No Model, No

Inference: A Bayesian
Primer on the Grue
Problem1
Ellwtt Sobm

1. Introduction
Stripped to its bare quintessence, the grue problem reduces to two issues.
The first concerns the relationship between generalizations and their in-
stances. If I've observed a number of emeralds before now and have found
each to be green, what epistemic difference is there between the following
two generalizations?
(ALLGREEN) All emeralds are green.
(ALLGRUE) All emeralds are grue.2
The second problem concerns the relationship betweenpredictions and their
precedents. If I've observed a number of emeralds before now and have found
each to be green, what epistemic difference is there between the following
two predictions?
(NEXTGREEN) The next emerald I examine will be green.
(NEXTGRUE) The next emerald I examine will be grue.
The phrase "epistemic difference" is vague and requires clarification.
However it is understood, it is meant to rule out some obvious differences
between the two hypotheses. The GREEN hypotheses are stated in a
familiar vocabulary, whereas the GRUE hypotheses make use of a made-up
word. Unless one is prepared to regard this difference between the
hypotheses as episizrnicafly relevant-as somehow relevant to what we ought
to believe-this obvious difference makes no difference as far as the grue
problem is c0ncerned.j Beyond this, it falls to any solution of the problem
to make the concept of epistemic difference precise.
I have med to state a minimum formulation of the grue problem because
the p b k m is too often conflated with some of the ideas that figured in
 A Baywian Primer on thc Grrrc Pmbkm 227
226 Elliott Sober

Goodman's proposed solution. Goodman held that ALLGREEN is con- reasoning, of other approaches. In addition, the broad lessons I will draw
firmed by its instances, whereas ALLGRUE is not. This goes beyond my from my Bayesian analysis are mbrrrt; they are very much in harmony with
minimal formulation of the problem, in that instance confirmation is just the conclusions that would be generated by non-Bayesian positions.
one concept of epistemic relevance. We should not assume at the outset that Bayesians think that observations confer probabilities on hypotheses and
this is where the relevant difference between the GREEN and GRUE that the mathematical idea of probability is the right way to measure the
hypotheses is to be found. Goodman also asserted that the two problems epistemic property of plausibility. Bayes's theorem describes how the
I've sketched bear an important connection to each other; he maintained ponrriorprobability of hypothesis H-the probability it has in the light of the
that ALLGREEN is confirmed by its instances only if NEXTGREEN is too. observation O-is a function of three other quantities:
Again, the grue problem should not be burdened with this assumption;
rather, this is a thesis that needs to be argued for explicitly. And finally,
Goodman claimed that these evidential issues bear an important connection
to the problem of distinguishing lawlike from accidental generalizations. P(O/H) is called the likelihood of H. It describes the probability that the
Goodman believed that lawlike generalizations (like ALLGREEN) are hypothesis H confers on the observations. Don't confuse P(O/H) with
confirmed by their instances, whereas accidental generalizations (like P(H/O); the likelihood of H and the probability of H are different. P(H) is
ALLGRUE) are not.' Having separated these Goodmanian theses from the termed the prior probability of H, meaning the probability the hypothesis
problem they are meant to address, I will argue that each is incorrect and can has before the observation 0 is made.
be seen to be so by giving the grue problem a probabilistic representation. The grue problem, whether it is understood in terms of generalizations
During the time that Goodman wrote about the grue problem, philoso- or in terms of predictions, involves comparing two hypotheses in the light of
phers often focused on the problem of qwlitativc confinnation. Just as a body of observations. By bringing together two applications of Bayes's
Goodman had his emeralds, Hempel (1965) had his ravens. Hempel's theorem, we can derive a comparative principle of the kind required:
problem was not to measure thc d g ~ c cof c o n . t i o n that observing red
shoes provides for the generalization "all ravens are black." Rather, he (S) P(H,/O) > P(Hz/O) if and only if P(O/H,)P(H,) >
wanted to say whether such observations provide any confirmation at all. P(O/HZ)P(HZ).
Now, quite apart from the merits of what Goodman and Hempel said about
emeralds and ravens, it seems undeniable that a theory of qualitative If Hl is to have a higher posterior probability than H,, this must be because
confirmation should be embeddable in a theory of quantitative confirma- Hl has the higher likelihood or the higher prior probability (or both). So if
tion. We need the notion of dtgrc of confrmation, not just the dichotomous we are to find a difference between the GREEN and the GRUE hypotheses
division of c m . d versus not c o n . . The scientific community em- in this Bayesian format, we know that there are exactly two places to look.
ploys probability concepts in its understanding of this quantitative notion. Principle (S) is a synchronic principle. It does not describe how much
So a natural place to begin discussion of the grue problem is with the c h a w the observation 0 engenders in the probabilities of the hypotheses; it
concept of probability. simply describes what it takes for the one to have a higher value than the
other, after the observations are obtained. However, the dkbronic issue is
also worth considering. If P(H/O) represents the plausibility that H has
a@ 0 is found to be true, and P(H) is the plausibility that H possesses
2. Bayesianism-The Basics bcfm that event, then it is natural t o describe the c h a w in plausibility that
Those who agree that probability is a useful concept for explicating H experiences across this change as follows:
epistemic concepts nonetheless disagree about how it should be deployed.
Bayesianism constitutes an influential school in this debate, but it is not the 0 confirms H if and only if P(H/O) > P(H).
only game in town.5 I'll mention later on a standard criticism of 0 disconfirms H if and only if P(H/O) < P(H).
Bayesianism. But for now, Bayesianism is a perfectly sensible place to begin.
For m'any inference problems, the Bayesian solution is not at all conuover- Notice that 0 can confirm H even though P(H/O) is quite low; merely let
sial, i n that it happens to coincide with the verdicts, if not the exact P(H) = 0.0000001 and P(H/O) = 0.01. And 0 can disconfirm H wen
 A Bay& Prinur on the Grrrc Pmbkm 229

though H remains quite probable in the light of 0; merely let P(H) = 0.95 3. Generalization-The Diachronic
and P(H/O) = 0.94. Question
Two applications of these definitions yield the following diachronic
principle: Let's begin with a simple point about Bayesianism. If H deductively entails
0 and if 0 was not certain to be true before the observation was made, then
(D) 0 confirms H, more than 0 confirms H2 if and only if
0 confirms H. This is true because Bayes's thwrem can be rewritten as
P(H,/O) - P ( H J > P ( H J 0 ) - P(H2). follows:
I take the d : j h n c e between the posterior and .prior probabilities, rather
than the n;rtio between them, to represent'.degiee of confirmation. Ellery
Eells has suggested to me the following argument for this choice. Suppose
that 0 confirms H precisely when the left-hand side is greater than 1, which must
be so, if P(O/H) = 1 and P(0) < 1.
What does this mean about the grue problem? Here we must tread
and that carefully. Let us begin with the formulation that focuses on generalizations.
Notice that ALLGREEN and ALLGRUE both deductively imply that the
emeralds examined before now are green.' If it was not a certainty
If degree of confirmation is measured by taking differences, H, is confirmed beforehand that those emeralds should have turned out to be green, we
by 0 more than H2is. But if degree of confirmation is measured by the ratio must conclude that both hypotheses are confirmed by the past observations.
of posterior to prior, the reverse is true. Surely a jump from 0.1 to 0.9 Where the proposition E says that the sampled past emeralds are green, the
reflects a larger change in plausibility than a jump from 0.0001 to 0.001. In relevant facts are
any event the main conclusions I will argue for in what follows do not turn
on my choice of measure.
Although Goodman formulated his puzzle as one about confinnation,
there is no reason to restrict our attention to the diachronic issue. We want
to know if observing green emeralds raises the probabilities of the GREEN and
hypotheses more than such observations raise the probabilities of the GRUE
hypotheses. But we also would like to know whether the GREEN hypothe-
ses are more probable, in the light of these observations, than the GRUE
Both generalizations are confirmed by their instances if this simply means
hypotheses are. So there are four cases to consider-the [Link]
that each has its probability increased by the past observations.
the prediction problems each need to be considered both diachronically and
But now let us consider the degree of confirmation that each generaliza-
synchronically.
tion experiences. I begin by noting that
Even though the ideal of minimalism has guided me in my description
of the grue problem, I have to admit that I introduced a substantive choice
in my formulation. This involves the way I have described the evidence. I
imagine that our evidence was obtained by examining some emeralds and
finding out what color they exhibit. This search strategy differs from that of if and only if
sampling the universe at random and noting whether the things we come up
with are emeralds and what their colors are. With fairly modest assump
tions, it turns out that sampling the emeralds is a better strategy than
sampling the world at random (if the goal is to test either of the two
generalizations described a b o ~ e )I. assume
~ in what follows that this is how This latter inequality is true precisely when
our evidence was obtained. But once again, many of the broad lessons I will
draw fiom my analysis do not turn on the details of this assumption. P(ALLGREEN)(l/P(E) - 1) > P(ALLGRUE)(l/P(E) - l),
230 Elliott Sober A Baywian P r i m on the Grne Probkm 231
which simplifies to P(ALLGREEN) > P(ALLGRUE), if P(E) < 1. In other unattractive as general doctrines, I do not think we must concede that the
words, if we assume that the observations were not certain beforehand and problems at hand are insoluble. If we find ALLGREEN more plausible than
that each generalization implies the observations, then ALLGREEN is ALLGRUE, this is because we hold various substantive, if hard to articulate,
confirmed more than ALLGRUE if and only if the former has the higher theories about the world. Perhaps we expect emeralds to be alike in color
prior probability. because we think that they are alike in physical structure, and we believe that
What would make it plausible to assign ALLGREEN the higher prior color supervenes on physical structure. Of course, these convictions involve
probability? Rather than addressing that question here, I postpone it until assumptions about the future. But there is no escaping such commitments;
the next section. based solely on our experience of the past, the generalizations cannot be
shown to differ in their probabilities.
4. Generalization-The Synchronic
Question 5. Prediction-The Diachronic Question
As noted before, if we are sampling from the population of emeralds, then Let us now shift from the issue of generaliaion to the issue of prediction.
the generalizations ALLGREEN and ALLGRUE each deductively entail NEXTGREEN does not deductively imply that the emeralds examined
that the items sampled before now were green. Since the two hypotheses before now have been green, and neither does NEXTGRUE. So the
both have likelihoods of unity, principle (S) entails that a difference in arguments that solve the problems associated with the generalizations do
posterior probability must be due entirely to a difference in prior: not apply to the problems about prediction. Is there some other argument
that forces the same conclusion? Or are generalizationsand predictions not
If P(O/H,) = P(O/H,), then P(H,/O) > P(H,/O) if and only if as tightly coupled epistemologically ik Goodman and many others have
W , ) >Wl). thought?
To- answer this question. let us leave the strange and wonderful world of
-

So ALLGREEN is more probable than ALLGRUE, given the evidence E, grue behind for a ioment; and consider a rathe; more mundane inductive
precisely when the former hypothesis has the higher prior. problem. Imagine an urn that is filled with a thousand balls by drawing from
What could justify the belief that ALLGREEN has a higher prior a source whose composition is known. Suppose the source contains 50
probability than ALLGRUE? It is at this point in the story that I must percent red balls and 50 percent green balls. By random sampling from the
confess my anti-Bayesian sympathies. If prior probabilities are to be source, the urn is filled with a thousand balls. The problem is to sample
objective, I do not see how they can be assigned a priori. And if they are from the urn (with replacement) and to draw two inferences based on the
merely subjective-simply indicating the degree of belief of some agent- sample obtained. The first inference is to be a generalization concerning the
then I don't see that they have any epistemic relevance to this problem. One composition of the whole urn. The second is to be a prediction concerning
does not show that ALLGREEN is more plausible than ALLGRUE simply the color of balls that will be sampled in the future.
by giving voice to the autobiographical remark that one finds the former Since we know that the urn was composed by draws from the source, we
more plausible a priori than the latter. can assign prior probabilities to each of the possible compositions, from
Furthermore, I do not see how these hypotheses can have (objective) 1000 red and 0 green to 0 red and 1000 green. When we sample from
probabilities unless they describe possible outcomes of a chance process the urn, we can use Bayes's theorem to compute the posterior probability
(Edwards 1972).I do not know what it means to say that Newton's law of of the various hypotheses about the urn's composition. Suppose I take 250
gravitation or Darwin's thcory of evolution has an objective probability. draws from the urn and find that each ball I sample is green. These obser-
They were not made true by God's reaching into an urn that contained slips vations make the hypothesis that all the balls in the urn are green more
of paper on which candidate laws were inscribed. So if we can't specify a probable than it was initially. Just as was true for ALLGREEN and ALL
chance process that produces a coloration pattern for emeralds, I don't GRUE, the generalization is confirmed when a prediction deduced from the
know what it means to assign ALLGREEN and +[Link] objective prior generalization comes true, provided that the prediction was not certain
probabilities. . .
beforehand.
Having said that I find both objective and subjective Bayesianism But now let us consider the problem of predicting what the next ball will
232 EUwtt Sober A BayeGn Primer on the Gnu Pwbkm 233

be like, given information about the character of the sample. In this case, the require that one confirm a prediction about the next instance. Sampling
probability that the next ball will be green is 0.5, regardless of what the from the population of emeralds, it is inevitable (given the modest
previously sampled balls were like. Observing 250 green balls does not raise assumption that it was not a certainty that the sampled emeralds would turn
the probability that the next ball drawn from the urp will be green, even out to be green) that observing green emeralds before now should raise the
though the observations do raise the probability that all the balls are green. probability that all emeralds are green (ditto for grue). But no such
How can we make sense of the fact that- generalization and prediction part inevitability attaches to the prediction that the next emerald examined will
ways in this example? be green (or grue). Having separated these two problems, let us now explore
We can grasp the general point by considering the simple case in which the prediction problem on its own.
we make just two draws (with replacement) from the urn. There now are We have observed emeralds before now and found each to be green.
four equiprobable sequences of green (G) and red (R) draws-GG,' GR, What does it take for that body of observation (E) to raise the probability
RG, and RR. If the first ball sampled is green, then the probability of GG that the next emerald I examine will be green? A useful representation of
increases from 0.25 to 0.5. However, after this first observation, the when this is true is provided by the following:
probability that the second draw will be green is still 0.5, just as it was before P(NEXTGREEN/E) > P(NEXTGREEN) if and only if
any ball was drawn. I conclude that confirming a generalization and P(NEXTGREEN & E) > P(NEXTGREEN)P(E).
confirming a claim about the next instance are not always as intimately
connected as Goodman suggests. The right-hand side of this biconditional says that the cmariamc of
This example illustrates another defect in what Goodman said about the NEXTGREEN and E is positive. The covariance of A and B is defined as
grue problem. He claimed that a generalization is confirmable by its Cov(A, B) = P(A&B) - P(A)P(B). When A and B are independent of each
instances only if it is lawlike. This latter concept, whatever else it might other, Cov(A,B) = 0. Positive covariance means positive association.
mean, entails that the generalization "supports" a counterfactual. If "All So the GREEN prediction is confirmed by the evidence only if the
Xs are Ys" is lawlike, then if it is true, so is the statement "If a were an X, prediction and the evidence exhibit positive covariance. The 2 x 2 table
then a would be a Y." In the example just given, the generalization "All balls below represents this constraint on the confirmation of NEXTGREEN.
in the urn are green" is confirmed by its instances. Yet knowledge of the In Table 1, P(E) = p and P(NEXTGREEN) = q. A positive covariance
process whereby the contents of the urn were assembled guarantees means that c > 0.
that this generalization, if true, is only accidentally so. It is a mere fluke As noted earlier in connection with the urn example, it is not inevitable
if all the balls in the urn happen to be green. There is nothing about being a that past precedents should confirm a prediction. The information pre-
ball in this urn that makes something green, nor is it true that a ball would sented in that example concerning how the urn was filled made all the
not have been put into the urn unless it were green. "If this tennis ball difference. I suggest that a similar answer be given in connection with this
were in the urn, then it would be green" is as false as any counterfactual problem about NEXTGREEN. If NEXTGREEN is confirmed by E, this is
can get." because some sort of process induced a correlation (a positive covariance)
When questions of lawlikeness are considered, it makes all the difference
in the world whether the mechanism whereby the population is assembled is
known in advance or is inferred in the process of sampling. If I sample balls (4) (1 - Y)
from the urn and find that all are green, and I have no idea how the urn was NEXTGREEN NOT-NEXTGREEN
formed, the suspicion naturally arises that the homogeneous character of
my sample is not an accident. However, the fact that nomological connec-
tions are reasonably suspected in such cases does not show that lawlike-
ness is a presupposition of instance confirmation. Simply replace prior ig-
norance of process with a substantive process assumption (of the kind just
sketched for the urn problem), and the composition of a popula-
tion known to be fortuitously assembled can be confirmed by random
~arnpling.~
The urn example establishes that confirming a generalization does not Table 1
 A Baywian P n w on thc Gnrc P m b h

between past and future emeralds with respect to their color. P(NEXTGREEN/E) > P(N0T-NEXTGREEN/E)
What sort of process might this be? In this respect Goodman's example if and only if bq + c]/p > P(1 - q) - c]/p
is a bit unfortunate, since emeralds, I gather, are standardly said to be green if and only if c > p(l - 2q)/2.
by [Link]. But ignoring this wrinkle in Goodman's example, a natural
suggestion is that the relevant process assumption is that emeralds had their Here we have shifted +e problem to asking whether past observations make
color determined nr a m p . If emeralds share the same microstructure, and it more probable that the next emerald will be green or that it will be not
if microstructure determines color, then one has the basis for expecting, green. Since the probability of NOT-NEXTGREEN cannot be less than the
before even one emerald is examined, that emeralds will be alike in color. probability of NEXTGRLJE, the biconditional describes a necessary condi-
But suppose that no such process assumption is available. If we know tion for P(NEXTGREEN/E) > P(NEXTGRUE/E). And as before, if we
nothing about the process by which emeralds receive their colors (or assume that the next emerald will either be green or blue, then the condition
grulers), how are we to decide whether past precedents confirm a predic- cited is both necessary and sufficient.
tion? The answer, I think, is that wc cannot. As Hume argued, a description I noted in connection with the diachronic prediction problem that we
of the past, in and of itself, offers no guidance whatever as to what the future can't assume a prion' that c > 0. The point of interest here is that this
will be like. Here I don't mean just that we can't dcdwc what the future will assumption does not su@c for the GREEN prediction to be more probable
be like from a description that is solely about the past. The Humean point is than the G W E prediction. Here we see a difference between the synchronic
more profound: we can't even infer what the future will probably be like, and the diachronic versions of the prediction problem.
based solely on a description of the past. Nor can we say whether past
precedents confirm a prediction about the future unless we are willing to
make assumptions concerning how past and ,future are related (Sober 7. Summary
1988b). Let's take stock. Inferring the color of the next emerald and inferring what
I so far have explored what it takes for the prediction NEXTGREEN to all emeralds are like are different inference problems. Prediction and
be confirmed by the observation E. What can be said of the relationship of generalization present different issues. Likewise how much an observation
NEXTGWE to the same observation? If "grue" just meantp-cm bcfm m boosts the probability of a hypothesis is a different question from how high
n o t n c m aftm, then we could conclude that NEXTGREEN is confirmed that probability actually becomes. The diachronic and synchronic issues
just in case NEXTGRLJE is disconfirmed. But the usual definitions of need to be separated. The results obtained from applying this pair of
"grue" do not permit this simple conclusion to be drawn. It is possible for E dichotomies to the grue problem are summarized in Table 2. In each case,
to confirm both NEXTGREEN and NEXTGRUE. No first principle rules we can assert that an epistemic asymmetry obtains between the GREEN and
out the assumption that the past observations raise the probability of both the GRUE hypothesis only if we are prepared to make a substantive
predictions (even though they are incompatible with each other). assumption about the way the world is. What is more, the assumptions
Nonetheless, we have obtained a mccJsay (but not a sufficient) condi- change as we shift from problem to problem.
tion for NEXTGREEN to be confirmed more than NEXTGWE; for this to
be so,NEXTGREEN must be confirmed, which requires that c > 0. And if
we assume that the next emerald will be either green or blue, then c > 0 is 8. Concluding Comments
both necessary and sufficient for NEXTGREEN to be confirmed and
NEXTGRLJE to be disconfirmed. With these results in hand, it is worth stepping back for a moment to reflect
on the nature of the grue problem and on what kind of solution we can
hope to attain.
6. Prediction-The Synchronic Question Once again, a comparison with Hempel's raven problem is instructive.
In a paper of breathtaking brevity, Good (1967) showed that observing a
Given that the emeralds observed before now have all been green, is it more black raven (sampled at random from the world at large) can actually
probable that the next emerald will be green or that it will be p e ? If we dirconfirmthc gemralizatiwmthatatl m e n s arc black, provided o m adopts a
reformulate thii question a little, we can use the 2 x 2 table described'txfore few simple (if implausible) empirical assumptions about the inference
to identify an assumption on which this difference in probabilities depends: problem. Hempel (1968) replied that Good's argument miscontrued
 A Bayesian Primer on the Grue Probbm 237

Diachronic Synchronic philosophers believe that appeal to such "extra-empirical" virtues is part of
what it means to do science, I do not. The austere framework of
Bayesianism, and of other probabilistic epistemologies, accords no irreduci-
P(ALLGREEN/E) - P(ALLGREEN) > P(ALLGREEN/E) > P(ALLGRUE/E) ble role to simplicity, unification, non-ad hocms, and so on.1 What this
P(ALLGRUE/E) - P(ALLGRUE) if and only if means is that simplicity never provides an irreducible justification of any
if and only if P(ALLGREEN) > P(ALLGRUE). hypothesis. To be sure, it is sometimes true that the simpler theory is more
P(ALLGREEN) > P(ALLGRUE) and probable, or more likely, or more strongly confirmed by a body of data;
P(E) < 1. however, the simpler theory never has these properties becaust it is simpler
(Sober 1988b, 1990). In the present context, it does no harm to admit that
the GREEN hypothesis is "simpler" than the GRUE hypothesis. But if this
difference is to count as epistemically relevant, it will be necessary to appeal
to empirical matters of fact of the sort described in the above table. Once
If the next emerald will be If the next emerald will be these empirical assumptions are made explicit, any further mention of
either green or blue, then either green o r blue, then
P(NEXTGREEN/E) - P(NEXTGREEN) >
simplicity will be quite unnecessary.
P(NEXTGREEN/E) > P(NEXTGRUE/E)
P(NEXTGRUE/E) - P(NEXTGRUE)
For me, the fundamental lesson of the grue problem is that empirical
if and only if
if and only if c > 0. c > p(1-2q)/2.
assumptions that go beyond the content of past observations are needed to
establish an epistemic asymmetry between GREEN and GRUE. Whereas
philosophers often formulate this point by appealing to the need for
"auxiliary assumptions," scientists of a statistical bent often stress the
Table 2 importance of specifying a "model" of the relation of data to the various
hypotheses under test. Without assumptions of this sort, the data cannot be
interpreted. The slogan for scientists is: NO MODEL, NO INFERENCE.
This entirely familiar point from the practice of science should not be
the problem that he, Hempel, had wanted to pose. Hempel (1965) forgotten when we investigate the theory of that practice.
had indicated that he was interested in exploring the relationship of obser- If this is the right lesson to draw from the grue problem, we can reach an
vation to hypothesis within a "theoretically barren" background context. assessment of the solution to the problem that Goodman proposed-his
Assuming nothing at all about the world, the question is whether black theory of entrenchment. A predicate becomes entrenched when people use
ravens and red shoes both confirm the generalization that all ravens are it to formulate predictions and generalizations. It has always been a mystery
black. to me why the fact that people use a predicate should have any epistemic
Good took the view, and so do I, that almost nothing can be said about relevance. Why should our use of a predicate be evidence that this or that
confirmation in a background context of this sort. In all four of the hypothesis is true? This naive question is sometimes answered with the
problems surveyed above, an epistemic asymmetry between the GREEN response that the "new" riddle of induction involves describing our
hypothesis and the GRUE hypothesis is possible. But notice what the inductive practices, not trying to justify them. I have my doubts about this
asymmetries in these cases depend upon: for GREEN to be more probable descriptive claim as well. Is it really so obvious that human inference makers
than GRUE, or for GREEN to receive a greater boost in probability than think a hypothesis with unentrenched predicates is less plausible than a
GRUE does, m p P r k a fassumptions must be madc thatgo beyond the testimony hypothesis with entrenched predicates, all else being equal? But this reply to
of part obsmation. It isn't reason alone (or "the scientific method") that one side, I hope it is clear why Goodman's theory is the wrong kind of
induces an asymmetry here, but substantive assumptions about the way the theory, at least if one is interested in normative questions of evidence and
world is. confirmation. To describe how well entrenched a predicate nmp is involves
I therefore think it is misleading, at best, to claim that the GREEN describingpast events only. No such description can suffice to establish an
hypothesis is preferable to the G- hypothesis on the ground that the epistemic asymmetry between a GREEN hypothesis and a G W E hypothe-
former is "simpler." This appeal to simplicity gives the impression that the sis.
simplicity of a hypothesis is a reason to think that it is true. Although many I suspect that many philosophers who may have been skeptical of the
238 EUiott Sober
A Baye,kn Primer on the Gm? Pmblcm 239
. .
details of Goodman's entrenchment theory nonetheless thought that
Only a statement that is lawlike-regardless of its truth or falsity or its scientific
Goodman was at least looking for the right &ind of asymmetry; they
importance-is capable of receiving confirmation from an instance of it;
implicitly assumed that a solution to the problem could be found in some
accidental statements are not.
fact established solely by past experience and a priori reasoning, without
recourse to induction itself. To rest an asymmetry between GREEN and
5. Two alternatives should be mentioned here. First, there is likelihoodism,
GRUE on an assumption about the future was standardly said to "beg the
according to which hypotheses are evaluated solely in terms of the probabilities they
question." But what question are we thereby begging? It is the following
confer on observations (Edwards 1972). Second, there is the approach of Akaike
ill-formed question: Whichbclicfi kcdljcst on part cxprricnce and on a prioti,
(1973) and his school; see Forster and Sober (forthcoming) for discussion.
nasm'qg SU&C to show that thc GREEN bypothesir is morc rcasonablc than the
6. Discussion of the raven problem from a Bayesian point of view has made it
GRUE bypothesir? To me, this question resembles another: Whith types of
abundantly clear why looking at ravens and seeing whether they are black is a better
butter a n capable of cwttiqg a diamond?
strategy for testing "A1 ravens are black" than the strategy of looking at nonblack
University of Wiscm'nJMadison things and seeing if they are nonravcns (or, for that matter, the strategy of looking at
objects drawn from the whole universe and seeing if they are consistent with the
hypothesis). See, for example, Chihara (1981), Horwich (1982), Eells (1982),
Howson and Urbach (1989), and Earman (1992).
NOTES 7. Of course, ALLGREEN does not deductively imply that an object sampled
1. I thank Ellery Eells, Malcolm Forster, and Douglas Stalker for helpful at random from the whole universe will be both an emerald and green. Fkcall that I
comments on an earlier draft of this paper. I also am grateful to the editors of the am assuming that the sampling takes place within a restricted universe; it is part of
Philosophical R W for granting me permission to' reprint here a few paragraphs the sampling problem that the population of objects from which the sample is drawn
from my article (Sober 1988a). is composed entirely of emeralds. The relevant fact here is that for each object a
2. Since the year 2000 is fast approaching, I will define "grue" so that the grue sampled before now from this population, P(a is green / ALLGREEN & a is an
problem retains its timeless immediacy: An object is said to be grue at a given time emerald) = P(a is green / A L U R U E & a is an emerald) = 1.
precisely when it is either green and the time is before now, or it is blue and the time 8. I discuss what other philosophers have said about Goodman's proposed
is not before now. As has become somewhat customary, I delete the concept of connection between confirmation and lawlikeness in Sober (1988a).
"being examined" from Goodman's original definition. 9. Many sciences (e.g., population biology) test generalizations by drawing
3. Vocabulary differences between the two hypotheses cannot make any samples from populations that are known to be fortuitously assembled. This
epistemic difference, if a pincipk of lo&al quCalcncc is correct. As Goodman undertaking would be impossible if only lawlike statements could be confirmed by
pointed out in his exchange with Camap, once "bleen" is defined in tandem with their instances. It is worth remembering that Goodman's proposal was advanced
grue (bleen = blue before or green after), the GRUE hypotheses can be reformulated during a period in which philosophers often equated science with physics and
in familiar vocabulary and the GREEN hypotheses can be expressed by using the physics with the search for physical laws.
made-up words. 10. Of course, one might take this fact to be a rcdaurio of such probabilistic
4. Here is a passage from Goodman (1965,73) in which these three theses are epistemologies; one might argue that they are mistaken precisely because they fail to
asserted: accord an irreducible importance to these "epistemic virtues." An assessment of this
suggestion can be developed only by attending closely to the dialectics of theory
That a given piece of copper conducts electricity increases the credibility of choice in a variety of scientific controversies.
statements asserting that other pieces of copper conduct electricity, and thus
confirms the hypothesis that all copper conducts electricity. But the fact that a
given man now in this room is a third son does not increase the credibility of REFERENCES
statements asserting that other men now in this room arc third sons, and so docs
not confirm the hypothesis that all men now in this room are third sons . . . The Akaike, H. (1973). "Information theory and an extension of the maximum
difference is that in the former case the hypothesis is a k I r % statement; while in likelihood principle." In B. Petrov and F. Csaki, eds., Sccond Intrrnational
the latter case, the hypothesis is a merely contingent or accidental generality. Sympn'urn on Infmmarion T k q Jh d e m i a i Kiado, 267-81.
Chihara, C. (1981). "Quine and the confirmational paradoxes." In P. French, H.
 Wcttstein, and T. Uehling, eds., Midwest Studies in PMaroply, vol. 6, Univcrsity
Bavesian Projectibility
J

of Minnesota Press,425-452.
Earman, J. (1992). Buycs or Bwt. MIT Press. Brian S W
Edwards, A. (1972). LiCclibooA Cambridge University Pms.
Eells, E. (1982). Rational l k k b n and C a w d t y . Cambridge University Pms. Undoubtedly we do makc pndiEti0n.s by pojcctiw the
Forster, [Link] Sober, E. (forthcoming). "How to tell when simpler, morc unified, pattctw of the past into the fuFun, but in sclcctin~the
or less ad hoc theories will provide more accurate predictions." pattctw we pc'cct fmm a- all those that the past
Good, I. (1967). "The white shoe is a red herring." [Link] o u d fm the Philosoply exhibits, we use practical criteria that so jiw seem to have
of S c k z 17, 322. ucapcd ducomy and fmmuhim.
Goodman, N. (1965). b t , Fictia and Furccast. Bobbs-Merrill. Nelson Goodman, "A Query on
Hempcl, C. (1965). "Studies in the logic of confirmation." In Aspcca of Scicntafic Confirmation" (1946)
Eaplnnntion and Other Essays in the Phi[osoply of Scinrcc. Free Press.
Hempcl, C. (1968). "The white shoe-no red herring." Britirh J o u d fm the
Philosoply of Scicntt 18, 239-40.
Horwich, P. (1982). Pmbabil* and EPidcncc. Cambridge University Press.
Howson, C. and Urbach, P. (1989). Scimhjic Rcasm'n~:The Bay& Approach. 1. Introduction
Open Court. In 1946 Nelson Goodman raised the problem of the projectibility of
Sober, E. (1988a). "Confirmation and law-likeness." Philosophical Rtpitrp 97, hypotheses in a note addressing the confirmation theory of Rudolf Carnap.
93-98.
H e later gave a sensational illustration: the hypothesis "All emeralds are
Sober, E. (1988b). Rccarnrc~tiwthe Past: Parsitmy, EvolutionJ and Inference. MIT grue," which came to be discussed as the Goodman Paradox. The predicate
Press.
"grue" was so manifestly pathological many were led into thinking that the
Sober, E. (1990). ''Let's razor Ockham's razor." In D. Knowles, ed., ExpfuMcMn problem was to find some criterion which would exclude similar pathology
and Its Limits. Cambridge University Press. 73-94.
from inductive reasoning.
Goodman sees clearly that the problem of projectibility is much more
general, and that judgments of projectibility must be central to any adequate
theory of confirmation. He believes that a theory of projectibility should be
a pragmatic theory, and sketches the beginnings of such a theory in the last
chapter of F a t , Fiction and Emcast.
I believe that the broad outlines of Goodman's approach to a theory of
projectibility are just right. The theory will not attempt t o tell one in a
vacuum which predicates are projectible. Rather it will explain how present
judgments of projectibility should be based on past judgments of
projectibility together with the results of past projections. The circularity
involved in such a theory is not to be viewed as vicious. The theory can still
inform our understanding of projectibility. And a theory with this sort of
"virtuous circularity" is really the best that can be expected. However, the
implementation of Goodman's program for a theory of projectibility does
not seem very far advanced. There is, however, a preexisting pragmatic
framework which offers precise tools for addressing the question: the theory
of personal probability. How is projectibility represented within this
framework? This question leads straight to central concepts of Bayesian
statistics.