissenschaftslehre, Seidel Buchhandlung, Sulzbach. Theory
of Science, B. Torrel (trans.), J. Berg (ed.). The Netherlands, Dordrecht: D.
Reidel Publishing Company.
Burger, I.C. and J. Heidema: forthcoming, Degrees of Abductive Boldness. In L.
Magnani, N. Nersessian and C. Pizzi (eds.), Logical and Computational Aspects
of Model-Based Reasoning. Kluwer Aplied Logic Series.
LOGICS IN SCIENTIFIC DISCOVERY
361
Corcoran, J.: 1975, Meanings of Implication, Diaglos 9: 59–76.
Flach, P.A.: 1995, Conjectures: an Inquiry Concerning the Logic of Induction,
PhD Thesis, Tilburg University.
van Frassen, B.: 1980, The Scientific Image. Oxford: Clarendon Press.
Gabbay, D.M.: 1994, Classical vs Non-Classical Logics (the Universality of
Classical Logic). In D.M. Gabbay, C.J. Hogger and J.A. Robinson (eds.), Handbook of Logic in Artificial Intelligence and Logic Programming. Volume 2
Deduction Methodologies. Clarendon Press and Oxford Science Publications,
359–500.
Gabbay, D.M. and J. Woods: forthcoming, The Reach of Abduction.
Gärdenfors, P.: 1988, Knowledge in Flux: Modeling the Dynamics of Epistemic
States. MIT Press.
Gillies: 1996, Artificial Intelligence and Scientific Method. Oxford University
Press.
Groeneveld, W.: 1995, Logical Investigations into Dynamic Semantics. PhD
Dissertation, Institute for Logic, Language and Information, University of
Amsterdam.
Hanson, N.R.: 1961, Patterns of Scientific Discovery. Cambridge: The University
Press.
Josephson, J.R.: 1994, Abductive Inference. Cambridge University Press.
Kakas, A.C., R.A. Kowalski and F. Toni: 1993, Abductive Logic Programming,
Journal of Logic and Computation 2(6): 719–770.
Kalsbeek, M.: 1995, Meta Logics for Logic Programming. PhD Dissertation,
Institute for Logic, Language and Information, University of Amsterdam.
Konolige: 1990, A General Theory of Abduction. in Automated Abduction,
Working Notes. Spring Symposium Series of the AAA, Stanford University,
62–66.
Kowalski, R.A.: 1979, Logic for Problem Solving. New York: Elsevier.
Kuipers, T.: 1999, Abduction Aiming at Empirical Progress or Even at Truth
Approximation, Foundations of Science 4: 307–323.
Kuipers, T.: 2000. From Instrumentalism to Constructive Realism: On Some Relations Between Confirmation, Empirical Progress, and Truth Approximation.
Kluwer Academic Publishers: Synthese Library.
Kraus, S., D. Lehmann and M. Magidor: 1990, Nonmonotonic Reasoning,
Preferential Models and Cumulative Logics, Artificial Intelligence 44: 167–207.
Lakatos, I.: 1976, Proofs and Refutations. the Logic of Mathematical Discovery.
Cambridge University Press.
Langley, P., H.A. Simon, G.L. Bradshaw and J.M. Zytkow: 1987, Scientific
Discovery: Computational Explorations of the Creative Processes. Cambridge,
MA: MIT Press.
Mayer, M.C. and F. Pirri: 1993, First Order Abduction via Tableau and Sequent
Calculi, Bulletin of the IGPL 1: 99–117.
Meheus, J. and T. Nickles: 1999, Scientific Discovery and Creativity: Methodological Approaches, Foundations of Science (4): 3.
362
ATOCHA ALISEDA
Meheus, J.: forthcoming, Ampliative Adaptative Logics and the Foundation of
Logic-Based Approaches to Abduction. In L. Magnani, N. Nersessian and C.
Pizzi (eds.), Logical and Computational Aspects of Model-based Reasoning.
Kluwer Aplied Logic Series.
Michalski, R.: 1994, Inferential Theory of Learning: Developing Foundations
for Multistrategy Learning. In Machine Learning: A Multistrategy Approach.
Morgan Kaufman Publishers.
Paul, G.: 1993, Approaches to Abductive Reasoning: An Overview, Artificial
Intelligence Review 7: 109–152.
Peirce, C.S.: 1958, Collected Papers of Charles Sanders Peirce, volumes 1–6, C.
Hartshorne and P. Weiss (eds.). Cambridge: Harvard University Press, 1931–
1935; and volumes 7–8, A.W. Burks (ed.). Cambridge: Harvard University
Press.
Poole, D., R.G. Goebel and Aleliunas: 1987, Theorist: A Logical Reasoning
System for Default and Diagnosis. In Cercone and McCalla (eds.), The Knowledge Fronteer: Essays in the Representation of Knowledge. Springer Verlag
Lecture Notes in Computer Science, 331–352.
Perry, J. and J. Barwise: 1983, Situations and Attitudes. Cambridge, MA: MIT
Press.
Polya, G.: 1945, How to Solve It. A New Aspect of Mathematical Method.
Princeton University Press.
Popper, K.: 1963, Conjectures and Refutations. The Growth of Scientific Knowledge, 5th edn. London/New York: Routledge.
Popper, K.: 1958, The Logic of Scientific Discovery. London: Hutchinson.
Pople, H.E.: 1973, On the Mechanization of Abductive Logic. In Proceedings of
the Third International Joint Conference on Artificial Intelligence. San Mateo:
IJCAI; Stanford, CA: Morgan Kauffmann, 147–152.
Quine, W.V.: 1961, On What There Is. In From a Logical Point of View, 2nd edn.
rev. Cambridge and Harvard University Press.
Reichenbach, H.: 1938, Experience and Prediction. Chicago: University of
Chicago Press.
Reiter, R.: 1980, A Logic for Default Reasoning, Artificial Intelligence 13.
Rescher, N.: 1978, Peirce’s Philosophy of Science. Critical Studies in His Theory
of Induction and Scientific Method. University of Notre Dame.
Ruben, D-H.: 1990, Explaining Explanation. London/New York: Routledge.
Salmon, W.: 1977, A Third Dogma of Empiricism. In Butts and Hintikka (eds.),
Basic Problems in Methodology and Linguistics. Dordrecht: Reidel.
Salmon, W.: 1990, Four Decades of Scientific Explanation. University of
Minnesota Press.
Shoham, Y.: 1988, Reasoning about Change. Time and Causation from the
Standpoint of Artificial Intelligence. Cambridge, MA: The MIT Press.
Simon, H.A., P. Langley and G.L. Bradshaw: 1981, Scientific Reasoning as
Problem Solving, Synthese 47: 1–27.
Suppes, P.C.: 1979, Estudios de Filosofía y metodología de la ciencia. Alianza
Editorial.
LOGICS IN SCIENTIFIC DISCOVERY
363
Tan, Y.H.: 1995, Non-Monotonic Reasoning: Logical Architecture and Philosophical Applications, PhD Thesis, University of Amsterdam.
Thagard, P.: 1988, Computational Philosophy of Science. Cambridge: MIT Press.
Thagard, P.: 1992, Conceptual Revolutions. Princeton University Press.
Thompson, P.: 1981, Bolzano’s Deducibility and Tarski’s Logical Consequence,
History and Philosophy of Logic 2: 11–20.
Instituto de Investigaciones Filosóficas
UNAM, México
E-mail: atocha@filosoficas.unam.mx
and
Faculteit der Wijsbegeerte
RuG, Groningen
The Netherlands
E-mail: atocha@philos.rug.nl