Locating IBE in the Bayesian Framework | Jonathan Weisberg
- Van Fraassen's argument against compatibility "causes more trouble for compatibilists than is typically thought" (125)
- "Bayesianism in its dominant, subjectivist form, can only be made compatible with IBE if IBE is made subservient to conditionalization in a way that robs IBE of much of its substance and interest." (125)
- "If Bayesianism and IBE are to be fit together, I argue, a strongly objective Bayesianism is the preferred option." (125)
Just like Lipton, Weisberg emphasizes that we should not go for an easy compatibilism by making IBE trivial.
Good read, interesting paper. Too bad there's no concrete proposals for a general objectivist approach, but still a convincing case why it may be worth it to explore such a perspective.
Summary (from end of paper)
I have been arguing that the compatibility of IBE and Bayesianism is not so simple a matter as one might think. Van Fraassen’s argument that IBE conflicts with Subjectivist Conditionalization demonstrates a real conflict between the two rules, since they track different things. IBE tracks explanatory goodness or beliefs about explanatory goodness, while Subjectivist Conditionalization tracks prior conditional degree of belief. Moreover, resolving the conflict by demoting IBE to heuristic status robs IBE of much of its appeal and interest. IBE not only inherits the subjectivism of subjective Bayesianism, but also becomes inapplicable in many cases where explanationists and avowed compatibilists want it to apply. In response, I urged explanationists and compatibilists to abandon compatibility with Subjectivist Conditionalization, and embrace the objective Bayesian framework instead. To bolster the appeal and promise of that approach, I offered some preliminary suggestions for fitting IBE and objective Bayesianism together. (142)
Introduction
Van Fraassen's position is "that IBE is bad epistemology because violating conditionalization is irrational" (125).
Main compatibilist strategy: does IBE really clash with conditionalization?
Weisberg's analysis on why this compatibilist approach has not been satisfactory:
Unfortunately for compatibilists, however, IBE cannot be made consistent with conditionalization, at least not without sapping much of the substance from IBE. To a first approximation, the problem is this: IBE tracks explanatory goodness, while conditionalization tracks prior conditional degree of belief. For the two rules to agree, ‘explanatory goodness’ must be understood in a way that makes it subservient to prior conditional degree of belief, which robs IBE of much of its appeal and interest. (126)
"IBE tracks explanatory goodness, while conditionalization tracks prior conditional degree of belief." -> it's not entirely clear to me what a "prior conditional degree of belief" is. Does he simply mean: it tracks prior belief as it is updated when provided with new evidence? I would personally not call this belief "prior" anymore, as this is what is commonly known as the "posterior" degree of belief.
And the “what if the best isn’t very good?” concern no longer applies once IBE is translated into degree-of-belief terms, which we are about to do.
Some complications
W.r.t. IBE: "What makes an explanation better? Should we accept the explanation that actually is best, or just the one that we think is best? What if there may be explanations we haven’t thought of yet? What if the best explanation isn’t very good?" (126)
Weisberg tries to circumvent these issues.
- Sticks to clear-cut examples with clear intuitions. Then we don't necessarily need a definition of explanatory goodness to follow the argument.
- Do we favor the best explanation or the one we think best? See "burglar case" on p. 8.
- Unconceived explanations -> cases of non-existent priors.
- What is the best is not good? Weisberg argues this does no longer apply once IBE is translated into degree-of-belief terms. (REVIEW whether I agree)
Incompatibility argument
In Weisberg's (normative) reading of IBE: "IBE is a claim about what we ought to believe, but bare bones Bayesianism has nothing to say about qualitative belief; it deals solely in degrees of belief." (Weisberg 2009, 127)
Two options:
- Introduce a qualitative notion into Bayesianism
- Frame IBE in terms of degrees of belief
However, several authors reject the idea of Bayesian qualitative beliefs, e.g. Jeffrey 1970, Christensen 2004.
Weisberg departs from the second option, also because this is more common in the literature.
Weisberg does point out that an alternative could be to view qualitative beliefs as explanatory reasoning + degrees of belief. (see Levi 1980, Maher 1993).
Summary of Van Fraassen:
"Van Fraassen (1989) interprets IBE as a policy of favoring explanatorily superior hypotheses by giving their subjective probabilities a post-conditionalization boost."
What is subjectivist conditionalization?
Subjectivist Conditionalization makes your new degree of belief in a hypothesis a function of two things: the new evidence and your conditional degree of belief prior to acquiring that evidence. This is a “subjective” policy in two senses. First, probabilities here are understood as the subjective state of an agent, namely her degrees of belief. Second, to the extent that rational agents may have different prior degrees of belief, there is room for rational inter-subjective disagreement. If you rationally hold degree of belief p'(H|E)=x, while I rationally hold p'(H|E)!=x, then we will be rationally obliged to draw different conclusions given evidence E. (127)
Basically: standard Bayesian conditionalization for updating degrees of belief, after seeing evidence E, where rational conclusions may differ based on different background knowledge. There are no further requirements and rational inter-subjective disagreement if perfectly fine.
You can add more (relatively weak) constraints on the probabilities for them to be interpreted as degrees of belief (whereas objective Bayesians impose very strong constraints):
- Van Fraassen: principle of Reflection (Van Fraassen 1984, 1995)
- Lewis' Principal Principle (Lewis 1980)
Foreshadowing the main point: objective Bayesians are already comfortable with violating subjective conditionalization
Weisberg argues that mainly the subjective Bayesian is bothered by Van Fraassen's argument because it relies heavily on showing that subjective conditionalization is violated.
However, the objective Bayesian does not need to be bothered by this as much: because she places strong constraints on priors and conditionalization anyways, it is not a problem to say: hey, if your (prior) beliefs violate the objective constraints, then you should adjust your prior beliefs or otherwise subjective conditionalization will not lead to rational results anyways.
So the more objective Bayesian will not be troubled if IBE requires violations of Subjectivist Conditionalization, since she already sees violating Subjectivist Conditionalization as legitimate when necessary in order to respect other principles she holds. (128)
Challenges for compatibilism (based on subjective conditionalization)
The main answer of compatibilists is to deny there is a conflict with IBE. You can construe a simple coin-flipping example where Bayesian conditionalization and IBE agree.
However, Weisberg argues, such examples heavily rely on the prior and conditionals.
compatibilism depends crucially on a tight correlation between an explanation’s superiority and its prior conditional degree of belief. In the coin-flip example, applying Subjectivist Conditionalization yielded a higher posterior degree of belief for H because H started with a higher conditional degree of belief. If instead we had had p(H|E) < p(H'|E), applying Subjectivist Conditionalization would have yielded the explanationist-unfriendly result that q(H) < q(H'). That is, after all, the content of Subjectivist Conditionalization. The agreement between prior conditional credence and explanatory superiority is thus crucial; to the extent that these two factors diverge, so will IBE and Subjectivist Conditionalization. (130)
In other words, those examples are not that convincing in the sense that they are silent about the assumptions they make on priors. (I suspect they will often say "this is a natural prior to have"; but then we should ask, why is it natural?).
Weisberg will defend the following claims.
Claims to be defended
- "prior conditional credence and explanatory goodness do diverge" (129)
- possible divergences are not just because IBE is an approximate heuristic (the most common compatibilist stance); and cannot simply be ignored as exceptions
- the heuristic view makes IBE "inapplicable in important cases" (129)
- The heuristic view also "forces IBE to take on the radical subjectivity of subjective Bayesianism, diminishing one of IBE's central appeals, namely its agreement with common sense and paradigm instances of good scientific inference." (129)
Where may IBE and subjective Bayesianism diverge?
Weisberg gives a nice example on page 130 about the most likely explanation of finding your house with the front door open, broken lock, valuables missing, mess in the living room.
First explanation: burglar broke in and stole your valuables.
Second explanation: a second burglar breaks the lock, only to encounter another burglar that came in through the window. They fight, which explains the mess. A police officer shows up, but he can't catch the burglars as they run away, and to mask his own failure, he opportunistically steals the valuables before disappearing.
The question here is: do we not agree that the first explanation is better even when we assign higher prior conditional belief on the second explanation?
Claim of Weisberg: "So explanatory goodness, whatever it is, looks to be at least somewhat independent of prior conditional degree of belief." (130)
Counter 1: explanatory goodness is relative to background belief
Counter 1: "explanatory goodness is relative to background belief" (130), so isn't it dependent on priors?
Weisberg's point rather is that the subjective Bayesian is completely free to have whatever prior credences, however unreasonable they may seem, and that when such a case occurs, we do tend to reason using IBE that the conclusion of using subjective conditionalization does not lead to the better explanation.
Later Weisberg says it very explicitly: "Subjectivist Conditionalization tracks conditional belief, which is something different." (131)
In other words, subjective conditionalization does not place enough constraints on background beliefs and therefore gets into trouble. So in a way Weisberg says that Van Fraassen has a point that IBE violates the subjective conditionalization of purely subjective Bayesianism. However, compatibilists should not be bothered too much by this and try to show how IBE in fact does not violate it, but instead convert to objective Bayesianism.
Counter 2: IBE tracks beliefs about explanatory goodness, not actual explanatory goodness
Counter 2: "IBE tracks beliefs about explanatory goodness, not actual explanatory goodness"
Two ways of understanding IBE (131):
- "internalist": you should prefer the explanation that you judge better
- "externalist": you should prefer the explanation that is in fact better
Weisberg ansers this distinction does not matter for his point.
While the internalist reading might make your posterior credence depend on the idiosyncrasies of your prior credences, it makes it depend on a different idiosyncrasy than the one that matters to Subjectivist Conditionalization. It is possible to be very confident that H is a better explanation of E than H' is, and yet have p(H|E) < p(H'|E); this does not contradict the probability axioms or any other standard Bayesian principle. So, whether IBE tracks real explanatory goodness or merely believed explanatory goodness, Subjectivist Conditionalization tracks conditional belief, which is something different.
Very interesting, but I have a bit of trouble pinpointing exactly what the main point is. Hmm, ok, I think that Weisberg is mainly arguing against the claim that only the externalist view would cause trouble here w.r.t. subjective conditionalization. His point instead is more general: in the internalist view, it is not even guaranteed that subjective conditionalization tracks the explanation that you judge to be better, because, Weisberg argues, it tracks an orthogonal matter: conditional belief.
It doesn't matter that in the internalist view, the posterior is made dependent on idiosyncrasies of the prior belief. Even then, what matters is that subjective conditionalization tracks the conditional belief in H (so given evidence E), which depends on P(E|H), P(H), P(E); and this is not the same as tracking which H is judged to be a better explanation.
I have some doubts about this point, but I can't succesfully formulate them yet.
Counter 3: the best explanation just is the one with the highest conditional prior
Weisberg has used this term before but:
Q: what exactly is the conditional prior? I think Weisberg just means P(H|E) before subjective conditionalization on the evidence.
In terms of the burglar example, a compatibilist may claim that we simple are "in the habit of having a higher prior conditional degree of belief for explanations like the one-burglar explanation." (131)
If you are conceding that we judge what is the best explanation out of habit (very weak account of IBE!) then this can be accomodates with subjective conditionalization:
$$p(H|E) > p(H'|E) \rightarrow p(H)p(E|H) > p(H')p(E|H')$$
I.e. the explanation with higher prior conditional credibility, has a higher prior and/or likelihood. When we "reverse engineer" Bayes' rule like this, we may identify p(H) with the background beliefs and the likelihood as a fit with the data.
This amounts to making IBE and subjective conditionalization compatible by "rigging" the account of explanatory goodness:
- Leaves out virtues like "simplicity, elegance, unification" (132) etc.
- Explanatory goodness is now entirely dependent on an agent's belief
- What's left of any intrinsic relation "to the hypothesis or its relationship to the evidence" (132)
It is quite obvious that this does not capture the spirit of IBE and doesn't match how we think of good explanations.
To me, it sounds a bit like saying: this explanation is the best because we already considered it best (i.e. we only consider previous evidence, not the current evidence, right?!)
Main takeaway:
- We should not understand explanatory goodness entirely in terms of the prior or the likelihood.
- What if there are no relevant prior degrees of belief (132)
The heuristic view sees divergence as an exception
Examples: Okasha (2000), Lipton (2004)
Okasha, for example, says that our explanationist inferences use “explanatory considerations as an aid to calculating the priors and likelihoods needed to apply Bayes’s theorem,” (Okasha, 2000, p. 703) and Lipton and McGrew make similar pronouncements. (cited in Weisberg 2009, 132)
If Bayes is the norm that IBE as a heuristic approximates, you need to at least show that explanatory goodness in IBE's account track (prior conditional?) Bayesian credence.
If explanatory goodness is fallibly but reliably correlated with prior conditional credence, then it is plausible that following IBE is generally legitimate as a way of trying to meet Bayesian standards. (133)
Isn't this something that Glass attempts? Review: Glass-Does_inference_to_the_best_explanation_track_truth.md In this view it's possible that the worse explanation (in IBE's terms) gets a higher prior conditional credence. That's possible because the break in scenario is quite exceptional. We are cognitively limited when it comes to doing actual subjective conditionalization, so here IBE can help us to approximate it to best of our ability.
Weisberg's claim: "they concede too much when they demote IBE to mere heuristic status, and grant normative primacy to Subjectivist Conditionalization." (133)
Why? Arguments:
- Subjective conditionalization only applies when prior degrees of belief exist, but this is not the case for many scientific breakthroughs that introduce new concepts or new hypotheses. (So I would summarize: abduction proper is not accounted for in the Bayesian framework!)
Response: "each of Okasha, Lipton, and McGrew intends for IBE to apply in cases where Subjectivist Conditionalization does not apply because the requisite priors do not exist. Okasha, for example, is explicit that "IBE illuminates where Bayesianism is silent"" (134)
But I fully agree with Weisberg that when IBE is supplementing Bayesianism, then you transgress the idea of IBE being a heuristic. It does not only approximate Bayesianism, but adds something to it. This is weird when at the same time the "normative force" of IBE depends on its approximation of subjective conditionalization. Make choice, basically!
You get into weird scenarios, because you're suddenly committed to saying that IBE suddenly does stand on its own in a more normative sense in cases where subjective conditionalization is not applicable. (134; Weisberg mentions that Okasha 2000 says something like this).
Really nice: Weisberg powns Lipton's favorite example:
Lipton is not so explicit that IBE should apply where Bayesianism is silent, but his favored exemplar of IBE is a case where Bayesianism is silent. Lipton’s example is Semmelweis’s realization that childbed fever was caused by the transmission of “cadaverous matter” from the hands of attending surgeons and students to patients in labor. Semmelweis was the house officer of the First Obstetrical Clinic at the Vienna General hospital in 1846, where the mortality rate from childbed fever was approximately 13%, as compared with 2% in the Second Obstetrical Clinic. After considering and eliminating several potential explanations, Semmelweis noticed that a friend of his died of similar symptoms when he pricked his finger during an autopsy. Semmelweis conjectured that there was a connection between the handling of cadavers and the infections in the First Clinic, and found that physicians and students would frequently go straight to the First Clinic from the autopsy room, but not so with the Second Clinic. He concluded that infectious material was being transmitted from the cadavers to the patients in the First Clinic, but not the Second. (134)
Weisberg's analysis, in defense of IBE and against the heuristic view:
Plausibly, it had never occurred to him, and before his friend’s death he had no prior degree of belief in either the evidence acquired or the hypothesis he ultimately settled on. And that is certainly true of those to whom he brought his results. Nevertheless, Semmelweis could readily appreciate the strength of the evidential support for his hypothesis, and it was on the force of that same strength that he ultimately convinced others. If Semmelweis’s case is an exemplar of IBE, then IBE is not subservient to Subjectivist Conditionalization. (134)
The heuristic approach makes IBE inherit the subjectivity of subjective Bayesianism
And this robs IBE of most its intuitive appeal (135-136).
One of the best things IBE has going for it is that it agrees with common sense and our ordinary practices of reasoning and argumentation in cases like this. And the fact that it completely disregards all these considerations in cases like the burglar example is one of the worst things Subjectivist Conditionalization has going against it. Faced with such a conflict between IBE and Subjectivist Conditionalization, if the explanationist yields to the Bayesian, she spurns common sense and strong intuitions, undermining much of the attraction of her view. (136)
Best of both worlds: give IBE primacy
I suspect that the explanationist has much to offer the Bayesian, but the potential benefits of a marriage of the two views are lost when we make IBE subservient to Subjectivist Conditionalization. If we grant IBE primacy instead, and use it to shape a more objective Bayesianism that rejects Subjectivist Conditionalization, we are in a position to develop a Bayesianism that is free of the limitations that come with Subjectivist Conditionalization. Such a view would be a truer friendship between the explanationist and the Bayesian than the heuristic approach to compatibilism, since each party would stand to benefit from the other. The explanationist can help the Bayesian fill in the gaps in her view, by providing a way to construct priors for new theories, for example. And the Bayesian provides the explanationist with a powerful and well-studied formal framework in which to formulate the details of her view. (Weisberg 2009, 136)
IBE and objective Bayesianism
The main difference: priors do not just come from prior degrees of belief, but "the correct prior probabilities to use are determined by rules like the Principle of Indifference, rather than by whatever degrees of belief you happen to have just before the evidence comes in." (137)
The current proposal is that the explanationist should adopt the objectivist framework, and use explanationist thinking to fix objective prior probabilities, either in conjunction with, or in place of, existing objectivist principles. Ideally, explanationist considerations would complement existing objectivist principles like the Principle of Indifference, yielding a more applicable and reasonable brand of Bayesianism. (137)
Definition of objectivist conditionalization:
At any given time, your credence in an arbitrary proposition H ought to be p(H|E), where p is the correct a priori probability distribution, and E is your total evidence at that time. (137)
So note that
- p is here a prior distribution
- E is the total evidence so far, not new evidence (we are talking priors here).
In this case, the posterior uses a prior that is not just whatever happens to be your prior belief, but what your prior belief should be if you take into account how the currently available "evidence bears on a hypothesis" (137)
The explanationist's role in the Bayesian framework is to characterize prior distribution p.
To ensure compatibility with the explanationist, this prior distribution p must be further constrained such that, for background assumptions B, we have:
p(H|E /\ B) > p(H'|E/\B) if H is a better explanation (in IBE terms) against the background assumptions.
Weisberg proposes this as a project; he himself does not have a concrete proposal.
The lack of concerete proposals in the objective camp (evaluate this) may also be why the main compatibilist position is the heuristic view.
He does offer some directions
Possible directions for the objectivist project
- Translate "existing explanationist treatments of problems in confirmation theory, translating them into objective Bayesian terms" (138)
- E.g. Goodman's riddle, see 138.
- Stability as virtue
Weisberg does three case studies that are each relatively complicated. Read later in more detail if necessary.
The third one seems most interesting, but is still questionable in my opinion.
I think he is trying to formulate a solution to Bertrand's Paradox of the Principle of Indifference: depending on your formulation of a problem to define priors over, you can get very different results; so the principle of indifference does not straightforwardly describe a unique prior distribution. See the section on Bertrand's Paradox in Easwaran-Bayesianism II .
Q: In what sense is compatibility with subjective conditionalization dropped? After constraining the priors, you still use Bayes' rule right?