Josephson Abduction
Josephson Abduction
Josephson Abduction
ABDUCTIVE INFERENCE
some of its relationships with other traditionally recognized forms of inference. This is followed in chapter 2 by an orientation to our view of AI as a
science and to . . <?ur_apprpach to buildin_g knowledge systems. The remainder
of the book traces the development of Six generations of abduction machines
and describes some of the discoveries that we made about the dynamic logic
of abduction.
What is abduction?
Jo s-Q_r~~cn--, To ~IA
Svst<.~
oc.,,_~
J a ~(jlA.wi-,..
rp\-..~ laSolf~;
cu..
W\
1tor ,--tl<:P
et:
"'-C("
cJl "t=e:zix
bed-,_ ~ j
u'V\ ~ v-R.
rs)
7j C?~
ABDUCTIVE INFERENCE
pretations, and perceivings of ordinary life and in the more critically selfaware reasonings upon which scientific theories are based. Sometimes abductions are deliberate, such as when the physician, or the mechanic, or the
scientist, or the detective forms hypotheses explicitly and evaluates them to
find the best explanation. Sometimes abductions are more perceptual, such
as when we separate foreground from background planes in a scene, thereby
inaking sense of the disparities between the images formed from the two
eyes, or when we understand the meaning of a sentence and thereby explain
the presence and order of the words.
Abduction in ordinary life
Abductive reasoning is quite ordinary and commonsensical. For example, as
Harman (I 965) pointed out, when we infer from a person's behavior to some
fact about her mental state, we are inferring that the fact explains the behavior better than some other competing explanation does. Consider this specimen of ordinary reasoning:
JoE: Why are you pulling into the filling station?
TrDMARSH:
Because the gas gauge indicates nearly empty. Also, I have no reason to
think that the gauge is broken, and it has been a long time since I filled the tank.
TroMARSH:
Uiider the circuriistarices, the nearly empty gas tank is the best available
explanation for the gauge indication. Tidmarsh's other remarks can be understood as being directed to ruling out a possible competing explanation
(brnken gauge) and supporting the plausibility of the preferred explanation.
Consider another example of abductive reasoning: Imagine that one day you
are driving your car, and you notice the car behind you because of its peculiar
shade of bright yellow. You make two turns along your accustomed path home. ward and then notice that the yellow car is still behind you, but now it is a little
farther away. Suddenly, you remember something that you left at the office and
decide to turn around and go back for it. You execute several complicated manetivers to reverse your direction and return to the office. A few minutes later
you notice the. same yellow car behind you. You conceive the hypothesis that
you .are being followed, but you cannot imagine any reason why this should be
so that seems to have any significant degree of likelihood. So, you again reverse
direction, and observe that the yellow car is still behind you. You conclude that
you are indeed being followed (reasons unknown) by the person in the dark
glasSes in the yellow car. There is no other plausible way to explain why the car
remains continually behind you. The results of your experiment of reversing
direction a second time served to rule out alternative explanations, such as that
the other driver's first reversal of direction was a coincidence of changing plans
at the same time.
Harman (1965) gave a strikingly insightful analysis of law court testimony, which argues that when we infer that a witness is telling the truth, we
are using best-explanation reasoning. According to Harman our inference
goes as follows:
(i) We infer that he says what he does because he believes it.
(ii) We infer that he believes what he does because he actually did witness the
situation which he describes.
looked at it, and it looked just like honey. "But you never can tell," said Pooh. "I
remember my uncle saying once that he had seen cheese just this colour." So he put
his tongue in, and took a large lick. (pp. 61-62)
Pooh's hypothesis is that the substance in the jar is honey, and he has two
pieces of evidence to substantiate his hypothesis: It looks like honey, and
"bunny" is written on the jar. How can this be explained except by supposing that the substance is honey? He considers an alternative hypothesis: It
might be cheese. Cheese has been observed to have this color, so the cheese
hypothesis offers another explanation for the color of the substance in the
jar. So, Pooh (conveniently dismissing the evidence of the label) actively
seeks evidence that would distinguish between the hypotheses. He performs
a test, a crucial experiment. He takes a sample.
The characteristic reasoning processes of fictional detectives have also
been characterized as abduction (Sebeok & Umiker-Sebeok, I 983). To use
another example from Hannan (1965), when a detective puts the evidence
together and decides that the culprit must have been the butler, the detective
is reasoning that no other explanation that accounts for all the facts is plausible enough or simple enough to be accepted. Truzzi ( 1983) alleges that at
least 217 abductions can be found in the Sherlock Holmes canon.
"There is no great mystery in this matter," he said, taking the cup of tea which I had
poured out for him; "the facts appear to admit of only one explanation."
- Sherlock Holmes (Doyle, I 890, p. 620)
Abduction in science
ABDUCTIVE INFERENCE
ability to explain not only the motion of the planets, but also the occurrence
of the tides. In On the Origin of Species by Means of Natural Selection
Darwin presented what amounts to an extended argument for natural selection as the best hypothesis for explaining the biological and fossil evidence
at hand. Harman (1965) again: when a scientist infers the existence of atoms
and subatomic particles, she is inferring the truth of an explanation for her
various data. Science News (Peterson, 1990) reported the attempts of astronomers to explain a spectacular burst of X rays from the globular cluster
MIS on the edge of the Milky Way. In this case the inability of the scientists
to come up with a satisfactory explanation cast doubt on how well astronomers understand what happens when a neutron star accretes matter from an
orbiting companion star. Science News (Monastersky, 1990) reported attempts
to explain certain irregular blocks of black rock containing fossilized plant
matter. The best explanation appears to be that they are dinosaur feces.
Abduction ond history
Knowledge of the historical past also rests on abductions. Peirce (quoted in
Fann, 1970) cites one example:
Numberless documents refer to a conqueror called Napoleon Bonaparte. Though we
have not seen the man, yet we cannot explain what we have seen, namely, all those
documents and monuments without supposing that he really existed. (p. 21)
:~
In this section we show by example how the abductive inference pattern can
be used simply and directly to des.cribe diagnostic reasoning and its justifications.
In AI, diagnosis is often described as an abduction problem (e.g., Peng &
Reggia, 1990). Diagnosis can be viewed as producing an explanation that
best accounts for the patient's (or device's) symptoms. The idea is th~t the
task of a diagnostic reasoner is to come up with a best explanation for the
symptoms, which are typically those findings for the case that show abnormal values. The explanatory hypotheses appropriate for diagnosis are malfunction hypotheses: typically disease hypotheses for plants and animals
and broken-part hypotheses for mechanical systems.
The diagnostic task is to find a malfunction, or set of malfunctions, that
best explains the symptoms. More specifically, a diagnostic conclusion should
explain the symptoms, it should be plausible, and it should be significantly
better than alternative explanations. (The terms "explain," "plausible," and
"better" remain undefined for now.)
Taking diagnosis as abduction determines the classes of questions th~t are
fair to ask of a diagnostician. It also suggests that computer-based diagnostic systems should be designed to make answering such questions straightforward.
10
ABDUCTIVE INFERENCE
3.
Hypotheses were incorrectly judged to be implausible. Perhaps venous congestion should have been considered more plausible than it was, due to faulty
knowledge or missing evidence.
4a. Hypotheses were incorrectly thought not to explain important findings. For
example, .obstruction might explain findings that the physician thought it could
not, possibly because the physician had faulty knowledge.
4b. The diagnostic conclusion was incorrectly thought to explain the findings.
Neoplasm might not explain the findings, due to faulty knowledge or to overlooking important findings.
5a. The diagnostic conclusion was incorrectly thought to be better than it was.
Neoplasm might have been overrated, due to faulty knowledge or missing
evidence.
5b. The true answer was underrated, due to faulty knowledge or missing evidence.
The real choice here seems to lie between an infection of the liver and neoplasm of
the liver. It seems to me that the course of the illness is compatible with a massive
hepatoma [neoplasm of the liver] and that the hepatomegaly, coupled with the biochemical findings, including the moderate degree of jaundice, are best explained by
this diagnosis.
thus trying to convince the questioner that venous congestion was correctly
ruled out. If asked, "Why not consider some toxic hepatic injury?" the physician could reply:
[It would not] seem to compete with a large hepatoma in explaining the massive
hepatomegaly, the hypoglycemia, and the manifestations suggestive of infection.
thus trying to convince the questioner that the differential is broad enough.
Interestingly, in this case Bordley's diagnosis was wrong. Autopsy revealed
that the patient actually had cancer of the pancreas. (To be fair, the autopsy
also found tumors in the liver, but pancreatic cancer was considered the
primary illness.) One significant finding in the case was elevated amylase,
which is not explained by neoplasm of the liver. So, if we asked the physician, "How do you account for the sharply elevated amylase?" his only possible reply would be:
Oops.
II
The diagnosis was inadequate because it failed to account for all the important findings (item 4b in the previous numbered list).
This analysis tells us that if we build an expert system and claim that it
does diagnosis, we can expect it to be asked certain questions. These are the
only questions that are fair to ask simply because it is a diagnostic system.
Other questions would not be about diagnosis per se. These other questions
might include requests for definitions of terms, exam-like questions that
check the system's knowledge about some important fact, and questions about
There was something wrong with the data such that it really did not need to be
explained. In this case, hepatomegaly might not have actually been present.
The differential was not broad enough. There might be causes of hepatomegaly
that were unknown to the physician, or that were overlooked by him.
-~
ABDUCTIVE INFERENCE
12
the implications of the diagnostic conclusion for treatment. Thus, the idea
13
ety of possible explanations but cannot be sure that we have covered all
plausibles. Under these circumstances we can assert a proposition of the
agnostician and the client. It defines a set of questions that any person, or
form of the first premise of the syllogism, but assert it only with a kind of
qualified confidence. Typically, too, alternative explanations can be discounted for one reason or another but not decisively ruled out. Thus abductive
inferences, in a way, rely on this particular deductively valid inference form,
but abductions are conclusive only in the limit.
Of course disjunctive syllogism fits any decision by enumeration of.alternatives and exclusion, not just abductions (where explanatory alterna_tives
are considered). From this it can be seen that abduction cannot be idenJified
with disjunctive syllogism.
One power of this analysis lies in controlling for error, in making explicit
the ways in which the conclusion can be wrong. A challenging question
implies that the questioner thinks that the answer might be wrong and that
the questioner needs to be convinced that it is not. A proper answer will
reassure the questioner that the suspected error has not occurred.
Doubt and certainty
Inference and logic
Inferences are movements of thought within the sphere ofbelief. 4 The function of inference is the acceptance (or sometimes rejection) of propositions
on the basis of purported evidence. Yet, inferences are not wholly or merely
all possible explanations for the data and that all but one of the alternative
explanations has been decisively ruled out. Typically, however, we will have
reasons to believe that we have considered all plausible explanations (i.e.,
those that have a significant chance of being true), but these reasons stop
short of being conclusive. We may have struggled to formulate a wide vari-
Ampliative inference
Like inductive generalizations, abductions are ampliative inferences; that is, at
the end of an abductive process, having accepted a best explanation, we -may
have more information than we had before. The abduction transcends the information of its premises and generates new information that was not previously
encoded there at all. This can be contrasted with deductions, which can be thought
of as extracting, explicitly in their conclusions, information that was already
implicitly contained in the premises. Deductions are truth preserving, whereas
successful abductions may be said to be truth producing.
This ampliative reasoning is sometimes done by introducing new vocabulary in the conclusion. For example, when we abd~ce that the patient. has
hepatitis because hepatitis is the only plausible way-to explain the jaundice,
we have introduced into the conclusion a new term, "hepatitis," which is
from the vocabulary of diseases and not part of the vocabulary of symptoms.
By introducing this term, we make conceptual connections with the typical
progress of the disease, and ways to treat it, that were unavailable before.
Whereas valid deductive inferences cannot contain terms in their co~clu
sions that do not occur in their premises, abductions can "interpret" the given
data in a new vocabulary. Abductions can thus make the leap from "observation language" to "theory language."
1641). Since low-plausibility alternative explanations can be generated indefinitely, doubt cannot be completely eliminated.
On the way to a satisfactory explanation, an abductive process might seek
further information beyond that given in the data initially to be explained.
For example, there may be a need to distinguish between explanatory alternatives; for help in forming hypotheses; or for help in evaluating them. Often abductive processes are not immediately concluded, but are suspended
to wait for answers from information-seeking processes. Such suspensions
cif processing can last a very long time. Years later, someone may say, "So
that's why she never told me. I was always puzzled about that." Centuries
later we may say, "So that's the secret of inheritance. It's based on making
copies of long molecules that encode hereditary information."
Abductive conclusions: likelihood and acceptance
Emergent certainty
16
ABDUCTIVE INFERENCE
its links (although separate deductions can converge for parallel support).
For example, I may be more sure of the bear's hostile intent than of any of the
details of its hostile gestures; I may be more sure of the meaning of the sentence
than of my initial identifications of any of the words; I may be more sure of the
overall theory than of the reliability of any single experiment on which it is
based. Patterns emerge from individual points where no single point is essential
to recognizing the pattern. A signal extracted and reconstructed from a noisy
channel may lead to a message, the wording of which, or even more, the intent
of which, is more certain than any of its parts.
This can be contrasted with traditional empiricist epistemology, which
does not allow for anything to be more certain than the observations (except
maybe tautologies) since everything is supposedly built up from the observations by deduction and inductive generalization. But a pure generalization is always somewhat risky, and its conclusion is less certain than its
premises. "All goats are smelly" is less certain than any given "This goat is
smelly." With only deductive logic and generalization available, empirical
knowledge appears as a pyramid whose base is particular experiments or
sense perceptions, and where the farther up you go, the more general you
get, and the less certain. Thus, without some form of certainty-increasing
inference, such as abduction, traditional empiricist epistemology is unavoidably committed to a high degree of skepticism about all general theories of
science.
Knowledge without certainty
The conclusion of an abduction is "logically justified" by the force of the
abductive argument. If the abductive argument is strong, and if one is persuaded by the argument to accept the conclusion, and if, beyond that, the
conclusion turns out to be correct, then one has attained justified, true, belief, the classical philosophical conditions of knowledge, that date back to
Plato. 8 Thus abductions are knowledge producing inferences despite their
fallibility. Although we can never be entirely sure of an abductive conclusion, if the conclusion is indeed true, we may be said to "know" that conclusion. Of course, without independent knowledge that the conclusion is true,
we do not "know that we know," but that is the usual state of our knowledge.
Summary: Abductions are fallible, and doubt cannot be completely eliminated. Nevertheless, by the aid of abductive inferences, knowledge is possible even in the face of uncertainty.
Explanations give causes
There have been two main traditional attempts to analyze explanations as
deductive proofs, neither attempt particularly successful.Aristotle maintained
that an explanation is a syllogism of a certain form (Aristotle c. 350 B.C.)
17
that also satisfies various informal conditions, one of which is that the "middle
term" of the syllogism is the cause of the thing being explained. (B is the
middle term of "All A are B ; All Bare C; Therefore, All A are C .") More
recently (considerably) Hempel (1965) modernized the logic and proposed
the "covering law" or "deductive nomological" model of explanation. 9 The
main difficulty with these accounts (besides Hempel 's confounding the question of what makes an ideally good explanation with the question of what it
is to explain at all) is that being a deductive proof is neither necessary nor
sufficient for being an explanation. Consider the following:
QUESTION:
EXPLANATION:
18
ABDUCTIVE INFERENCE
and fonnal cause (Aristotle, Physics , bk. 2, chap. 3). Let us take the example of my coffee mug. The efficient cause is the process by which the
mug was manufactured and helps explain such things as why there are ripples
on the surface of the bottom. The material cause is the ceramic and glaze,
which compose the mug and cause it to have certain gross properties such as
hardness. The final cause is the end or purpose, in this case to serve as a
container for liquids and as a means of conveyance for drinking. A finalcause explanation is needed to explain the presence and shape of the handle.
Formal cause is somewhat more mysterious - Aristotle is hard to interpret
here - but it is perhaps something like the mathematical properties of the
shape, which impose constraints resulting in certain specific other properties. That the cross-section of the mug, viewed from above, is approximately
a circle, explains why the length and width of the cross-section are approximately equal. The causal story told by an abductive explanation might rely
on any of these four types of causation. 11
When we conclude that a finding f is explained by hypothesis H, we say
more than just that H is a cause off in the case at hand. We conclude that
among all th~ vas~ causal ancestry off we will assign responsibility to H.
Typically, our reasons for focusing on Hare pragmatic and connected rather
directly with goals of production or prevention. We blame the heart attack
on the blood clot in the coronary artery or on the high-fat diet, depending on
: our interests. Perhaps we should explain the patient's death by pointing out
that the patient was born, so what else can you expect but eventual death?
We can blame the disease on the invading organism, on the weakened im. mune system that permitted the invasion, or on the wound that provided the
'route of entry into the body. We can blame the fire on the presence of the
combustibles, on the presence of the spark, or even on the presence of the
oxygen, depending on which we think is the most remarkable. I suggest that
it comes down to this: The things that will satisfy us as accounting for /will
depend on why we are trying to account for f; but the only things that count
as candidates are parts of what we take to be the causal ancestry off
19
Harman (1965) argued that "inference to the best explanation" (i.e., abduction) is the basic form of nondeductive inference, subsuming "enumerative induction" and all other forms of nondeductive inferences as special
cases. Harman argued quite convincingly that abduction subsumes sampfeto-population inferences (i.e., inductive generalizations [this is my way of
putting the matter]). The weakness of his overall argument was that other
forms of nondeductive inference are not seemingly subsumed by abduction,
most notably population-to-sample inferences, a kind of prediction. The main
problem is that the conclusion of a prediction does not explain anything, so
the inference cannot be an inference to a best explanation.
This last point, and others, were taken up by Ennis (1968). In his reply to
Ennis, instead of treating predictions as deductive, or admitting them as a
distinctive form of inference not reducible to abduction, Harman took the
dubious path of trying to absorb predictions, along with a quite reasonable
idea of abductions, into the larger, vaguer, and less reasonable notion of
"maximizing explanatory coherence" (Harman, 1968). In this I think Harman
made a big mistake, and it will be my job to repair and defend Harman's
original arguments, which were basically sound, although they proved somewhat less than he thought.
-~
-~
Inductive generalization
First, I will argue that it is possible to treat every good (i.e., reasonable,
valid) inductive generalization as an instance of abduction. An inductive
generalization is an inference that goes from the characteristics of some
observed sample of individuals to a conclusion about the distribution of those
characteristics in some larger population. As Harman pointed out, it is useful to describe inductive generalizations as abductions because it helps to
make clear when the inferences are warranted. Consider the following inference:
All observed A's are B's
Induction
.~eirce's
view was that induction, deduction, and abduction are three distinct
of inference, although as his views developed, the boundaries shifted
somewhat, and he occasionally introduced hybrid forms such as "abductive
induction" (Peirce, 1903). In this section I hope to clear up the confusion
about the relationship of abduction to induction. First I argue that inductive
ge.neralizations can be insightfully analyzed as special cases of abductions.
I also argue that predictions are a distinctive form of inference, that they are
not abductions, and that they are sometimes deductive, but typically not.
The result is a new classification of basic inference types.
~~ypes
This inference is warranted, Harman (1965) writes," ... whenever the hypothesis that all A's are B's is (in the light of all the evidence) a better,
simpler, more plausible (and so forth) hypothesis than is the hypothesis, say,
that someone is biasing the observed sample in order to make us think that
all A's are B's. On the other hand, as soon as the total evidence makes some
other competing hypothesis plausible, one may not infer from the past correlation in the observed sample to a complete correlation in the total population."
20
ABDUCTIVE lNFEKENCE
If this is indeed an abductive inference, then "All A's are B's" should explain "All observed A's are B's." But, "All A's are B's" does not seem to
explain why "This A is a B," or why A and B are regularly associated (as
pointed out by Ennis, 1968). Furthermore, I suggested earlier that explanations give causes, but it is hard to see how a general fact could explain its
instances, because it does not seem in any way to cause them.
The story becomes much clearer if we distinguish between an event of
observing some fact and the fact observed. What the general statement in
the conclusion explains is the events of observing, not the facts observed.
For example, suppose I choose a ball at random (arbitrarily) from a large hat
containing colored balls. The ball I choose is red. Does the fact that all of
the balls in the hat are red explain why this particular ball is red? No. But it
does explain why, when I chose a ball at random, it turned out to be a red
one (because they all are). "All A's are B's" cannot explain why "This A is a
B" because it does not say anything at all about how its being an A is connected with its being a B. The information that "they all are" does not tell
me anything about why this one is, except it suggests that if I want to know
why this one is, I would do well to figure out why they all are.
A generalization helps to explain the events of observing its instances,
but it does not explain the instances themselves. That the cloudless, daytime
sky is blue helps explain why, when I look up, I see the sky to be blue (but it
doesn't explain why the sky is blue). The truth of "Theodore reads ethics
books a lot" helps to explain why, so often when I have seen him, he has
been reading an ethics book (but it doesn't explain why he was reading ethics books on those occasions). Seen this way, inductive generalization does
have the form of an inference whose conclusion explains its premises.
Generally, we can say that the frequencies in the larger population, together with the frequency-relevant characteristics of the method for drawing a sample, explain the frequencies in the observed sample. In particular,
"A's are mostly B's" together with "This sample of A's was drawn without
regard to whether or not they were B's" explain why the A's that were drawn
were mostly B's.
Why were 61 % of the chosen balls yellow?
Because the balls were chosen more or less randomly from a population that was
two thirds yellow (the difference from 2/3 in the sample being due to chance).
21
teristics in the sample is explained, not why these particular balls are yellow
or why the experiment was conducted on Tuesday. The explanation explains
why the sample frequency was the way it was, rather than having some markedly different value. In general, if there is a deviation in the sample from
what you would expect, given the population and the sampling method, then
you have to throw some Chance into the explanation (which is more or less
plausible depending on how much chance you have to suppose ). 12
The objects of explanation - what explanations explain - are facts about
the world (more precisely, always an aspect of a fact, under a description).
Observations are facts; that is, an observation having the characteristics that
it does is a fact. When you explain observed samples, an interesting thing is
to explain the frequencies. A proper explanation will give a causal story of
how the f'l'equencies came to be the way they were and will typically refer
both to the population frequency and the method of drawing the samples.
Unbiased sampling processes tend to produce representative outcomes;
biased sampling processes tend to produce unrepresentative outcomes. This
"tending to produce" is causal and supports explanation and predictio11. A
peculiarity is that characterizing a sample as "representative" is characterizing the effect (sample frequency) by reference to part of its cause (population frequency). Straight inductive generalization is equivalent to concluding that a sample is representative, which is a conclusion about its case.
This inference depends partly on evidence or presumption that the samp~ing
process is (close enough to) unbiased. The unbiased .~ampling process is Part
of the explanation of the sample frequency, and any independent evidence
for or against unbiased sampling bears on its plausibility as part of the. explanation.
If we do not think of inductive generalization as abduction, we are at a
loss to explain why such an inference is made stronger or more warranted, if
in collecting data we make a systematic search for counter-instances and
cannot find any, than it would be if we just take the observations passively.
Why is the generalization made stronger by making an effort to examine a
wide variety of types of A's? The inference is made stronger because the
failure of the active search for counter-instances tends to rule out various
hypotheses about ways in which the sample might be biased.
In fact the whole notion of a "controlled experiment" is covertly based on
abduction. What is being "controlled for" is always an alternative way of
explaining the outcome. For example a placebo-controlled test of the efficiency of a drug is designed to make it possible to rule out purely psychological explanations for any favorable outcome.
Even the question of sample size for inductive generalization can be seen
clearly from an abductive perspective. Suppose that on each of the only two
occasions when Konrad ate pizza at Mario's Pizza Shop, he had a stomachache the next morning. In general, Konrad has a stomachache occasionally
ADDUCTIVE lNFERENCE
22
Observations--> At least generally A's are B's--> The next A will be a B."
but not frequently. What may we conclude about the relationship between the
pizza and the stomachache? What may we reasonably predict about the outcome of Konrad's next visit to Mario's? Nothing. The sampleis not a large
enough. Now suppose that Konrad continues patronizing Mario's and that after
every one of 79 subsequent trips he has a stomach ache within 12 hours. What
may we conclude about the relationship between Mario's pizza and Konrad's
stomachache? That Mario's pizza makes Konrad have stomachaches. We may
predict that Konrad will have a stomachache after his next visit, too.
A good way to understand what is occurring in this example is by way of
abduction. After Konrad's first two visits we could not conclude anything
because we did not have enough evidence to distinguish between the two
competing general hypotheses:
1.
2.
Lj
This inference is stronger in that it establishes its 'conclusion with more certainty, which it does by hedging the generalization and thus making it more
plausible, more likely to be true. It could be made stronger yet by hedging
the temporal extent of the generalization:
Observations -+ At least generally A's are B's, at least for the recent past
and the immediate future -+ The next A will be a B.
The analyses of inductive projection with the hedged generalizations are
better than the first analysis because they are better at making sense of the
inference, which they do by being better at showing the sense in it (i.e., they
are better at showing how, and when, and why the inference is justified - or
"rational" or "intelligent"). Reasonable generalizations are hedged. Generally the best way to analyze "A's are B's" is not "All A's are B's," as we are
taught in logic class, but as "Generally A's are B's," using the neutral, hedged,
universal quantifier of ordinary life. 13
We have analyzed inductive projections as inductive generalizations followed
by predictions. The inductive generalizations are really abductions, as was argued before. But, what kind of inferences are predictions? One thing seems
clear: Predictions from hedged generalizations are not deductions.
Predictions from hedged generalizations belong to the same family as statistical syllogisms which have forms like these: 14
The eating pizza- stomachache correlation was accidental (i.e., merely coincidental or spurious [say, for example, that on the first visit the stomach ache
was caused by a virus contracted elsewhere and that on the second visit it was
caused by an argument with his mother]).
There is some connection between eating pizza and the subsequent stomach
ache (i.e., there is some causal explanation of why he gets a stomach ache
after eating the pizza [e.g., Konrad is allergic to the snake oil in Mario's
Special Sauce]).
By the time we note the outcome of Konrad's 79th visit, we are able to
decide in favor of the second hypothesis. The best explanation of the correlation has become the hypothesis of a causal connection because explaining
the correlation as accidental becomes rapidly less and less plausible the longer
the association continues.
Prediction
Another inference form that has often been called "induction" is given by
the following:
Therefore, Sis a B.
and
Observations --> All A's are B's --> The next A will be a B.
Predictions have traditionally been thought of as deductive inferences. However, something is wrong with this analysis. To see this, consider the alternative analysis of inductive projections, as follows:
-"
24
ABDUCTIVE INFERENCE
'
25
abduction-
prediction
data space
---
available and unnecessary for reasoning. People are good abductive reasoners
without close estimates of confidence. In fact it can be argued that, if confidences need to be estimated closely, then it must be that the best hypothesis
is not much better than the next best, in which case no conclusion can be
confidently drawn because the confidence of an abductive conclusion depends on how decisively the best explanation surpasses the alternatives. Thus
it seems that confident abductions are possible only if confidences for hypotheses do not need to be estimated closely.
Moreover, it appears that accurate knowledge of probabilities is not commonly available because the probability associated with a possible event is
not very well defined. There is almost always a certain arbitrariness about
which reference class is chosen as a base for the probabilities; the larger the
reference class, the more reliable the statistics, but the less relevant they
are; whereas the more specific the reference class, the more relevant, but the
less reliable. (See Salmon, 1967, p. 92.) Is the likelihood that the next patient has the flu best estimated based on the frequency in all the people in
the world over the entire history of medicine? It seems better at least to
control for the season and to narrow the class to include people at just this
particular time of the year. (Notice that causal understanding is starting to
creep into the considerations.) Furthermore, each flu season is somewhat
different, so we would do better to narrow to considering people just this
year. Then, of course, the average patient is not the same as the average
person, and so forth, so the class should probably be narrowed further to
something such as this: people of this particular age, race, gender, and social status who have come lately to doctors of this sort. Now the only way
the doctor can have statistics this specific is to rely on his or her own most
recent experience, which allows for only rough estimates of likelihood because the sample is so small. There is a Heisenberg-like uncertainty about
the whole thing; the closer you try to measure the likelihoods, the more
approximate the numbers become. In the complex natural world the longrun statistics are often overwhelmed by the short-term trends, which render
the notion of a precise prior probability of an event inapplicable to most
common affairs.
Taxonomy of basic inference types
Considering its apparent ubiquity, it is remarkable how overlooked and
underanalyzed abduction is by almost 2,400 years of logic and philosophy.
According to the analysis given here, the distinction between abduction and
deduction is a distinction between different dimensions, so to speak, of inference. Along one dimension inference can be distinguished into deductive
and nondeductive inference; along another dimension inferences can be distinguished as abductive and predictive (and mixed) sorts of inferences. Ab-
28
ABDUCTIVE INFERENCE
duction absorbs inductive generalization as a subclass and leaves the predictive aspect of induction as a separate kind of inference. Statistical syllogism is a kind of prediction. This categorization of inferences is summarized in Figure 1.2.
From wonder to understanding
Learning is the acquisition of knowledge. One main form of learning starts
with wonder and ends in understanding. To understand something is to grasp
Peirce's taxonomv
deduction
induction
abduction
other
deduction
l
..
Notes
I
2
3
4
inference
abduction
/
inductive
generalization
prediction
mixed
/
statistical
"'-
syllogism
prediction
deductive
29
"'-
inductive
projection
5
6
7
8
30
ABDUCTIVE INFERENCE
9 For a brief summary of deductive and other models of explanation see Bhaskar (1981), and for a
history of recent philosophical accounts of explanation, see Salmon ( 1990).
I 0 For a well-developed historical account of the connections between ideas of causality and expla-
nation see Wallace (1972, 1974). Ideas of causality and explanation have been intimately linked
for a very long time.
11 What the types of causation and causal explanation are remains unsettled, despite Aristotle's best
efforts and those of many other thinkers. The point here is that a narrow view of causation makes
progress harder by obscuring the degree to which alt forms of causal thinking are fundamentally
similar.
12 "It is embarrassing to invoke such a wildly unlikely event as a chance encounter between the
entry probe and a rare and geographically confined methane plume, but so far we have eliminated all other plausible explanations" (Planetary scientist Thomas M. Donahue of the University of Michigan on the analysis of chemical data from a Pioneer probe parachuted onto the
planet Venus, reported in Science News for Sept. 12, 1992).
13 Note that this analysis suggests an explanation for why traditional mathematical logic has been
so remarkably unsuccessful in accounting for reasoning outside mathematics and the highly
mathematical sciences. The universal quantifier of logic is not the universal quantifier of ordinary life, or even of ordinary scientific thought.
14 I have not put likelihood qualifiers in the conclusions of any these forms because doing so would
at best postpone the deductive gap.
The science of AI
The field of artificial intelligence (AI} seems scattered and disunited with
several competing paradigms. One major controversy is between proponents
of symbolic AI (which represents information as discrete codes) and proponents of connectionism (which represents information as weighted connections between simple processing units in a network). Even within each of
these approaches there is no clear orthodoxy. Another concern is whether AI
is an engineering discipline or a science. This expresses an uncertainty about
the basic nature of AI as well as an uncertainty about methodology. If AI is
a science like physics, then an AI program is an experiment. As experiments,
perhaps AI programs should be judged by the standards of experiments. They
should be clearly helpful in confirming and falsifying theories, in determining specific constants, or in uncovering new facts. However, if AI is fundamentally engineering, AI programs are artifacts, technologies to be used. In
this case, there is no such reason for programs to have clear confirming or
falsifying relationships to theories. A result in AI would then be something
practical, a technique that could be exported to a real-world domain and
used. Thus, there is confusion about how results in AI should be judged,
what the role of a program is, and what counts as progress in AL
It has often been said that the plurality of approaches and standards in
AI is the result of the extreme youth of AI as an intellectual discipline. This
theory implies that with time the pluralism of AI wil1 sort itself out into a
normal science under a single paradigm. I suggest, instead, that as it ages
AI will continue to include diverse and opposing theories and methods.
This is because the reason for the pluralism is that programs are at the heart
of AI, and programs can be approached in four fundamentally different
ways: (I) An AI program can be treated as a technology to be used to solve
practical problems (AI as engineering); (2) it can be treated as an experiment or model (AI as traditional science); (3) an AI program can also be
treated as a real intelligence, either imitative of human intelligence (AI as
art) or (4) non-imitative (AI as design science). Because these four ways
The first section of this chapter on the science of AI is by Susan G. Josephson; the remaining sections are by B. Chandrasekaran.
31