Inference to the Best Explanation, Bayesianism, and Knowledge | Alexander Bird in "Best Explanations" eds. Kevin McCain and Ted Poston
My general impression is that Bird requires a "super-objective" Bayesianism because he has a very strong account of IBE. So he is on the complete opposite spectrum from those (subjective) Bayesians who make compatibility easy by making IBE too trivial, i.e. "best" is simply "most probable".
Many arguments revolve around "E=K".
- How can Bayesian reasoning result in knowledge K?
- Phrased differently: how can the result of a Bayesian inference result in evidence E for further inference?
Conclusion
In full, because it's a great summary:
The heuristic approach to compatibilism between Bayesianism and IBE can see off the challenge from the subjective Bayesian. Explanationist considerations do not supplement Bayesian considerations by giving a boost to the probability of a hypothesis when conditionalized on the evidence. Rather we can see explanationist factors as corresponding to the various elements of Bayes’s theorem, in particular the prior P(h) and the ratio of the likelihood to the expectation of the evidence, P(e|h)/P(e). Still, that cannot be all that explanationist considerations achieve, for IBE reasonably aspires to give us knowledge, whereas subjective Bayesianism does not. This may motivate a move from subjective to objective Bayesianism. But that position faces the problem that no scientific proposition can reasonably be accorded credence of 1, in which case no proposition can ever be known or be evidence. So we need to move beyond the personalism of both subjective and objective Bayesianism: we should not understand the probabilities with which we are concerned as credences but rather as something else. That something else must allow it to be possible to have evidence and knowledge and to accord the relevant propositions probability 1. The proposal here is that this is achieved by identifying the probabilities with assessments of plausibility in the light of the evidence (thus drawing on Williamson’s evidential probability).
Consideration of how a hypothesis acquires plausibility on explanationist grounds (e.g. explanationist values that feed into plausibility are generated by assessments of similarity to successful, highly truthlike theories) shows that it can be objective and be correlated with the truth. IBE does face an additional problem in that if it is to lead to knowledge, explanationist considerations must in fact be able to rule out competing hypotheses, i.e. reduce their plausibility to 0. While showing that takes us beyond this chapter, I have argued that so doing does not face any obvious and insurmountable obstacle.23
Positions
Standard Bayesianism
- Personalism: evaluations are credences (may be seen as willingness to accept set of bets on the truth of propositions )
- Probabilism: evaluations can be represented as probabilities if they are rational
- (Strict, synchronic) Conditionalization
The difference with standard epistemology is that the Bayesian allows rational partial belief.
Dutch book arguments shows that a personalist must also be a probabilist.
Standard Bayesianism is normative w.r.t. reasoning, on top of a "descriptive kernel (personalism)" (98)
Objective Bayesians add additional constraints on choice of priors etc.
Descriptive Explanationism
People and scientists "often hold the explanatory character of a hypothesis to be relevant to its epistemic evaluation." (98)
IBE:
Inference to the Best Explanation (IBE), understood as a description of an inferential practice central to science, is explanationism par excellence: it holds that such subjects come to accept a hypothesis because it provides a better explanation of the evidence than its rivals. (98)
One may additionally claim that scientists that apply IBE in fact reason as the ought to. Then IBE makes not only descriptive, but also normative claims.
For those only seeing IBE as descriptive, it's relatively unproblematic to see IBE as a description of what Bayesians normative prescribe.
In the case IBE entails normative claims, compatibilism with Bayesianism becomes much harder.
Heuristic compatibilism
Main proponents: Lipton, who sees IBE as a heuristic for applying conditionalization, and Ted Poston (2014, 149-81).
When is it rational to apply a heuristic?
- "if the saving of efforts outweighs the loss in reliability" (99)
- "the heuristic may be more reliable than the optimal method as poorly implemented by us inadequate thinkers" (99)
Given that we suck at computing probabilities, the rational thing to do may be to use a heuristic instead. We then need to show that the results of the heuristic roughly tend to agree with those of Bayesian conditionalization.
Bird's summary of the 'lovely' heuristic:
(H) If hypothesis A is a lovelier potential explanation of the evidence than B, then judge A to be more likely to be true than B. (100)
This heuristic should roughly correspond to the elements of Bayesian conditionalization. This is Bird's summary of Lipton's view on how this is done:
(I) If propositions h and e are ones that the subject has not given thought to before, then considering both as hypotheses, we make an assessment of their explanatory loveliness, taking into account such features as unification and simplicity.5 This corresponds to the Bayesian process of assigning values to the priors Po(h) and Po(e).
(II) We evaluate the explanatory loveliness of h in relation to the evidence e, and we then change the overall loveliness assessment of h accordingly. This corresponds to conditionalization, (S), whereby we move from the prior probabilities, Po(h) and Po(e) to the posterior probability, Pn(h).
We do not consider the total evidence, but only what is deemed relevant:
(III) Explanationist considerations determine which evidence is relevant to the hypothesis—the injunction to consider the total evidence, although implicit in Bayesianism, is not one we can actually implement, so we need a heuristic to guide us to the relevant evidence, viz. whether the evidence could be explained by the hypothesis.
Internal and external explanatory virtues
Some aspects of a theory, e.g. a hypothesis, can be tested for 'simplicity' etc. independently. Bird coins this (following Einstein) 'internal' explanatory virtues. (101)
Other virtues concern how a hypothesis relates to evidence, which may be called 'external' explanatory virtues. (101)
How can these "internal" and "external" explanatory virtues be accounted for in Bayes theorem?
Bird takes the account of Lipton and Poston and fills in some details.
How are priors guided by explanatory considerations?
Internal explanatory virtue of a hypothesis is its prior.
This holds "insofar as h is a novel hypothesis" (101, my emphasis).
That's kind of an interesting assumption though, due to problems of "old evidence" in the Bayesian approach.
In general, the subjective Bayesian can reasonably claim that we assign higher priors to more explanatory explanations:
Nonetheless, as a matter of descriptive fact, our actual assessment of the probability of priors does correlate with our assessment of their explanatory value. In advance of the evidence we do think that simpler, more unifying explanations are more likely to be true. (101)
Bird gives a nice argument about why explanatory hypotheses are favored in the Bayesian approach. His argument revolves around scepticism vs. the more desirable inductivism with explanatory hypotheses. Following Huemer (2009).
Case: we know we have a biased coin, but we do not know biased to which side. So applying the principle of indifference then prescribes that we assign equal probability to heads as to tails for the next throw.
But as also Bertrand's Paradox points out, there is no unique way to apply the principle of indifference. This is because we can make different choices in what our hypothesis space is (i.e. what we apply our "indifference" to and over which we distribute our credences). In this example Bird discusses two approaches, assuming we throw the coin 10 times. (101-102).
- If our hypothesis space consists of each possible outcome, we assign each hypothesis 1/1024 credence. Probability of various proportions of heads and tails can be computed.
- P(10 heads)=0.001
- If instead, we make hypothesis about the fairness of the coin (0, 0.1, 0.2, 0.3 ... 1), each hypothesis has prior credence 1/11.
- P(10 heads)=0.09
The first option is sceptical because it does not allow let's say finding heads nine times to increase your belief in the coin being more biased towards heads. This is not how Bayesian updating should proceed.
Huemer (2009) and Poston (2014) then argue that this is why Bayesianism needs the more explanatory hypothesis of case 2. (102)
Bird argues that the second scenario is also more descriptively accurate! I.e. when humans reason about this, we indeed consider hypothesis on the fairness of coins (from which the observed sequence can be inductive evidence), rather than making hypotheses directly over the sequences. (I find this convincing).
As a consequence, if this is descriptively the case, then it seems people "are indeed moved by explanatory considerations in the assignment of priors to hypotheses." (102) We are thus, arguably, explanationists about priors.
Put another way, we are as a matter of fact inductive thinkers, and since being a Bayesian inductivist requires (or at least motivates) explanationist thinking about priors, then we should expect that we are explanationists about priors. (102)
External explanatory virtues do not only correspond to the likelihood
Likelihood P(e|h) "can be very high even if h does not explain e well--if e is itself very probable: famously John Jones's taking a birth-control pill does not provide a good explanation of his not getting pregnant, since it was certain anyway that he would not get pregnant; still P(e|h) is very high." (102)
The opposite also holds: the likelihood can be low while the hypothesis does offer a good explanation, if the probability of evidence occurring is even lower, i.e. some very rare evidence with a low likelihood can nevertheless be very explanatory.
For example, Holmes suspects Professor Moriarty of having murdered Major Arbuthnot and of having disposed of his body, meticulously erasing evidence of the crime. Holmes then discovers a minute trace of the highly unusual Uzbek tobacco smoked by Arbuthnot in the cuff of Moriarty’s coat. Now given the meticulousness that the hypothesis ascribes to Moriarty, finding that tobacco trace is itself highly unexpected, so P(e|h) is low. But Holmes’s hypothesis is still an excellent explanation. The latter fact is still consistent with the demands of Bayes’s theorem, since P(e) is even lower. (102)
Bird's proposal:
I propose therefore that the external explanatory virtue of h in connection to e corresponds to the ratio P(e|h)/P(e), i.e. the measure of how much more probable e becomes when conditioned on h. (102)
So, we have that external explanatory virtues correspond to P(e|h)/P(e).
Using Bayes Theorem, this corresponds in turn to the ratio P(h|e)P(h), which is a measure of confirmation.
This was not directly clear to me, but I did the following derivation.
We trivially have:
$$P(h|e)=\frac{P(h|e)}{P(h)} * P(h)$$
With Bayes rule, we have:
$$P(h|e)=\frac{P(e|h)}{P(e)} * P(h)$$
So we see that
$$\frac{P(h|e)}{P(h)} = \frac{P(e|h)}{P(e)}$$.
Because P(h|e)/P(h) is a degree of confirmation (i.e. how much does our confidence in h decrease/increase after seeing e), we can also interpret P(e|h)P(e) as a degree of confirmation.
If we thus start with Bird's definition of 'external explanatory virtue' then we see that this can be construed as a measure of confirmation in Bayesian terms! (103)
Note however that the heuristic view does not say the explanatory and Bayesian counterparts are the same. They are still distinct. Rather, the claim is that following explanationism you roughly get the same results as assigning P(e|h) and P(e) to compute P(e|h)/P(e):
So a subject forming a judgement on the basis of explanationist considerations forms a judgement that corresponds to a Bayesian value for P(e|h)/P(e) without that being the result of preceding judgements corresponding to P(e|h) and P(e). (103)
The heuristic also shows how people can generally reach the correct solution (i.e. that some evidence becomes more likely were the hypothesis true) without knowing the exact probabilities; the important point is that the general conclusions are the same, e.g. a ranking of preferences
Even so, he might be able to be rather more precise about the degree to which the Moriarty hypothesis increases the probability of the presence of the ash. For the heuristic view to hold, it should be the case that as a matter of fact, subjects’ judgements of the external explanatory virtues of a hypothesis, i.e. when the evidence confirms a hypothesis by explanationist lights, should correlate with their assessments that the hypothesis makes the evidence more likely than it would otherwise be. And that looks correct: as the Holmes example and others show, if a hypothesis explains some evidence, then we think that the evidence proposition is more likely if the hypothesis is true than if we don’t make that assumption. (103)
How do you show that the conclusions based on the explanatory heuristics (using both external and internal explanatory values) indeed correspond roughly to the Bayesian calculations? This is hard (if not impossible) to do precisely. However, given judgement based on internal and external explanatory values, the involved reasoning seems to be multiplicative like in Bayes rule, rather than additive:
On the other hand, the way we combine those judgements does seem to mirror multiplication rather than addition or averaging. For example, if a hypothesis is known to be false then even unlikely evidence it explains well will not make us think that the hypothesis is even moderately likely—that corresponds to the product of a finite amount and zero, whereas averaging would suggest that a known false hypothesis could become likely on favourable evidence. (103)
Summary heuristic compatibilism
If the explanationist approach gives different results than what the Bayesian approach prescribes as rational, the explanationist approach can be said to be incoherent. (103)
If the explanationist approach fully agrees with the Bayesian results, then explanationism/IBE seems to be the worse theory, because it's not as precise and also controversial (104).
The heuristic approach solves this 'dilemma':
The heuristic approach advocated by Lipton and Poston resolves this challenge, by showing that explanationism is a useful substitute for Bayesian reasoning. For various reasons it is unnatural and difficult for human thinkers to implement the probabilistic reasoning required by standard Bayesianism. So we use explanationist considerations as a heuristic—a manner of reasoning that will by and large get the same result as Bayesian reasoning but without its costs and pitfalls. (104)
Bird's contribution is trying to specify the details of 'internal' and 'external' explanatory considerations.
Three challenges for compatibilism
Main assumption: IBE can lead to knowledge.
If IBE plays a role in science, than denying that assumption entails "scepticism about those parts of science that employ those practices" (104)
Knowledge and objectivity
So far we showed how subjective Bayesians may be content: explanationism is a heuristic for Bayesian conditionalization.
However, Bird considers that explanationists are perhaps not satisfied!
For the latter thinks that there is more to epistemology than is found in subjective Bayesianism. The explanationist thinks that on some occasions IBE can give us knowledge. She holds explanationist considerations to be a guide to the truth. So the explanationist cannot regard standard Bayesianism as a complete normative epistemology. She should say that priors and likelihoods informed by explanationism are better than those not so informed. Explanationist Bayesianism must reject subjective Bayesianism, and a strong form of objective Bayesianism looks more suitable, with constraints on which priors are rationally acceptable. (104)
In other words, in that case IBE is not entirely descriptive and has its own normative claims to bring in!
This is interesting: Bird argues that Lipton himself also secretly brings in some normative role for IBE. Consider Voltaire's problem:
Lipton himself needs to say something like this. Otherwise many of the prima facie problems that he raises for IBE (and addresses) would not be concerns at all. For example, Voltaire’s problem asks why it is that we should think that the explanatorily loveliest explanation is the actual (true) explanation. Should we think that the actual world is the loveliest of all possible worlds? But that’s irrelevant if all the epistemic work is done by Bayes’s theorem while explanationist concerns are merely the source of the values of the priors and likelihood, without making any normative contribution. However, concern with Voltaire’s problem is required if we want the resulting epistemic attitude to correlate with truth and IBE possibly to lead to knowledge. (105)
So again, Lipton's position may satisfy subjective Bayesians but not necessarily explanationists. (I'm also thinking about the challenge Glass posed: compatibilism without making IBE too trivial).
Knowledge, evidence, and credence
This section mainly deals with the problems of old evidence or background knowledge and the weird consequences of probabilities being 1 for the position of personalism.
E.g. for background knowledge we would have to commit to the following:
For a credence of 1 in p means that one would be willing to take a bet that would yield a one penny gain if p is true and unending torture if p is false. But few if any propositions are such that reasonable people would take such a bet. Certainly the contingent propositions that we take to be knowledge or to be our evidence are not typically of that sort. That looks to lead to the sceptical conclusion that we do not know those propositions after all and that no such propositions can ever really be our evidence.(105)
The last claim would also invalidate conditionalization (106).
Bird then discusses Jeffrey conditionalization and derives it from Kolmogorov's third axiom. Cf. Jeffrey-Probability_Kinematics.md
Refer to the paper for the details, but in any case the interpretation of Bird is very interesting. Conditionalization is just a special case of Jeffrey's rule, the difference is rather the question: when do we update? Strict conditionalization only updates when evidence is absolutely certain, whereas Jeffrey's rule updates already using uncertain evidence.
So what is distinctive about Jeffrey conditionalization is not the updating rule, which is in use whenever there is a change in the probability distribution. For example, as a result of strict conditionalization, some probabilities will change and ( F.) and its kin will be used iteratively to update further propositions. So what is at issue is not the updating rule but instead which changes in the probability distribution should rationally initiate updating. (107)
So the epistemological questions that Bayesianism (and most obviously: strict conditionalization) do not answer are (both on p. 107):
- When does something count as evidence?
- The only constraint for strict conditionalization is that P(E)=1
- When is the transition of some proposition to evidence rational, thereby triggering conditionalization?
- If this transition is not rational, then it also does not have to trigger the rational procedure of strict conditionalization.
Bird foreshadows answers that however require dropping personalism:
- Evidence propositions are propositions that are known
- If one's knowledge changes, probability distributions should be updated using conditionalization
Why does this require dropping personalism?
Problem with Jeffrey conditionalization
An additional issue for Jeffrey conditionalization is that the interpretation of evidence as "what is known" does not work, because evidence does not have a probability of 1 using Jeffrey conditionalization.
If the answer to above questions is that it's rational to update if something new becomes known (with P(E)=1), this does not help determining when it is rational to initiate Jeffrey conditionalization.
Jeffrey conditionalization allows for the retention as epistemically possible (probability > 0) propositions inconsistent with the evidence, but that makes no sense if the evidence propositions are known: propositions inconsistent with the subject’s knowledge are not epistemically possible. That is really just to say that there is no natural place for an account of evidence as knowledge in Jeffrey conditionalization. (108)
I find this sentence complicated: "Jeffrey conditionalization allows for the retention as epistemically possible (probability > 0) propositions inconsistent with the evidence". Isn't it rather better to say that there is uncertainty in the evidence? Bird's phrasing is more like 1) there is unambigious evidence, but 2) Jeffrey conditionalization takes several options into account, given epistemic uncertainty about the evidence; 3) many of these epistemic possibilities are inconsistent with the evidence.
So Jeffrey's account lacks internal motivation with regards to when updating is rational and is in need of external motivation:
So Jeffrey conditionalization needs an account of evidence that allows evidence to be less than knowledge, yet explains why it is rational to update in response to changes in evidence. So far no such account is forthcoming. (108)
Perhaps, Bird says, we can weaken the criterium: we can also rationally update when this is rationally permissible rather than required (by new knowledge with P(E)=1).
Bird then gives a quite technical account of when a change is permitted, of which I do not entirely follow the urgency; the point is that updating may be trigged by changes in probabilities that are independent of other probabilities (otherwise they are derived and then updating should be triggered by the probabilities on which it is dependent).
Then Bird spends some time arguing that even Jeffrey's famous candlelight example may be said to still rely on credences in the reliability of your own eyesight etc.
This is the summary and bottom line:
Jeffrey conditionalization looked liked a plausible option to take account of the fact that if personalism is true, then hardly ever, if at all, will a proposition have a probabil ity of 1—certainly not the propositions of interest to science. Bayesian updating, whether strict or Jeffrey, needs an account of the circumstances in which a change in a probability distribution may rationally initiate updating of other propositions, since coherence can always be achieved by reversing the initial change. Such an account will come from outside the Bayesian framework. Strict conditionalization can be supplemented by a natural account of when that is: when a proposition becomes known, that proposition may (and should) initiate updating. No such account is available to the Jeffrey conditionalizer. (109) Jeffrey conditionalizer.
- Conditionalizaion needs external account of when updating is justified
- Bird's proposal: when we have new knowledge
- This invalidates Jeffrey conditionalization, as it does not have such a strict concept of knowledge
Knowledge and IBE
Bird's summary of IBE:
When hypothesis A is a much lovelier potential explanation of the evidence than all its competitors (and it is in itself a good explanation, there is a good quantity of evidence, and alternative explanations have been sought assiduously), then infer that A is the actual explanation of the evidence and so is true. (109)
Bird calls the rule when to infer a proposition a "rule of detachment" (109).
Bayesianism on the other hand does not in itself tell you when to infer a statement! It just provides probabilities over a hypothesis space.
Having an extra rule (pick the hypothesis > 0.95 ) has many problems.
The problem Bird raises here again departs from the statement that the IBE explanationist has a rule of inferring a hypothesis such that that hypothesis is held to be true and adds knowledge.
The question is specifically how Bayesianism deals with the fact that an explanationists inference can be made whereas the evidence can nevertheless be consistent with multiple hypotheses, i.e. no hypothesis has a prior of 1.
(Right now, I actually do not see yet why this is a problem)
The problem:
- Explanationist inference will infer hypothesis will p < 1
- This means that the inferred hypothesis cannot be regarded as knowledge in a Bayesian account
- Does this mean that explanationism entails scepticism?
Jeffrey Dunn (2014), cited on p. 110, gives a nice example w.r.t. evidence that I think is very clear:
Can a proposition that is learned by inductive inference itself be evidence? Let p be the proposition that a new Pope has been elected, and s be the proposition that there is white smoke coming from the Sistine Chapel. My credences are P(p) = 0.5, P(s) = 0.5, and P(p|s) = 0.9. I then see white smoke over the Sistine chapel, so s is now in my evidence set. So, with conditionalization, my credences are now P(p) = 0.9, P(s) = 1, and P(p|s)=0.9. Can I now use p as evidence (e.g. with regard to some proposition concerning the immediate future of the Church)? According to Bayesianism, I cannot. For if p were now evidence, then it would have to have a probability of 1, whereas its probability is 0.9. (110)
Bird's summary of the challenges
Mere compatibilism with subjective Bayesianism answers Bayesian objections to IBE but fails to articulate how IBE can lead to the truth and so be a source of knowledge. (110)
Bayesianism requires that evidence propositions and other known propositions have probability equal to 1. If that probability is a credence then, since no one reasonably has credence 1 in most non-logical propositions, very few propositions will ever be known/evidence. The challenge is a consequence of Bayesianism’s commitment to personalism. (110-111)
If IBE selects among hypotheses that are all consistent with the evidence, Bayesian compatibilism implies that the inferred hypothesis has a probability of less than 1. But a known proposition cannot have a probability of less than 1. The challenge is a consequence of Bayesianism’s commitment to probabilism plus the claim that IBE selects among hypotheses that are all consistent with the evidence. (111)
Probability and plausibility
Interestingly, Bird argues that we must not only reject subjective Bayesianism as a candidate for compatbility with IBE due to not being compatible with the required objectivity for IBE, but also objective Bayesianism!
Objective Bayesians impose further constraints, e.g.
- Principle of Indifference
- Calibration (take known frequencies into account for priors)
- Jaynes' Maximum Entropy Principle
These have some a priori appeal but are not enough according to Bird.
Further a priori principles seem too thin to be able to do that—if they did, that would amount to an a priori justification of induction. So objective Bayesianism, as it is typically articulated, turns out to be an unsatisfactory answer to problem 1 after all. (112)
I.e. also objective Bayesians cannot sufficiently show "that there is a correlation between higher posterior probabilities and the truth." (112)
Bird's proposes a "'super-objective' Bayesianism where good priors and likelihoods are determined by explanatory considerations that cannot be captured as a priori principles but instead have a de facto reliability." (112)
So the first big difference is:
- objective Bayesianism supplements with additional a priori constraints
- Bird instead goes further and thinks a priori general principles are too weak; we should allow explanatory considerations with de facto reliability to play a role
Then we should drop personalism to solve challenge 2. Bird instead speaks of plausibilities instead of credences (i.e. they are no longer strictly understood as subjective credences).
See p.112 for Bird's relation to Salmon's objective Bayesianism that uses plausibilities. Apparently this style also fits Kuhn's philosophy of science.
(I do wonder however: plausibilities obviously ar still related to some subjective judgement? How strictly can we separate them from credences? Bird does also seen to gloss over this in footnote 14 on page 113, arguing against Williamson by saying that indeed plausibility cannot really be thought of independent of any "psychological make-up" even if it's the one of an idealized scientist)
This avoids problem 2 because "plausibilities can be given values of 0 and 1" (113) Bird does retain probabilism w.r.t. plausibilities (112).
What's an example of explanatory considerations with a de facto role?
Bird gives an extensive example that revolves around analogical reasoning, see page 113. He specifically shows explanatory elements that do not directly occur in the Bayesian formula's. (I'm wondering then, how he still hopes to integrate IBE and Bayesianism at this point).
So a hypothesis can gain plausibility by its connection to elements of the scientific context in a way that is independent of conditionalization on evidence. Analogy with well-confirmed hypotheses is an example of how this occurs. (114)
If we have such a de facto consideration (about where evidence-independent plausibility comes from) then the next question is how we measure how reliable that consideration is. (114)
Or said differently: we discussed "internal" explanatory considerations, but what about "external" considerations?
I have just argued that we make judgements of plausibility of a hypothesis that are independent of its conditionalization on the evidence—in so doing we respond to internal explanatory virtues, and these may be influenced, for example, by analogy with other theories. (114)
Assessing Objective Plausibility ('external explanatory virtues')
How much does the plausibility of a hypothesis change when we obtain evidence?
Inspiration from Kuhn:
Indeed, we can draw on a central lesson of the work of Thomas Kuhn (1970), which is that scientists’ assessment of any hypothesis is crucially influenced by their prior experience of other successful theories. According to Kuhn, such theories, which he called ‘exemplars’ or ‘paradigms’, are models that inform the development and assessment of newer hypotheses. Because Newton’s theory of gravitation was so successful, it became a model (in multiple ways) of what a good theory might look like, and so, as we saw, Coulomb’s proposed law acquired plausibility because of its similarity to that model. (114)
Link to explanatory virtues and learning by exemplars; idea is that learning by exmample can lead to good theories (that may be true). Inspired by Walker (2012):
Scientists acquire their sense of explanatory virtues from their training with and use of exemplars. Let us assume some degree of scientific realism—the theories in question have a high degree of truthlikeness. Let us also assume that theories similar to true or highly truthlike theories are more likely to be truthlike than less similar theor ies. If these assumptions are correct (and they are themselves at least plausible), then the practice of allowing one’s preferences among theories to be influenced by their explanatory virtues will be a truth-conducive practice.
The made assumptions seem very strong to me.
How does this relate to IBE and knowledge?
So now let us imagine a scientist who considers some candidate hypothesis in advance (or independently) of the evidence. Her assessment of its plausibility is influenced in part by the degree to which it possesses the explanatory virtues acknowledged by the community. In the light of the foregoing, and if the two assumptions are correct, we may expect there to be some correlation between the plausibility assessments and truth. Consequently, probabilities interpreted as (quantified) plausibilities will be objective in the sense relevant to challenge 1: probabilities will relate to truth, in that hypotheses assessed as more plausible will be more likely to be true. (115)
The general idea is that imperfections in the plausibility assessment of hypotheses will be "washed out" if more evidence comes in. Bird argues (and this point does seem fair to me) that for the subjective Bayesian any prior is permissible; here plausibility (in this explanatory sense) is an additional constraint on priors, so we will not have priors that are so extreme (assuming they are not themselves extremely plausible or implausible) that they cannot be "washed out" anymore. (115)
Bird's position is still not entirely clear to me w.r.t. how we now apply Bayesian computations. Is his position affected my counters against the "boosting" argument of Van Fraassen? That they break the rationality of strict conditionalization?
Plausibility, Knowledge, and Eliminativism.
Plausibility vs credence.
Working with plausibility has some advantages. Probabilities of 0 and 1 do no longer involve problems of people accepting crazy bets on the truth of a proposition. This solves challenge 2.
A plausibility of 1 means that the proposition is maximally plausible given the evidence and background knowledge. Rationally assessing a proposition as maximally plausible does not seem unachievable in the way that a rational credence of 1 looks unachievable for most propositions. (115)
Hmm okay okay.
Speaking in terms of plausibility does however have another property, namely that a proposition with plausibility 1 excludes competing hypotheses (i.e. they should have plausibility 0).
This gives a tension with IBE:
This looks to be in tension with IBE, since the latter is typically understood as making inferences on the basis of one explanation being a better explanation than the competitor hypotheses. (115)
But I again don't really see the exact issue at stake here. So I agree: IBE chooses the best explanation, which we can understand as the one that's most plausible (Bird's terminology). Then Bird states:
If that is right, then the inferred proposition will not be knowledge, since knowing h requires P(h)=1 and P(h')=0 for all competitor hypotheses h'. (116)
Then Bird reasons for eliminative abduction:
So knowledge of a proposition inferred on the basis of explanationist considerations requires that the competing hypotheses must be ruled out by the evidence and background knowledge: what may be called eliminative abduction (Bird 2005). It might be wondered whether explanationist considerations can ever achieve this. Perhaps the most that can be achieved is that a hypothesis is shown to be highly implausible. (116)
Almost all arguments in Bird's paper depend on the claim that IBE can lead to knowledge. But perhaps his account of IBE is too strongly formulated in this regard.
Bird does show (p. 117) some quite convincing cases where it's quite clear why and how a scientist reject a certain hypothesis.
Additionally to ruling out hypotheses, one must also "know that the true hypothesis is among the competing hypotheses she considers." (118)
E=K is safe either way
Returning to Dunn’s case, if the subject knows that a new Pope has been elected when she sees white smoke above the Sistine Chapel, then it must also be the case that the subject knows that if there is white smoke above the Sistine Chapel then a new Pope has been elected. (This could well be a case of eliminative abduction: another explanation for the smoke is that there is an accidental fire. But if so, the smoke would not be white and there would be fire alarms. So that alternative hypothesis is ruled out by the evidence.) If the subject does know that conditional proposition, then her conditional probability P(p|s) = 1, not 0.9 as in Dunn’s example. In which case, P(p) = 1, and Dunn’s problem disappears; there is no Bayesian problem with p being evidence. On the other hand, if P(p|s) = 0.9 then the subject does not know that if there is white smoke then there is a new Pope, in which case acquiring the evidence that there is white smoke cannot lead to knowledge that there is a new Pope (and so that proposition is not an evidence proposition). Either way, E = K is safe. More importantly, in the first case, where P(p|s) = 1 and so P(p) = 1, the inferred proposition is known (and so is evidence and has probability = 1).