I. Douven - 2013 - Inference To The Best Explanation, Dutch Books, and Inaccuracy Minimisation

INFERENCE TO THE BEST EXPLANATION, DUTCH
BOOKS, AND INACCURACY MINIMISATION

BY IGOR DOUVEN
Bayesians have traditionally taken a dim view of the Inference to the Best Explanation (IBE),
arguing that, if IBE is at variance with Bayes rule, then it runs afoul of the dynamic Dutch
book argument. More recently, Bayes rule has been claimed to be superior on grounds of condu-
civeness to our epistemic goal. The present paper aims to show that neither of these arguments
succeeds in undermining IBE.
The inference rule called Inference to the Best Explanation (IBE) assigns
confirmation-theoretic import to explanatory considerations. According to
some, IBE is the cornerstone of scientific methodology.
1
But critics have
argued that, if IBE is at variance with Bayes rule, then, like any other
such rule, it is to be rejected as leading to irrational belief updates.
2
For
many years, the standard argument for this claim has been the so-called
dynamic Dutch book argument, which purports to show that updating by
any rule other than Bayes makes one liable to sure financial losses. How-
ever, even Bayesians themselves have increasingly come to regard this ar-
gument as addressing the wrong issue, to wit, that of whether it is
rational from a practical, rather than an epistemic, viewpoint to deviate
from Bayes rule. This has led some theorists to pursue a different strategy
1
See, for example, R. Boyd, The Current Status of Scientific Realism in J. Leplin
(ed.), Scientific Realism (University of California Press, 1984), pp. 4182, and E. McMullin,
The Inference that Makes Science (Marquette University Press, 1992).
2
We will be throughout concerned with update rules applicable to learning events in
which an agent becomes certain of a proposition of which he or she was previously uncer-
tain. Bayesians acknowledge that other types of learning event call for different update
rules. See, for example, R. Jeffrey, The Logic of Decision (University of Chicago Press, 2nd
ed., 1983), Ch. 11, and B. C. van Fraassen, Laws and Symmetry (Oxford UP, 1989), Ch. 13.
The Philosophical Quarterly Vol. 63, No. 252 July 2013
ISSN 0031-8094 doi: 10.1111/1467-9213.12032
2013 The Author The Philosophical Quarterly 2013 The Editors of The Philosophical Quarterly
Published by John Wiley & Sons Ltd, 9600 Garsington Road, Oxford ox4 2DQ, UK, and 350 Main Street, Malden, MA 02148, USA
in defence of Bayes rule, a strategy that is purported to offer a distinc-
tively epistemic argument in favour of that rule, one spelled out in terms
of inaccuracy minimisation. Roughly, the argument is that by updating
via any non-Bayesian rule, ones degrees of belief are not as accurate as
they would have been had one updated via Bayes rule.
This paper aims to show that neither of the aforesaid arguments succeeds
in undermining IBE. The first part of the paper argues that, while current
developments in mainstream epistemology may help to deflect some of the
criticism the dynamic Dutch book argument has met with, the argument
fails nonetheless, since it rests on an unfounded (and unstated) premise. The
second part focuses on the inaccuracy-minimisation defence of Bayes rule,
arguing that there appear to be several equally legitimate ways to interpret
the notion of inaccuracy minimisation, and using computer simulations to
show that under some of them it may be IBE rather than Bayes rule that
does best with regard to inaccuracy minimisation.
I. THE DYNAMIC DUTCH BOOK ARGUMENT REVISITED
For many decades, the Ramseyde Finetti Dutch book argument has
been viewed as key to the Bayesian account of rationality. According to
this argument, we are susceptible to Dutch bookscollections of bets
ensuring a negative net pay-off come what mayprecisely if our
degrees of belief violate the axioms of probability. From this, Ramsey
and de Finetti concluded that rational degrees of belief are formally
probabilities.
Ian Hacking may have been the first to observe that the Ramseyde
Finetti argument in fact does nothing to justify Bayes rule.
3
However, a
few years after the publication of Hackings paper, Paul Teller reported a
Dutch book argumentwhich he attributed to David Lewisaimed at
justifying Bayes rule as the only rational update rule.
4
This dynamic
Dutch book argument (as it is now called) purports to show that if a per-
son updates by some rule other than Bayes, she can be offered a series of
bets at different points in time such that each bet will seem fair at the
time it is offered, yet jointly the bets guarantee a financial loss. What is
worsethe argument continuesthe person could have seen this loss
3
I. Hacking, Slightly More Realistic Personal Probabilities, Philosophy of Science, 34,
1967, pp. 31125.
4
P. Teller, Conditionalization and Observation, Synthese, 26, 1973, pp. 21858. The
source of the argument reported by Teller was later published as D. Lewis, Why Condi-
tionalize? in his Papers on Metaphysics and Epistemology (Cambridge UP, 1999), pp. 4037.
INFERENCE TO THE BEST EXPLANATION 429
coming. This vulnerability to dynamic Dutch books has convinced many
that non-Bayesian updating is a mark of irrationality.
5
Recently, however, the Dutch book approach to defending Bayesianism
has come under a cloud. Critics have argued that when we are concerned with
the rationality of degrees of belief as well as the change thereof over time,
we are concerned with questions of epistemic rather than practical rationality.
Given that being vulnerable to cunning bookies seems primarily a practical
liability, the Dutch book arguments have been said to be beside the point.
6
In response to this, some (e.g., Brian Skyrms
7
) have claimed that Dutch
book vulnerability does flag an underlying epistemic defect: it is a mani-
festation of the fact that a person deems one and the same bet or series
of bets as both fair and not fair and thus is in an inconsistent state of
mind. Even if this claim is true for the (static) Ramseyde Finetti argu-
ment,
8
the point does not carry over to the dynamic Dutch book argu-
ment. An agent may be susceptible to engage in the kind of betting over
time that figures in that argument without at any one time holding incon-
sistent views on the fairness of any bets. Naturally, after a learning event
she may regard a bet as unfair that previously she regarded as fair, but
the same would have been true had she been a Bayesian learner.
Still, Bayesians may not be altogether defenceless against the above cri-
tique. In particular, they may be able to get some mileage out of the prag-
matic turn that a number of epistemologists are currently taking. The
epistemic status of a belief has traditionally been thought to depend solely
on matters that bear on the truth of the belief, like the quality of ones evi-
dence or whether or not one is reliably connected to what the belief is
about. But over the past years, various authors have argued that the episte-
mic status of a belief is inextricably bound up with the believers practical sit-
uation, in particular, with what is at stake for her in believing correctly.
9
Bayesians wishing to maintain the integrity of the Dutch book defence may
not want to buy into any particular one of the arguments that have been
5
We should actually speak of putative vulnerability to dynamic Dutch books: that a
non-Bayesian updater is bound to regard all bets in a dynamic Dutch book as fair has been
disputed in I. Douven, Inference to the Best Explanation Made Coherent, Philosophy of
Science, 66, 1999, pp. S42435; see also M. Tregear, Utilising Explanatory Factors in Induc-
tion?, British Journal for the Philosophy of Science, 55, 2004, pp. 50519.
6
See J. Joyce, A Nonpragmatic Vindication of Probabilism, Philosophy of Science, 65,
1998, pp. 575603, Sect. 2, and references given there.
7
B. Skyrms, Coherence in N. Rescher (ed.), Scientific Inquiry in Philosophical Perspective
(University Press of America, 1987), pp. 22541.
8
But see Joyce, A Nonpragmatic Vindication of Probabilism, p. 585 f, for a critique.
9
See, for example, J. Fantl and M. McGrath, Evidence, Pragmatics, and Justification,
Philosophical Review, 111, 2002, pp. 6794, and J. Stanley, Knowledge and Practical Interests
(Oxford UP, 2005).
430 IGOR DOUVEN
advanced in favour of this pragmatic encroachment view (as it has been
called). However, it may suffice for them to argue that the mere prominence
in current epistemology of the debate on pragmatic encroachment is enough
to call into question the existence of the clear-cut divide between the episte-
mic and the pragmatic that the critics of the Dutch book arguments are pre-
supposing. Forit may be saidif there were such a clear-cut divide,
contributions to this debate should have gone down like lead balloons. And
Bayesians may concludeif there is no such divide, then little is left of the
charge that Dutch book arguments address the wrong type of rationality.
Be this as it may, there is a deeper problem with the dynamic Dutch
book argument, one that remains even if the pragmatic encroachment
view is endorsed. For note that Dutch book invulnerability is only one
among an in principle indefinite number of practical interests that people
may have. What if IBE, or some other non-Bayesian rule, serves other
practical interests better than Bayes rule? Could the Bayesian in that case
still maintain that, in view of the dynamic Dutch book argument, Bayes
rule is the only rational update rule? Surely there exist practical goals
whose achievement would more than make up for running the risk of
being fleeced by a Dutch bookieespecially in view of the fact that
Dutch bookies only occur as fictional characters in philosophers tales!
To show that IBE may indeed have compensating practical advantages,
we use a particular probabilistic version of IBE that we apply in the context
of a simple statistical model. Let fH
i
g
iOn
be a set of self-consistent, mutually
exclusive, and jointly exhaustive hypotheses, and Pr ones probability function
prior to learning E. Then, according to the version of IBE to be considered,
ones new probability for H
i
after learning E (and nothing stronger) equals
PrH
i
PrEjH
i
f H
i
; E
P
n
j 1
PrH
j
PrEjH
j
f H
j
; E

;
where f is a function that assigns a bonus point (
>
0) to the hypothesis (or
hypotheses, in case of a tie) that explain(s) E best in light of the back-
ground knowledge, and assigns nothing to the other hypotheses.
10
For present concerns, all we need is a definition of best explanation for
the following statistical model. Let fH
i
g
0OiO10
be a set of bias hypotheses
concerning a given coin, where H
i
is the hypothesis that the bias for heads
10
A number of authors, including Peter Lipton and Jonathan Weisberg, have proposed
versions of IBE that are compatible with Bayesianism; see P. Lipton, Inference to the Best
Explanation (Routledge, 2nd ed., 1004), Ch. 7, and J. Weisberg, Locating IBE in the Bayes-
ian Framework, Synthese, 167, 2009, pp. 12543. Note that the present version is equivalent
to Bayes rule if, and only if, there is no best explanation among the hypotheses and
f assigns 0 to all of them.
is i/10. These hypotheses are supposed to be jointly exhaustive. Let E
j
indi-
cate that the outcome of the j-th of a series of tosses with the given coin is E.
Then we say that H
i
best explains E
j
iff i/10 is closer to the actual frequency
of heads in the first j tosses than k/10, for all k {0,,10} different from i.
For definiteness, in this case let f H
i
; E
j
:1. If, for some i and k, i/10 and
k/10 are equally close to the actual frequency of heads, and closer than l/10
for all l different from i and k, the bonus is split: f H
i
; E
j

f H
k
; E
j
:05. No other hypotheses receive bonuses. So, for instance, if
681 heads have been observed in a series of 999 tosses and the 1000th toss is
again heads, then in this model, and absent any further information about
the coin, H
7
provides the best explanation of that last outcome and thus
receives a bonus if updating proceeds by the above rule.
We will continue to use IBE as a label for the broad idea that explana-
tory considerations have confirmation-theoretic import and use IBE (in sans
serif font) to designate our probabilistic explication of that idea. It is to be
emphasised that nothing in the following will hinge on whether IBE is the
best or even a satisfactory explication of IBE. The main role of IBE will be in
bringing into relief hidden, problematic premises in the dynamic Dutch
book and inaccuracy-minimisation arguments. If we have failed to capture
the notion of best explanation even for the above simple model, that will
not make those arguments valid without the hidden premises, nor will it
make these premises appear unproblematic.
Now, suppose the Bayesian and the explanationist (as we shall henceforth
call the IBE-updater) are watching the same sequence of coin tosses to which
the above model pertains. Both have started with a flat probability distribu-
tion over the eleven bias hypotheses, and they update their probabilities for
these hypotheses via their respective update rules. It has been said that a
proposition is assertable to the extent that it has high subjective probability
for its assertor.
11
If so, we may ask whothe Bayesian or the explanationist
11
F. Jackson, On Assertion and Indicative Conditionals, Philosophical Review, 88, 1979,
pp. 56589, at p. 565. It has been argued that a proposition is assertable if it is rationally
acceptable; see, for example, I. Douven, Assertion, Knowledge, and Rational Credibility,
Philosophical Review, 115, 2006, pp. 44985 and Assertion, Moore, and Bayes, Philosophical
Studies, 144, 2009, pp. 36175. This yields Jacksons view cited in the text if high subjective
probability suffices for rational acceptability. That might seem too simplistic, in particular
in view of so-called lottery propositions (propositions to the effect that a given ticket in a
large fair lottery with only one winner will lose), for, although highly probable, such propo-
sitions do not at all seem rationally acceptable. But see I. Douven, The Lottery Paradox
and the Pragmatics of Belief, Dialectica, 66, 2012, pp. 35173, and Putting the Pragmatics
of Belief to Work in A. Capone, F. Lo Piparo, and M. Carapezza (eds), Perspectives on Prag-
matics and Philosophy (Springer, in press), for a proposal on which high probability is enough
for rational acceptability, with the exception of lottery propositions, which are excluded on
principled grounds.
432 IGOR DOUVEN
is more likely to first be in a position to assert the truth about the bias of
the coin, supposing the sequence is long enough for both eventually to be in
that position.
To answer this question, we ran computer simulations of sequences of
coin tosses that were long enough for both the Bayesian and the explana-
tionist to assign a probability greater than .99 to the true bias hypotheses
(following Richard Foleys suggestion that .99 is a reasonable threshold
for high subjective probability
12
). More exactly, we ran 1000 sequences
of (simulated) tosses of a coin with bias .1 for heads, 1000 sequences of
tosses of a coin with bias .2 for heads, and so on, up to bias .9 (the results
for bias 0 and bias 1 were determined analytically). It was further simu-
lated that, after each toss, the Bayesian and the explanationist updated
their probabilities for the various bias hypotheses, and it was registered
who was first in assigning a probability greater than .99 to the true
hypothesis and also how many more updates the other party needed to
assign a probability greater than .99 to the truth as well. Table 1 summa-
rises the results of these simulations. As can be seen in this table, the ex-
planationist was, on average, much faster to come to a position to assert
the truth about the bias of the coin than the Bayesian.
If this sounds unremarkable, imagine that the hypotheses concern some
scientifically interesting quantitysuch as the success rate of a medical treat-
Bias X SD n<100 100n<0 n = 0 0<n100 n>100
.0 24 0 0 0 0 1000 0
.1 64 66 1 86 0 709 204
.2 128 83 0 6 0 384 610
.3 180 101 0 0 0 184 816
.4 211 115 0 0 0 99 901
.5 220 117 0 0 0 111 889
.6 212 111 0 0 0 106 894
.7 172 101 0 0 0 195 805
.8 126 86 0 5 3 393 599
.9 65 64 4 69 0 716 211
1.0 24 0 0 0 0 1000 0
Table 1. Results of 1000 simulations of sequences of coin tosses for each bias value: X gives the average
over 1000 simulations of the number n of updates by which the explanationist was faster than the Bayesian
in assigning a probability above .99 to the correct bias hypothesis; SD gives the standard deviation of the
sample; the other numbers indicate in how many of the simulations the explanationist was more than 100,
between 0 and 100, etc., updates faster than the Bayesian.
12
R. Foley, The Epistemology of Belief and the Epistemology of Degrees of Belief,
American Philosophical Quarterly, 29, 1992, pp. 11124, at p. 113.
ment, or the probability of depressive relapserather than the bias of a coin,
and the tosses are observations or experiments aimed at determining that
quantity. Which researcher would not want to use an update rule that
increases her chances of being in a position to make public a scientific theory,
or a new medical treatment, before the (Bayesian) competition is?
However, there is a possible downside to the apparent success of IBE,
described above. The reason why the explanationist so often beats the
Bayesian in these simulations in assigning a high probability to the truth
is that she is, in a clear sense, bolder in her responses to new information,
due to the fact that she adds a bonus to the best explanation. It does not
take much to see that this same feature makes the explanationist more
prone to assign a high probability to some false hypothesis: a row of con-
secutive tosses producing a subsequence in which the relative frequency
of heads starkly deviates from the probability for heads is more likely to
push the explanationists probability for some false bias hypothesis over
the .99 threshold than it is to push the more cautious Bayesians probabil-
ity for some false hypothesis over that threshold. Thus, even if IBE may
put one more rapidly in a position to assert truths, it may also put one
more rapidly in a position that licenses the assertion of falsehoods.
Consider, then, an arguably fairer comparison between IBE and Bayes
rule. A magician has a set of eleven coins in his pocket, one coin with a
perfect bias for heads, one with a bias of .1 for heads, one with a bias of
.2 for heads, and so on. He entices a Bayesian and an explanationist into
playing a game against one another. In this game, the magician picks one
of the coins and starts tossing it, showing each outcome to the two play-
ers. A player scores a point by raising her hand and asserting the truth
about the bias of the coin. However, if the player gets the bias wrong, the
point goes to the other player. If both the players raise their hands after
the same update and both are right about the bias, they both receive a
point; if they are both wrong, neither receives a point; if one is right and
the other wrong, the former receives two points. Once a player has
asserted a bias hypothesiswhether correctly or incorrectlythe magi-
cian puts the coin back into his pocket and picks a (not necessarily differ-
ent) coin. Then everything starts all over. This procedure is repeated 100
times, after which the player with the highest score is declared the win-
ner. Note that this game seems to treat the possible advantages and disad-
vantages of IBE-updating in an evenhanded manner: because of her
boldness, the explanationist might often be first to identify the truth, but
by the same token she might earn quite a few points for her opponent.
Who is more likely to win the game, the explanationist or the Bayesian?
Or are their chances of winning equal?
434 IGOR DOUVEN
To answer these questions, we ran 1000 simulations of the game. In
these simulations, the explanationist always won, and typically did so by a
wide margin. To be more exact, the mean of the differences between the
explanationists score and the Bayesians score was 27 (SD = 9); the small-
est difference between the two players score at the end of a game was 5,
and the greatest was 52.
This is not meant to demonstrate the superiority of IBE. Sets of trick
coins of the variety required for playing the above game are hardly more
common than Dutch bookies. Moreover, nothing said so far, or to be said
hereafter, excludes the possibility that there are circumstances under
which Bayesian updating offers practical advantages more consequential
than Dutch book invulnerability. Instead, the point is to draw attention to
a hidden assumption in the Dutch book approach, to wit, that Dutch
book invulnerability trumps all other practical goals we may have. Absent
an argument for this assumption, the Dutch book approach does not pose
a threat to IBE or any other probabilistic explication of IBE, even granting
that the pragmatic pervades the epistemic.
II. IBE AND OUR EPISTEMIC GOAL
The relation between the pragmatic and the epistemic is immaterial to
an approach to justifying Bayes rule, and more generally the tenets of
Bayesianism, that has emerged over the past fifteen years or so. Various
theorists, disconcerted by the apparently pragmatic focus of the Dutch
book arguments, have sought to justify the Bayesian tenets in strictly epi-
stemic terms. Specifically, they have sought to show that those tenets are
most conducive to the achievement of our epistemic goal, as spelled out
for graded beliefs.
This development owes much to Jim Joyces A Nonpragmatic Vindica-
tion of Probabilism, in which a first attempt is made to formulate an epi-
stemic goal in terms of graded beliefs, and which argues that, for every
degrees-of-belief function that violates the probability axioms, there is a
degrees-of-belief function that obeys those axioms and that is closer to the
epistemic goal. Joyce thought that this vindicated the synchronic part of
Bayesianism, according to which degrees of belief ought to be probabili-
ties, though in a later paper he admits that the earlier argument was not
airtight.
13
For present purposes, however, it is mostly Joyces conception
of our epistemic goal as pertaining to graded beliefs that matters.
13
J. Joyce, Accuracy and Coherence: Prospects for an Alethic Epistemology of Partial
Belief, in F. Huber and C. Schimdt-Petri (eds), Degrees of Belief (Springer, 2009), pp. 26397.
Among mainstream epistemologists, who tend to be concerned first
and foremost with categorical beliefs, it is almost universally held that our
epistemic goal is to believe all that is true and only what is true. Until
recently, there was no equally clear conception of an epistemic goal in
terms of graded beliefs. This changed when Joyce proposed an epistemic
goal for graded beliefs that was explicitly meant to be analogous to the
aforementioned epistemic goal for categorical beliefs.
14
The proposal is
that our system of degrees of belief ought to be gradationally accurate,
and in fact as gradationally accurate as any other system of degrees of
belief that we might adopt. The notion of gradational accuracy is defined
in technical terms, in particular in terms of so-called scoring rules (about
which more below). But the basic intuition underlying it is clear enough,
to wit, that the higher ones degree of belief in a true proposition is, the
more accurate one is, ceteris paribus, and also the lower ones degree of
belief in a false proposition is, the more accurate one is, ceteris paribus.
In the same vein, Hannes Leitgeb and Richard Pettigrew have tried to
give a nonpragmatic justification not only of the synchronic part of Baye-
sianism but also of its diachronic part, that is, of Bayes rule.
15
They share
Joyces view of our epistemic goal, which they put as follows:
16
Accuracy: An epistemic agent ought to approximate the truth. In other words, she
ought to minimise her inaccuracy.
Also like Joyce, Leitgeb and Pettigrew make the notion of inaccuracy pre-
cise by reference to scoring rules. What they then argue is that Bayesian
updating minimises expected inaccuracy. More exactly, they argue that if,
and only if, an agent updates via Bayes rule, she minimises the expected
inaccuracy of her post-update probability function, where the expectation
is minimal according to her pre-update probability function.
17
What Leitgeb and Pettigrew aim to show for Bayes rule is in an
important respect disanalogous with what Joyce aims to show for the syn-
chronic part of Bayesianism. While Joyce argues that an agent whose
degrees of belief are not probabilities fails to minimise actual inaccuracy,
Leitgeb and Pettigrew argue that an agent who updates by a rule other
than Bayes fails to minimise expected inaccuracy. Nothing they say pre-
cludes the possibility that some non-Bayesian rule outperforms Bayes rule
14
Joyce, A Nonpragmatic Vindication of Probabilism, p. 578 f.
15
H. Leitgeb and R. Pettigrew, An Objective Justification of Bayesianism I: Measuring
Inaccuracy, Philosophy of Science, 77, 2010, pp. 20135, and An Objective Justification of
Bayesianism II: The Consequences of Minimizing Inaccuracy, Philosophy of Science, 77, 2010,
pp. 23672.
16
Leitgeb and Pettigrew, An Objective Justification of Bayesianism I, ibid., p. 202.
17
Leitgeb and Pettigrew, An Objective Justification of Bayesianism II, ibid., p. 249 f.
436 IGOR DOUVEN
with respect to actual inaccuracy minimisation. More importantly, Leitgeb
and Pettigrew are concerned only with the inaccuracy of the immediate
post-update belief state. But an equallyif not morelegitimate question
one may ask about update rules is which of them will lead to the most
accurate belief state in the long run. Indeed, given that the epistemic goal
for graded beliefs is supposed to be analogous to the epistemic goal as dis-
cussed by mainstream epistemologists, the long-run question is the more
natural one to ask. After all, in mainstream epistemology the epistemic
goal is mostly conceived as an ultimate goal to which all our epistemic
endeavours are geared and in light of which they are to be assessed.
18
This points to what seems to be a major problem for the inaccuracy-
minimisation approach, to wit, that Accuracy permits of a number of differ-
ent interpretations. For instance, it can be interpreted as demanding that
every single update minimise expected inaccuracy, as Leitgeb and Petti-
grew do, or that every update minimise actual inaccuracy, or that every
update be aimed at realising the long-term project of coming to have a
minimally inaccurate representation of the world, even if individual
updates do not always minimise inaccuracy or expected inaccuracy. What
makes this problematic is that all of these sound like legitimate epistemic
goals, and that it is by no means obvious that if rule R is most conducive
to the realisation of one goal and rule R is most conducive to the realisa-
tion of a second, then R = R, or at least the rules are equivalent in that
they always yield the same output, given the same input.
In fact, there is not just the question of whether the goal of inaccuracy
minimisation is meant as pertaining (only) to the next update or is rather
meant as a long-term project. Even if we assume that it is conceived as a
long-term project, there are still further distinctions that can be made.
One notable distinction concerns the question of whether we should aim
to have a minimally inaccurate probability function in the long run, how-
ever far in the future that may be, or whether it is better to have a mod-
erately accurate probability function in the shorter run. It is certainly
imaginable that, for purely epistemic reasons, one might opt for the latter,
18
See, for example, A. Latus, Our Epistemic Goal Australasian Journal of Philosophy, 78,
1998, pp. 2839, at p. 29, who says that in asking what our epistemic goal is we are inter-
ested in what we want the overall result of our various ways of finding outto be. Alvin
Goldman also argues that (what he calls) epistemic systems are to be evaluated in terms of
their accuracy, which he phrases in terms of degrees of truth-possession, which in turn is
defined by reference to a scoring rule; see his Epistemic Relativism and Reasonable Dis-
agreement, in R. Feldman and T. Warfield (eds), Disagreement (Oxford UP, 2010), pp. 187
215, Sect. 2. However, he is very clear that what he has in mind is long-term accuracy:
Epistemic system E is better than epistemic system E
iff conformity to E would produce

(in the long run) a higher total amount of degrees of truth-possession than conformity to
E
would produce (p. 194).

supposing there is a choice to be made. We are, after all, curious about
the truth,
19
and curiosity typically comes with a sense of urgency. As a
result, we might prefer an update rule that is more likely to take us fairly
close to the truth in a reasonably short time span over one that is more
likely to take us extremely close to the truth in the long run but less likely
to take us even fairly close to the truth in (say) the middle-long run.
20
We could go on for some time disambiguating Accuracy in this way, if
only because speed and accuracy of convergence to the truth can be
traded off in an indefinite number of ways. However, our aim here is not
to catalogue all reasonable precisifications of Accuracy but rather to show
that the inaccuracy-minimisation argument poses no real threat to IBE.
As Leitgeb and Pettigrew argue, Bayesian updating minimises the
expected inaccuracy of our next belief state relative to our current belief
state, and thereby may be said to be most conducive to one epistemic goal,
but that does little to undermine IBE as long as it is understood that IBE
(or a particular explication of it) is most conducive to some other episte-
mic goal or goals, where possibly the different epistemic goals can both
(or all) be regarded as explications of the same broad idea that we should
strive for a minimally inaccurate belief system.
At this juncture, Bayesians might try to argue that expected inaccuracy
minimisation of our next belief state as judged from our present one
trumps any other epistemic goal we may have. The prospects of this strat-
egy are about as bleak as the prospects of showing that Dutch book invul-
nerability trumps any other practical goal that we may have; bleaker still,
for it would seem absurd to claim that it is epistemically more important
to have an update rule that minimises expected inaccuracy than to have
one that actually minimises inaccuracy.
21
Alternatively, Bayesians might try to show that Bayes rule does best
with respect to all epistemic goals. This may be hard to establish in an
19
See, for example, C. Hempel, Aspects of Scientific Explanation, in his Aspects of Sci-
entific Explanation and Other Essays in the Philosophy of Science (Free Press, 1965), pp. 331496, at
p. 333, and R. Foley, The Theory of Epistemic Rationality (Harvard UP, 1987), at p. 11.
20
I. Douven, Simulating Peer Disagreements, Studies in History and Philosophy of Science,
41, 2010, pp. 14857, Sect. 4, makes a parallel point in terms of the qualitative notion of
truth approximation for update procedures that take into account peer opinions. (See G.
Oddie, Truthlikeness, The Stanford Encyclopedia of Philosophy (Fall 2008 Edition), available at
http://plato.stanford.edu/entries/truthlikeness/, for a useful overview of the literature on truth
approximation.)
21
This seems especially true in view of Robbie Williams observation that expected
inaccuracy minimisation defences of Bayes rule raise the question of why [you should]
trust an outdated belief state to tell you how to fix your beliefs now you have new informa-
tion; J. R. G. Williams, Generalized Probabilism: Dutch Books and Accuracy Domina-
tion, Journal of Philosophical Logic, 41, 2012, pp. 81140, at p. 835.
438 IGOR DOUVEN
a priori manner, supposing that among the relevant goals are ones whose
realisation requires actual inaccuracy minimisation. How well a rule does
in terms of truth approximation may plausibly depend not only on the
rule but also on which world is the actual one. For example, even if ex-
planationists have so far failed to produce a reason for believing that best
explanations tend to be true, we might happen to inhabit a world in
whichperhaps purely coincidentallybest explanations do tend to be
true. Worse yet (for the Bayesians), the simulations to be reported below,
which compare Bayes rule and IBE in terms of gradational inaccuracy in
the setting of our earlier statistical model, give reason to doubt that Ba-
yes rule is the most conducive to the realisation of every epistemic
goal.
22
In these simulations, gradational inaccuracy is measured by means of
two scoring rules. One is the popular Brier scoring rule. Given a set
fH
i
g
iOn
of self-consistent, mutually exclusive, and jointly exhaustive
hypotheses, the Brier score penalty for an agent whose probabilities are
given by Pr equals
P
n
i1
sH
i
t PrH
i

2
, where sH
i
t designates the
semantic value of H
i
.
23
The other is the log score rule, another popular
scoring rule.
24
In contrast to the Brier rule, the log score rule penalises an
agent only on the basis of her probability for the true hypothesis: where
H
i
is the true hypothesis, an agents log score penalty equals ln PrH
i
.
For each possible bias value, we ran 1000 simulations of a sequence of
1000 tosses. As previously, the explanationist and the Bayesian updated
their degrees of belief after each toss. We registered in how many of those
1000 simulations the explanationist incurred a lower penalty than the
Bayesian at various reference points, at which we calculated both Brier
penalties and log score penalties. The outcomes of these simulations are
displayed in Table 2. They show that, on either measure of inaccuracy,
IBE is most often the winnerit incurs the lowest penaltyat each refer-
ence point. Hence, at least in the present kind of context, IBE seems a bet-
ter choice than Bayes rule.
22
An anonymous referee noted that, while the simulations suffice to make my case
against the inaccuracy minimisation defence of Bayes rule, an analytical investigation of
IBE would be worthwhile in its own right. I fully agree, but this is a task better dealt with
in a separate paper.
23
Leitgeb and Pettigrew, An Objective Justification of Bayesianism I: Measuring Inac-
curacy, argue that the Brier score is the scoring rule via which to measure gradational inac-
curacy. However, their argument rests on several premises the defence of which they
relegate to future work.
24
In his The Justification of Induction, Philosophy of Science, 59, 1992, pp. 52739, Roger
Rosenkrantz also uses the log score rule in an attempt to vindicate Bayes rule that fore-
shadows Leitgeb and Pettigrews.
Or maybe notfor we also calculated averages, taken over the 1000
simulations for each bias value, of the penalties incurred by the explana-
tionist and the Bayesian at the designated reference points. Table 2 gives,
for each of these points, the mean of the IBE penalties minus the mean of
the Bayes rule penalties (in parentheses). As can be seen from those num-
bers, if there is a difference between the two means, it is always in favour
of Bayes rule. The reason for this is manifest from the spread of the sim-
ulation outcomes (not represented here), which shows that, although IBE
wins in most instances, it is typically by a relatively small margin, whereas
in some of the runs in which IBE loses, it incurs considerably greater
Bias 100 250 500 750 1000
Brier score
.0 1000 (0) 1000 (0) 1000 (0) 0 (0) 0 (0)
.1 923 (.02) 993 (0) 1000 (0) 1000 (0) 1000 (0)
.2 748 (.10) 937 (.02) 991 (0) 1000 (0) 1000 (0)
.3 717 (.11) 918 (.03) 984 (0) 998 (0) 999 (0)
.4 690 (.11) 865 (.06) 968 (.01) 990 (0) 997 (0)
.5 679 (.12) 879 (.05) 972 (0) 992 (0) 998 (0)
.6 660 (.14) 904 (.03) 977 (0) 992 (0) 995 (0)
.7 698 (.12) 907 (.03) 978 (.01) 995 (0) 1000 (0)
.8 754 (.09) 947 (.02) 996 (0) 999 (0) 999 (0)
.9 910 (.04) 990 (0) 1000 (0) 1000 (0) 1000 (0)
1.0 1000 (0) 1000 (0) 1000 (0) 0 (0) 0 (0)
log score
.0 1000 (0) 1000 (0) 0 (0) 0 (0) 0 (0)
.1 926 (.18) 993 (.01) 1000 (0) 924 (0) 515 (0)
.2 750 (.75) 937 (.47) 991 (.12) 1000 (0) 1000 (0)
.3 717 (.90) 918 (.58) 984 (.12) 998 (.05) 999 (0)
.4 691 (1.08) 865 (1.08) 968 (.66) 990 (.36) 997 (.14)
.5 679 (1.18) 879 (.94) 972 (.56) 992 (.24) 998 (.09)
.6 664 (1.20) 904 (.86) 977 (.24) 992 (.20) 995 (.21)
.7 698 (.99) 907 (.72) 978 (.28) 995 (.12) 1000 (0)
.8 756 (.68) 947 (.39) 996 (.08) 999 (.01) 999 (0)
.9 910 (.25) 990 (0) 1000 (0) 932 (0) 513 (0)
1.0 1000 (0) 1000 (0) 0 (0) 0 (0) 0 (0)
Table 2. Results of 1000 simulations of sequences of 1000 coin tosses for each bias value: the columns give
the number of simulations in which IBE incurred a lower Brier score penalty (above) or log score penalty
(below) than Bayes rule after 100, 250, 500, 750, or 1000 tosses; in parentheses is the mean of the IBE penal-
ties minus the mean of the Bayes rule penalties, taken over the 1000 simulations (rounded to two decimal
places).
440 IGOR DOUVEN
penalties than does Bayes rule. This, in turn, is due to the fact, noted
earlier, that the explanationist reacts in a bolder fashion to the evidence
than does the Bayesian and is therefore more easily led astray by rows of
tosses which produce subsequences with deviating relative frequencies.
The graphs in Figure 1 illustrate these connections. The left panel in
the middle row represents the explanationists (gray) and the Bayesians
(black) probabilities after each toss of a sequence of 250 tosses with a fair
coin. Both are early on pushed off the right track, but the explanationist
much more so than the Bayesian. This has a clearly recognizable effect
on the differences between the penalties that they incur, as can be seen in
the middle panel of the same row for the Brier rule and in the right panel
for the log score rule.
From these simulations the two update rules under consideration come
out doing better in different respects; this fact further buttresses the point
of this section. For there seems to be no clear answer to the question of
whether it is better, epistemically speaking, to use an update rule that in
general achieves greater accuracy than other update rules, even if typi-
cally not much greater accuracy; or to use an update rule that is less likely
50 100 150 200 250
0.02
0.04
0.06
0.08
50 100 150 200 250
0.05
0.1
0.15
0.2
0.25
0.3
50 100 150 200 250
0.5
1
1.5
50 100 150 200 250
0.5
1
1.5
2
2.5
3
50 100 150 200 250
0.7
0.75
0.8
0.85
0.9
0.95
50 100 150 200 250
0.2
0.4
0.6
0.8
1
50 100 150 200 250
0.86
0.88
0.9
0.92
0.94
0.96
0.98
50 100150200250
0.0025
0.005
0.0075
0.01
0.0125
0.015
0.0175
50 100 150 200 250
0.025
0.05
0.075
0.1
0.125
0.15
Fig. 1. Simulations with a randomly chosen sequence of 250 tosses with a coin with
bias = .1 (top row), bias = .5 (middle row), and bias = .9 (bottom row) Left column:
probabilities assigned to the true bias hypothesis; middle column: Brier penalties; right
column: log score penalties. Gray squares: values for the explanationist; black
triangles: values for the Bayesian.
than another update rule to ever make one vastly inaccurate, even though
the former typically makes one somewhat more inaccurate than the latter.
Naturally, we would prefer a rule that offered the best of both worlds, or
better yet, of all worldsin other words, one that was most conducive to
any epistemic goal we might have. But there may be no rule that fits this
bill, and in any event, the above simulations give reason to believe that
Bayes rule is not that rule. To be sure, these simulations do not warrant
a more positive verdict about IBE either. But the aim of this section was
not to make a case for IBE, or even for IBE. Rather, it was to highlight
and question a hidden assumption in the inaccuracy-minimisation defence
of Bayes rule, to wit, that there is but one way in which an update rule
can be said to minimise inaccuracy.
In closing, I would like to consider a possible Bayesian response to the
foregoing. In section I, it was argued that what makes the dynamic Dutch
book argument objectionable may not so much be the fact that the argu-
ment addresses the wrong kind of rationalitypractical rationalitybut
rather that it focuses on just one possible practical concern to the neglect
of all others, apparently without good reason. If pragmatic considerations
can be legitimately invoked in the present discussion, then one might try
to argue that, even if the goal or goals that Bayes rule serves best are not
privileged from an epistemic viewpoint, they are privileged from a practi-
cal viewpoint. Specifically, one might try to argue that, if the results of
the above simulations have some general validity, and do not just hold
for the particular type of statistical model considered in this paper, then
the fact that updating via Bayes rule leads, on average, to a lower pen-
alty is enough to justify this rule. And indeed, it does follow from the said
fact that were one to keep betting on consecutive tosses with a given coin,
each time posting betting odds in accordance with ones degrees of belief
at that time, then one would maximise ones expected payoff by updating
via Bayes rule.
But it would be false to think that payoffs and scoring rule penalties
must always be so strictly related as in this example. To see how they can
come apart, suppose that you are the owner of one of two engineering
firms in a city. The two firms are in permanent competition for contracts
from the local authorities. As a rule, a contract goes to the firm that pro-
poses the best solution to whatever the engineering problem is that the
authorities want to be solved. Given that the engineers employed by your
firm are about as competent as the engineers employed by the other firm,
which firm comes up with the best plan typically depends on the accuracy
of the information on which they could base their proposals. Note that,
under these circumstances, being able to base ones proposal on informa-
442 IGOR DOUVEN
tion that is just slightly more accurate than the information available to
the competition is enough to get the contract. By contrast, if the available
information is less accurate, it is immaterial whether it is only slightly less
accurate or much less accurate: one will not get the contract either way;
in this case, a miss is as good as a mile. Again supposing the results of
our simulations to have some general validity, they suggest that if you
and your team update via IBE while the competition updates via Bayes
rule, then you are likely to get the vast majority of the contracts. Once in
a whilevery rarelya contract will go to the other firm. When it does,
the information on which your proposal was based was probably much
more inaccurate than the information on which the rival plan was based.
However, this has no financial consequences beyond the fact that you
miss out on the contract, which would also have happened if the informa-
tion available to you had only been slightly less accurate than the infor-
mation available to the other firm.
Thus, there are circumstances under which it is clearly better to update
via IBE, practically speaking, given that it greatly increases the chance of
ending up with a more accurate belief system than one would have ended
up with had one updated via Bayes rule. Naturally, this is only one sense
in which one can aim to minimise ones inaccuracy, and the foregoing
considerations in fact suggest that, from a pragmatic viewpoint, there may
be no one best update rule. Depending on the circumstances and on
what exactly ones interests are, Bayes rule may serve ones interests best,
or IBE may do so, or, perhaps, sometimes yet another rule may do so. In
epistemology, contextual approaches have been much in the limelight
lately. I am not aware of anyone suggesting that which update rule to go
by may have no context-independent answer. The suggestion is worth
exploring, I think, but this must wait for another occasion.
III. CONCLUSION
We have considered the (currently) main arguments for the claim that
Bayes rule is the only rational update rule. Against the dynamic Dutch
book argument, it was argued that if updating via Bayes rule has certain
practical advantages in comparison with other rules, a version of IBE
may have different, possibly more important, practical advantages in com-
parison with Bayes rule. Equally, against the inaccuracy-minimisation
approach, it was argued that if Bayes rule minimises inaccuracy in one
or in some senses, a version of IBE may minimise inaccuracy in another
sense or senses, where for all anyone has shown, no one of these senses
can be said to be privileged in the sense of capturing the epistemic
goal. Consequently, neither the dynamic Dutch book approach nor the
inaccuracy-minimisation approach succeeds in challenging the rationality
of IBE-updating.
25
University of Groningen
25
I am grateful to Jake Chandler, Sylvia Wenmackers, and two anonymous referees for
helpful comments. An earlier version was presented at the Institut JeanNicod (Paris). I
thank the audience for stimulating questions.
444 IGOR DOUVEN

I. Douven - 2013 - Inference To The Best Explanation, Dutch Books, and Inaccuracy Minimisation

Transféré par

Informations du document

Titre original

Copyright

Formats disponibles

Partager ce document

Partager ou intégrer le document

Options de partage

Avez-vous trouvé ce document utile ?

Ce contenu est-il inapproprié ?

Droits d'auteur :

Formats disponibles

I. Douven - 2013 - Inference To The Best Explanation, Dutch Books, and Inaccuracy Minimisation

Transféré par

Droits d'auteur :

Formats disponibles

INFERENCE TO THE BEST EXPLANATION, DUTCH

BOOKS, AND INACCURACY MINIMISATION

iff conformity to E would produce

would produce (p. 194).

Vous aimerez peut-être aussi