Bayesian Explanationism | Ted Poston, Ch. 7 in "Reason and Explanation"
Intro:
The explanationist view I have defended maintains that propositional justification consists in a subject’s overall explanatory position. There is no purely autonomous justification that provides a subject with autonomous reasons, whether from experience or intuition. Rather a subject has justification for a claim only if that claim is part of a virtuous explanatory view which beats relevant competitors. One final challenge to this view comes from Bayesianism.
So Poston presents Bayesianism as a challenge to explanationism, I guess because it seems to be a competing theory that is in certain respects more precise.
In that sense, compatibilism is attractive: then you show that they are not competitors but instead supplementary, in the best case.
Yet the Bayesian structure is silent about explanation and its virtues. How do the explanatory virtues fit into a Bayesian story? An old methodological rule is to seek the most simple, powerful, and conservative explanation (i.e., the ‘best’ explanation) of some phenomena and to believe the best explanation when it is good enough. [ejw: ref to Harman 1965 here]. It is not apparent from the Bayesian apparatus why the norm to seek good explanations should be reflected in a Bayesian view of inference. (149)
In Poston's summary, the Bayesian view of scientific methodology is a way of "modeling inductive confirmation" (149)
Claims, goals, arguments:
- Argue for compatibilism (a new position according to Poston)
- "I argue that explanationism is consistent with Bayesian requirements of coherence and conditionalization." (149)
- "Furthermore, I argue that inductive confirmation requires explanatory information." (149-150)
- Supplementary: "Adding in explanatory considerations to undergird the Bayesian framework makes for a more powerful theory." (150)
Summary of his position, which he calls explanatory Bayesianism:
Main point: "It requires that a subject’s prior probability distribution reflect explanatory virtues." (150)
- Simplicity: "A Pr-function ought not have priors that give simple theories lower priors than complex theories." (150)
- Explanatory power: "A Pr-function should reflect the power of an explanatory hypothesis in the relevant likelihoods." (150)
- (Don't know how to name this): "A Pr-function should distribute probability over the most fundamental explanatory parameters instead of the Platonic heaven of all possible explanatory parameters." (150)
- "In light of new mysteries, one should seek a new prior distribution that departs least from one’s previous prior distributions while maximizing simplicity and explanatory power." (150)
Bayesianism
Commonalitites IBE and Bayesianism
- Bayesian holistic confirmation; depends on background knowledge (captured in prior beliefs)
-
"This Bayesian analysis of confirmation fits strongly with the explanationist’s contention that justified belief requires a background theory." (151)
-
"both Bayesians and explanationist stress the importance of coherence and that coherence is a holistic property of a belief system" (151)
- This is less clear to me at the moment.
Cognitive values associated Bayes rule
"A hypothesis is confirmed by a piece of evidence to the extent the hypothesis is plausible, it predicts the evidence, and also to the extent the evidence was surprising (151-152, my emphasis).
- Plausibility -> prior
- Prediction -> likelihood
- Surprise -> prior
The unification argument "appeals to these cognitive virtues." (152)
Q: I'm not familiar with the unification argument!
Bayes and interpretation of probabilities
Bayesians typically identify probabilities with a person’s degree of beliefs. Bayesians then put forward two characteristic claims: probabilism and conditionalization. The thesis of probabilism maintains that a rational subject has degrees of belief that satisfy the axioms of the probability calculus. Conditionalization is the thesis that a subject who acquires new evidence should update her degrees of belief in accord in Bayes’s theorem. (153)
Arguments for Bayesianism
I've seen most of these somewhere else, so I'll skip over it quickly.
Quick abstract:
Since Bayesians are typically personalists about the interpretation of probability, this amounts to the claim that humans have degrees of belief for propositions and that these degrees of beliefs should be coherent. To this end Bayesians offer Dutch Book arguments, representation theorem arguments, and scoring rule arguments for these conclusions.8 A final argument for Bayesianism is what I will call ‘the unification argument.’ This argument takes a different route by arguing for a Bayesian view of induction on the grounds that it unifies many natural methodological principles via the mathematical theory of probability. (154)
Dutch books
Dutch books: p.154-157
Representation theorem arguments
Representation theorem arguments: 157-159
What the representation theorem argument shows is that if a subject’s preferences satisfy these constraints then a subject can be represented as an expected utilitymaximizer with a coherent set of degrees of beliefs. Usually, a representation theorem argument assumes that if a subject can be represented by a set of utilities and degrees of belief then the subject has that set of utilities and degrees of belief. Consequently, such arguments show that if a subject’s preferences satisfy those constraints then the subject has a coherent set of degrees of belief. To the extent subject’s preferences ought to satisfy those constraints a subject should have a coherent set of degrees of belief. (157)
Preference consistency: preferences of rational agent obey constraints C
Representation theorem: if indeed an agents preference obey constraints C, then these preferences "can be represented as resulting from some unique set of utilities U and probabilistically coherent degrees of belief B relative to which they maximize expected utility.
To complete the argument, it assumes representation accuracy: if we can assume the preceding two points, then the agent actually hase U and B.
Probabilism (agent has probabilistically coherent set of beliefs) follows from these principles! (158)
The representation accuracy is a pretty huge metaphysical assumption. Many people (e.g. Christensen or Mark Kaplan) argue instead that the constraints for a coherent set of beliefs are normative rather than actual.
This is captured by informed presence: a rational agent should prefer A over B if A is more probable, assuming they are both equally desirable.
Most importantly, Poston claims this argument is consistent with explanationism!
The representation theorem argument for probabilism is consistent with explanationism. Nothing about Informed Preference requires violating explanatory constraints and the constraints arising from the explanatory virtues do not conflict with it. (159)
Moreover,
The constraints are justified on overall explanatory grounds in line with the argument in the previous chapter for an explanationist justification of the principle of noncontradiction. An agent who violates the constraints or Informed Prefence [type in original-ejw] is under a more general obligation to explain herself. (159)
Unification argument
"Kevin Kelly and Clark Glymour observe that the unification argument is what truly motivates Bayesian confirmation theory." (160)
Also: Curd, Cover, Pincock, cited in page 160:
A considerable part of the attraction of the Bayesian approach is that, on the basis of a simple theorem of probability—Bayes’s theorem–and a few assumptions about rationality and degrees of belief, Bayesianism promises a unified explanation of a wide range of accepted principles and truisms of scientific methodology.
Methodological principles that the Bayesian framework integrates:
- "Surprising predictions have special evidential value." (160)
- "Simple theories are more likely to be true" (160); N.B. this metaphysical assumption can in itself be questioned.
- "Ad hoc additions to theories to address problematic evidence reduce the plausibility of those theories." (160)
- "Diverse sets of evidence lend stronger support to theories than narrow sets of evidence." (161)
- "Not every theory is equally well confirmed by the evidence it entails." (161)
- "There is an elegant Bayesian solution to paradoxes of confirmation such as the raven paradox and Goodman’s riddle." (161)
The unification argument isn’t a reason to reject the old explanationist methodology; rather it’s a vindication of it. (161)
Bayesian Unification
- Surprising predictions (161-162)
- Value of diverse evidence (162-164)
- Def. of "confirmational independence" (163)
- Confirmation of evidence, that is independent of the confirmation by other evidence (i.e. we have 'diverse' evidence), will be additive on top over that other independent evidence.
- Bayesian treatment of raven paradox (164-166)
Van Fraassen's boosting argument
On page 168:
If IBE is to be made compatible with Bayesianism, then IBE must somehow be formulated in terms of probability, such that theories with high explanatory power should receive a high likelihood and theories with low explanatory merits a low one. But given that Bayesian conditionalization can be shown to always maintain coherent degrees of belief as it follows the axioms of probability theory, this only leaves two options for IBE. The first option is that IBE, understood in probabilistic terms, is reducible to Bayes' theorem. This breaks our requirement that IBE is not made trivial. The second option is that IBE does something else than conditionalization with Bayes' theorem by "boosting" the probabilities of more explanatory hypotheses, but this violates the coherency that Bayes' theorem guarantees. Hence, IBE would then be an incoherent account of inference and for example susceptible to Dutch books where agents engage in bets they are guaranteed to lose.
The heuristic view
Van Fraassen’s incompatibility argument assumes that probabilistic IBE requires that explanatory hypotheses are given an extra boost of probability. But there are good reasons to think that standard explanatory considerations—simplicity, fit with background evidence, and explanatory power—correlate with the prior probability and likelihoods of hypotheses. (169)
One argument against compatibilism may simply be that Bayesianism is entirely correct, both normatively and descriptively, as an account of induction. (cf. 170, Lipton 2004, 104)
Lipton has four routes to respond to this type of criticism against compatibilism:
- "Bayesianism is not a correct normative theory" (170)
- Discussion of standard problems of Bayesianism
- Bayesianism is a correct normative view, but not descriptively.
- "the views are compatible because Bayesianism is just logic for degrees of belief, and logic doesn’t conflict with any substantive epistemological theory." (170)
- "on this third response, Bayesianism doesn’t imply a dynamics to belief since if one learns that e and sees that updating by Bayes’s theorem leads one to have an unacceptable high posterior probability then one can change one’s priors to restore probabilistic consistency." (170)
- So explanation driven induction does not necessarily contradict Bayes
- Lipton 2004, 106
- IBE is actually a way of realizing Bayesian reasoning. See Lipton 2004, 107.
Lipton's weak heuristic view
Lipton's response is the fourth one.
Note, that Lipton's account acknowledges that "Bayesianism provides the correct normative and descriptive account of inductive practice" (170)!
Lipton’s idea is that even though we are not good at abstract probabilistic reasoning, as shown by the work of Tverksy and Kahneman, we use explanationist reasoning to help us respect the constraints of Bayes’s theorem. (171, cf. Lipton 2004, 112)
But whereas Tversky and Kahneman see heuristics as a replacement of Bayesian reasoning (and this use of heuristic leads to irrational fallacies etc.), Lipton (optimistically?) thinks of these heuristics as helping us meet the constraints that Bayesianism places on our beliefs, instead of replacing them. (Lipton 2004, 112) So
- Bayesianism indicates constraints for our degrees of belief to be rationally distributed
- IBE is a heuristic for meeting those constraints
Lipton's position (2004, 114-116):
- "First, explanatory considerations can help inquirers determine the likelihood—Pr(e|h)—which helps transition from a prior to a posterior. This may be done by considering how well h would explain e.60 The value of Pr(e|h) is directly proportional to how well h explains e. As we saw above, Newton’s theory offered no explanation of the bending of light in a strong gravitational field; hence P(e|Newton’s theory) is low. But the relevant likelihood for general relativity is high since it explains the phenomenon." (172)
- "Second, explanatory considerations can help inquirers determine prior probabilities.61 One obvious suggestion is that considerations of simplicity and fit with background evidence are primary determinants of a prior. Another suggestion is that considerations of the explanatory power of a hypothesis should be reflected in the prior of a hypothesis." (172)
- "Third, they can help determine which data are relevant to the hypothesis under investigation. Lipton observes that Bayes’s theorem does not indicate which evidence to use for conditionalization and that explanatory considerations can help to determine this." (172)
- How do we see which evidence is relevant for a hypothesis? Precisely because it would explain it. Bayesianism is silent here.
- "explanatory considerations can play a role in the judgment of which hypotheses promise to be fertile. Bayes’s theorem has nothing to say about the genesis of hypotheses. Lipton observes that scientists are interested in “fertile hypotheses with high content.”63 Explanationists, thus, can offer to the Bayesian a way of understanding the context of discovery that aims for theories with some well-studied Bayesian virtues. This latter consideration shows that explanationism captures an aspect of inferential practice where Bayesianism is silent.64"
- Very true.
Where is the weak heuristic view indeed... weak?
Okasha asks whether Bayesians are explaining the correctness of old methodological views or whether they are just representing those rules by adding probabilis tic constraints. This question gets at the essence of a problem with the weak heuristic view. It’s unclear just why a Bayesian should be moved by the explanationist retrofitting move. (173)
Poston's strong heuristic view
Claim: "the Bayesian ought to bring on board explanatory considerations as constraints on a Pr-function because explanatory considerations are required for inductive confirmation."
Response to Henderson's emergent compatibilism:
A recent paper (Henderson, forthcoming) argues for a new compatibilist response which differs from the constraint-based compatibilism I am urging. Henderson argues for an emergent compatibilism according to which IBE and Bayesianism track the same epistemic considerations. Henderson’s view makes use of a ‘natural prior’ to get inductive confirmation. On my view this ‘natural prior’ must come from explanatory considerations. In contrast to Henderson’s view, I see constraint-based compatibilism as itself compatible with her emergent compatibilism because explanatory constraints are necessary for a natural prior. (footnote 67, p.173-174).