It Can Be Rational to Change Priors
Ru Ye, Wuhan University
Forthcoming in Analysis
Abstract
According to a widely held norm of rationality, one should not
change prior credences without new evidence. An important argument for this norm appeals to accuracy considerations, which says that
changing priors doesn’t maximize expected accuracy. This is because
accuracy measures are strictly proper, and thus any probabilistically coherent person regards her own priors as uniquely maximizing expected
accuracy compared with other priors.
This paper attempts to resist the accuracy argument against changing priors. We argue that even if rational epistemic decisions maximize
expected accuracy according to strictly proper accuracy measures, it can
still be rational to change priors sometimes. The core idea of our argument is that changing priors can be rational if one wants to maximize
not just one’s current, short-term accuracy but also future, long-term accuracy. Our argument, if successful, shows that considering long-term
accuracy has significant ramifications for the accuracy-first project.
Keywords: Bayesian Epistemology; Epistemic Utility; Prior Probabilities
1
1 Introduction
According to a widely held norm of rationality, one should not change beliefs
without new evidence. In the Bayesian framework that models beliefs as obtained
by conditioning a prior credence function on evidence, the norm says that one
should not change priors. This norm has been motivated in a variety of ways.
Before listing the motivating arguments, we should first clarify the notion of
prior. The notion has been used in at least two different ways in the literature.
Sometimes, it means the credence held at the time before receiving a given piece
of evidence. Priors in this sense are fixed and cannot intelligibly change. Other
times, it means some hypothetical credence function that serves as one’s epistemic
standards, and one’s actual credence at a time is supposed to be determined by
conditioning this hypothetical credence on one’s new evidence at the time. Priors
in this sense can intelligibly change. In this paper, we use the notion of prior in the
second way.1
Here are some existing arguments against changing priors. First, changing
priors violates conditionalization: if one gains no new evidence, one should conditionalize on a tautology, which will result in the same priors (Meacham, 2014,
2015). Second, changing priors violates conservatism, which forbids changing beliefs without specific reasons to doubt them (Harman, 1995; Vahid, 2004, p. 97).
Third, there is a practical argument, which says that having stable priors is critical for carrying out one’s long-term plans (Titelbaum, 2015). Finally, and most
importantly, there is an argument based on accuracy considerations, which says
that changing priors doesn’t maximize expected accuracy. This is because accuracy
measures are strictly proper, and thus any probabilistically coherent person regards
her own priors as uniquely maximizing expected accuracy compared with other
1
Thanks to an anonymous reviewer for the urge to clarify the notion of prior.
2
priors (Schoenfield, 2014, forthcoming).2
It’s not too hard to find ways to resist some of the above arguments. For instance,
the conditionalization argument only works for some formulations of conditionalization but not for others.3 The argument from conservatism is debatable if
conservatism is questionable (Christensen, 1994). And the practical argument can
be doubted as to whether it shows that changing priors is irrational in the epistemic
sense of rationality; it also doesn’t show that changing priors is always irrational,
because sometimes you have no plans hanging on the belief in question.
The accuracy argument, on the other hand, may strike many as harder to resist.
The argument rests on three assumptions: accuracy is the primary goal in epistemic decisions, accuracy measures should be strictly proper, and rational decisions
should maximize expected utility. All three assumptions, while not uncontroversial, are broadly accepted in the present epistemic utility theory literature.
This paper tries to resist the accuracy argument against changing priors without challenging the three above-mentioned assumptions. We argue that even if
accuracy is the primary goal in epistemic decisions, and even if rational epistemic
decisions should maximize expected accuracy according to strictly proper accuracy measures, it can still be rational to change priors some times. The core idea
of our argument is that changing priors can be rational if one wants to maximize
not just one’s current, short-term accuracy but also future, long-term accuracy. We
explain this argument in Section 2, answer an objection in Section 3, and conclude
in Section 4 that considering long-term accuracy has significant ramifications for
the accuracy-first project.
2
For additional arguments against changing priors, see Titelbaum (2015).
More specifically, it doesn’t work for the wide-scope formulation of conditionalization
(Meacham, 2016).
3
3
2 Changing Priors to Be More Accurate in the Future
Suppose that you are ordering lunch at a restaurant. You have two options: a
chicken salad or a duck-egg soup with tofu, and all you care about is the taste.
You’ve had the chicken salad before, and you know that it tastes good. You’ve
never had the duck-egg soup before, but the menu photos tell you that you will
either love it or hate it. For concreteness, let’s suppose that you are certain that
the utility of the chicken salad (where the utility is entirely based on gustatory
pleasure) is 1; the utility of the duck-egg soup is either 5 or −5, and your confidence
in the two possibilities is even. Let’s also assume you’ll return to this restaurant
next week to choose between the same two dishes. Today, which dish?
If you are myopic—that is, if you only care about maximizing your immediate
gustatory pleasure and not your future self’s gustatory pleasure—then you should
choose the chicken salad today because it has greater expected utility: its expected
utility is 1, while the duck-egg soup’s expected utility is 0. And assuming that
your information on the two items doesn’t change, when you return to the same
restaurant next week, you should choose the chicken salad again on the basis of
the same expected-utility reasoning.
However, if you also care about your future self’s gustatory pleasure, then
you should choose the duck-egg soup today. Although choosing the soup doesn’t
maximize current expected gustatory pleasure, it will give you new information,
namely, information about what the duck-egg soup actually tastes like, and this
new information is beneficial for promoting your future gustatory pleasure—you
can use the new information to make a more informed choice when you return to
the same restaurant next week.
Here is a more detailed explanation. If you also care about your future gustatory
pleasure, then what you should maximize is expected total-utility, a utility function
that factors in not just the gustatory pleasure of your current self but also the
4
gustatory pleasure of your future selves. For the sake of simplicity, let’s assume
that this total-utility function is simply the sum of the current and the future utility.4
The expected total-utility of the chicken salad choice is therefore 1 + 1 = 2, while
the expected total-utility of the duck-egg soup choice is 0.5(5 + 5) + 0.5(−5 + 1) = 3:
if the soup has utility 5, then choosing it today will enable you to learn that it has
utility 5, and thus you will choose it again the next time, giving you a 5 in utility
again; if it has utility −5, then choosing it today will enable you to learn that it has
utility −5, and thus you will avoid the soup the next time and choose the chicken
salad instead, giving you a 1 in utility the next time; your confidence in the two
possibilities is even. As a result, the duck-egg soup choice has a greater expected
total-utility.
The lesson is this: for a person who wants to maximize utility for her future
selves as well as utility for her current self, sometimes an option is choiseworthy
because it promises to bring new information, and the new information might
allow one to maximize expected total-utility even if it doesn’t maximize expected
current utility.5
Now, we want to claim that this lesson generalizes to the decision of whether to
change priors. If you only care about the accuracy of your current belief state, then
switching priors is a bad idea, since it means adopting a less expectedly accurate
credence. But if you are not so myopic, that is, if you also care about the accuracy
of your future belief state, then sometimes switching priors is a good idea, since it
can bring you new information that you would not otherwise be able to get, and
getting new information can lead to a more accurate belief state in the future even
from your current perspective.
To drive home the point, consider the following dramatic scenario. Suppose
4
In a more realistic example, one might discount one’s future utility at a certain rate.
This is the main idea behind the exploitation-exploration tradeoff in reinforcement learning
(Sutton & Barto, 1998, pp. 26–30).
5
5
that your current prior cr at t0 distributed on W = {w1 , w2 } is (1/2, 1/2) and you
are considering whether to change it to a different prior (1/3, 2/3). God tells you
that if you make the change at t1 then he will tell you which world is actual,
so that you will update to an omniscient credence with the new information at
t2 . Suppose that your accuracy measure is given by the Brier score B(cr, wi ) =
P
− w j (cr(w j )–Iwi (w j ))2 . Then your expected total-accuracy of not changing priors is
−1/2[(1/2 − 1)2 + (1/2 − 0)2 + (1/2 − 1)2 + (1/2 − 0)2 ] − 1/2[(1/2 − 1)2 + (1/2 − 0)2 +
(1/2 − 1)2 + (1/2 − 0)2 ] = −1, whereas the expected total-accuracy of changing priors
is −1/2[(1/3 − 1)2 + (2/3 − 0)2 + 0] − 1/2[(1/3 − 0)2 + (2/3 − 1)2 + 0] = −5/9.
Of course, the above scenario is unrealistic. In real life, you are not guaranteed
to get new information by changing priors. But the fact remains that changing
priors can sometimes raise the probability of getting new information that you
would not otherwise be able to get (we will give some intuitive examples soon).
When the expected probability increase is great enough and the new information
is substantial enough,6 the expected total-utility of changing priors—where the
total-utility is the sum of the accuracy of your current beliefs and the accuracy of
your future beliefs—can be greater than the expected total-utility of maintaining
your current priors.
This argument rests on the assumption that changing priors can sometimes
raise the probability of gaining new information. This assumption is plausible for
two reasons. First, changing priors can help one gain higher-order evidence about
one’s old priors, specifically, evidence about the irrationality of the old priors.
Empirical research on motivated reasoning shows that people are generally better
at detecting flaws in reasoning that goes against their beliefs (Kahan et al., 2017).
So, by changing priors and reevaluating the old priors from a fresh perspective,
one can see more effectively whether the old priors are based on flawed reasoning
6
Informally, a piece of information is substantial enough relative to a decision problem when it’s
expected to make a great enough difference to one’s decision.
6
and thus irrational.
Second, changing priors can sometimes promote information gain by providing
a greater motivation for conducting inquiry. For instance, sometimes one is neutral
on a scientific theory, but one can see that if one becomes highly confident in the
theory, it will give one a greater motivation to conduct experiments in order to
confirm the belief, and thus one will get information that one would not otherwise
be able to get. And in order to be motivated to perform the inquiry and get the new
information, one must sometimes actually adopt the different priors, rather than
simply assuming or pretending to have them (Aronowitz, 2021). This phenomena of conducting inquiries in order to confirm a belief is not uncommon among
scientific communities.7
Of course, changing priors doesn’t always raise the probability of gaining new
significant information. If you know that you won’t live long, there isn’t much
time to gain new information; if you are at the late stage of your inquiry, where
relevant information is already abundant, there isn’t much new information to be
gained; and if you know that all the relevant new information will be generated in
an entirely passive manner, namely, in a manner that’s entirely insensitive to what
you believe, what you believe doesn’t affect what new information you gain. So,
we are not arguing that it’s always rational to change priors.8
In sum, the core idea of our argument is this: if you care about the long-term
accuracy of your future credences as well as the short-term accuracy of your current
7
Some have even argued that inquiring in order to confirm a belief is not irrational (Falbo, 2021).
As a reviewer notes, our argument that changing priors can help gain information might apply
only to agents who are not ideally rational, and this raises two worries. First, non-ideal agents
often don’t have direct control over their priors. Second, insofar as we are talking about non-ideal,
or ‘boundedly rational’ agents, why not just say that boundedly rational agents ought to become
ideally rational, instead of following these half-measures involving changing priors?
In response to the first worry, we think that rational requirements don’t assume ’direct’ control:
even if one cannot change one’s epistemic standards ‘at will,’ one might still be able to take some
intermediary steps to improve them over time. In response to the second worry, we think that
theorizing about half-measures can be interesting as ‘second-best epistemology’: it can be interesting
to ask questions like ’given that boundedly rational agents cannot become ideally rational, what’s
the second-best thing for them to do?’
8
7
credences, then your epistemic utility function can be represented by the sum of the
two accuracy scores;9 changing priors can allow you to gain new information and
thus increase long-term accuracy; if the expected increase in the long-term accuracy
is great enough, it can outweigh the expected loss in the short-term accuracy and
thus make changing priors rational.
Now, it’s clear that our argument doesn’t challenge the three assumptions made
in the accuracy argument against changing priors mentioned at the beginning of
this paper. Our argument accepts that accuracy is the most important goal in
epistemic decisions. It doesn’t require us to deviate from the decision rule of
maximizing expected utility. It also doesn’t challenge the assumption that accuracy
measures are strictly proper (as illustrated in the above God case, where we use
the Brier score as our accuracy measure.)
Of course, the epistemic total-utility function, which factors in the causal impact
of possessing a credence on the accuracy of one’s future credence, is not strictly
proper. But this doesn’t challenge the strict-propriety assumption about accuracy
measures, since the total-utility function is not an accuracy measure, although it’s
entirely based on accuracy considerations. More precisely, it’s not the kind of
accuracy measures that are at the heart of existing arguments for strict propriety.
Those kinds of accuracy measures are concerned with the accuracy of a credence
function when we view the credence function as an abstract mathematical entity,
without considering whether it’s had or who has it; they are not concerned with
the causal consequence (in terms of accuracy gain or loss) of having the credence
function (Carr, 2017, pp. 519–20).10 This is understandable: the causal consequence
9
For ease of exposition, we consider the accuracy scores from only two time slices. Our epistemic total-utility function can allow infinitely many time slices, if the utility of future accuracy
is discounted at appropriate rates, so that the sum of the utility of infinitely many time slices is a
convergent series.
10
The main argument for strict propriety is as follows (Joyce, 2009, pp. 277-9; Greaves & Wallace,
2006, p.621). Without strict propriety, some probability functions won’t maximize expected accuracy
relative to the function itself; such probabilities will be ruled out as irrational a priori, i.e., they will
be ruled out regardless of one’s evidential situation, because the person who has the probability
8
of having a mathematical function as one’s credence is highly individual, and thus
having a credence cannot be given a general accuracy score.
3 Violating Evidentialism?
We’ve argued that changing priors can promote long-term accuracy by promoting
information gain. Now, here is a worry: our argument assumes that we can trade
off short-term accuracy for long-term accuracy; such a trade-off can be problematic
because it violates evidentialism in some cases: there are cases where a credence
doesn’t conform with one’s current evidence, but adopting it promises to bring
valuable information and thus is permissible according to our proposal.11
We have two responses to this evidentialist worry. First, this paper only aims to
convince epistemic consequentialists, i.e., those who agree that whether changing
priors is rational depends on whether it maximizes expected accuracy. It’s not
intended to convince those who think about the rationality of changing priors from
an evidentialist point of view.12
function will be self-undermining. But we should not rule out a probability function as irrational a
priori, since each probability function may be rational in some evidential situations.
This line of reasoning works only if we consider a probability function as an abstract mathematical
function, regardless of whether it’s had or who has it as a credence function. If we consider the
causal effect of having a probability function for some person at some time, then saying that having
a probability function doesn’t maximize expected accuracy relative to the function itself for some
person at some time won’t rule out the function as irrational a priori—it might be still rational for
other people, or for the same person in a different situation, to have the probability function.
11
As a reviewer points out, considering information gain might threaten not just evidentialism
but almost all epistemic norms: for almost any plausible epistemic norm, we can imagine a situation
where violating the norm can help gain information. We share this concern, and we think that the
response developed in this section is applicable to other norms: considering information gain can
still be compatible with the dynamic versions of epistemic norms, which claim that, as one’s inquiry
progresses, one’s violation of those norms should be less and less frequent.
That said, we prefer to focus on evidentialism in this section, since gaining information by
changing priors is not just theoretically possible but also empirically supported, as we have argued.
12
Hedden (2015, pp. 475–6) has proposed an evidentialist argument for stable priors appealing
to the uniqueness thesis, which says that any evidence supports only one credence function. Given
uniqueness, a rational person doesn’t change priors because doing so leads to a credence that’s
unsupported by evidence. This paper doesn’t try to engage with this style of argument. For a
response to Hedden, see Titelbaum (2015, pp. 673–4).
9
Second, and more importantly, even from an evidentialist point of view, the
kind of tradeoff we advocate here is not egregious, because it’s compatible with
what can be called ‘dynamic evidentialism’: although there are times when one
must violate evidentialism, one’s dynamic epistemic behaviors in an inquiry should
exhibit an evidentialist trend in the following sense: as one’s inquiry progresses,
one’s violations of evidentialism will become less and less frequent. This is because
as one’s inquiry progresses, one gains more and more information, and thus the
probability of gaining new valuable information decreases; so, the probability that
changing priors promote information gain decreases. This means that, as one’s inquiry progresses, one should increasingly focus on ‘maximizing expected accuracy
given the information one already has’ rather than ‘seeking new information.’ As
a result, as one’s inquiry progresses, one should behave more and more like an
evidentialist.13
4 Conclusion
We have argued in this paper that changing priors can sometimes be rational by
promoting long-term accuracy. Our argument can be used to illustrate a broader
point, namely, that considering long-term accuracy has significant ramifications for
the accuracy-first project. This is because taking into account long-term accuracy
opens up new possible ways of defining the epistemic utility function, where utility
is still entirely based on accuracy considerations. In this paper, we take epistemic
utility as a simple sum of short-term and long-term accuracy. However, we can also
impose additional structures on the utility function. For example, we may want
the utility function to reflect not just whether the total accuracy score is excellent,
13
This idea is nicely captured by the popular epsilon-greedy algorithm with a decaying epsilon
in reinforcement learning, which says that as one’s inquiry progresses, the frequency of random
exploration decreases and the frequency of exploitation given one’s evidence increases (Sutton &
Barto, 1998).
10
but also some global features, such as whether the accuracy of beliefs over time
shows an improving trend and whether the improvement is fast enough.
Considering these new structures that we might impose on epistemic utility
functions has many benefits. It will not only provide a fresh understanding of
our accuracy goals, but also provide new resources to be used in the endeavor of
recovering rationality norms from accuracy considerations. As it happens, some
fruitful attempts in this direction have already been made in the research on formal
learning theory, a research program that aims to recover rationality norms from
long-term accuracy goals.14 So, this paper can also be viewed as a call for further
research in this direction.15
Bibliography
Aronowitz, S. (2021). Exploring by believing. Philosophical Review, 130(3), 339–383.
Briggs, R. (2009). Distorted reflection. Philosophical Review, 118, 9–85.
Carr, J. R. (2017). Epistemic utility theory and the aim of belief. Philosophy and
Phenomenological Research, 95(3), 511–534.
Christensen, D. (1994) Conservatism in epistemology. Nous, 28(1), 69–89.
Falbo, A. (2021). Inquiry and confirmation. Analysis, 81(4), 622–631.
Greaves, H. & D. Wallace (2006). Justifying conditionalization: Conditionalization
maximizes expected epistemic utility. Mind, 115(459), 607–632.
Harman, G. (1995). Rationality. In Osherson, D. N. & Smith, E. E. (eds.) An
Invitation to Cognitive Science: Thinking (3). Cambridge: MIT Press, 175–212.
14
A classic example is the learning-theoretic argument against counterinductive priors, which
says that such priors are irrational because a counterinductivist cannot converge to the truth in the
long run (Schulte, 2018). More examples of learning-theoretic arguments for rationality norms can
be found in Ye (2023, pp. 165–6).
15
An earlier version of this paper was presented at Zhejiang University and Wuhan University.
This work has benefited greatly from the thoughtful comments and suggestions provided by the
audience.
11
Hedden, B. (2015). Time-slice rationality. Mind, 124(494), 449–491.
Joyce, J. M. (2009). Accuracy and coherence: Prospects for an alethic epistemology
of partial belief. In F. Huber & C. Schmidt-Petri (Eds.), Degrees of Belief (V342,
pp. 263–297). Springer.
Kahan, D. M., Peters, E., Dawson, E. & Slovic, P. (2017). Motivated numeracy and
enlightened self-government. Behavioural Public Policy, 1, 54–86.
Meacham, C. J. G. (2014). Impermissive bayesianism. Erkenntnis, 79(Suppl 6),
1185–1217.
Meacham, C. J. G. (2015). Understanding conditionalization. Canadian Journal of
Philosophy, 45(5), 767–797.
Meacham, C. J. G. (2016). Ur-Priors, conditionalization, and ur-prior conditionalization. Ergo: An Open Access Journal of Philosophy, 3.
Schoenfield, M. (2014). Permission to believe: Why permissivism is true and what
it tells us about irrelevant influences on belief. Nous, 48(2), 193–218.
Schoenfield, M. (forthcoming). Meditations on beliefs formed arbitrarily. In T. S.
Gendler, J. Hawthorne & J. Chung (eds.), Oxford Studies in Epistemology, 7.
Schulte, O. (2018). Formal learning theory. The Stanford Encyclopedia of Philosophy
(Spring 2018 Edition), E. N. Zalta (ed.), URL = http://plato.stanford.edu/archives/spr2018/entries/
formal/.
Sutton, R. S. & Barto, A. G. (1998). Reinforcement Learning: An Introduction. Cambridge: The MIT Press.
Titelbaum, M. G. (2015). Continuing on. Canadian Journal of Philosophy, 45(5),
670–691.
Vahid, H. (2004). Varieties of epistemic conservatism. Synthese, 141(1), 97–122.
Ye, R. (2023). Permissivism, the value of rationality, and a convergence-theoretic
epistemology. Philosophy and Phenomenological Research, 106, 157–175.
12