Inference to the best explanation: does it track truth? | David H. Glass
This was really an excellent and very readable paper. It is not too lengthy, makes interesting conceptual distinctions, and supports them with clear results. The author also clearly situates the paper in relation to other authors, e.g. Lipton.
Note that this paper is explicitly about IBE as selecting amongst competing hypotheses and not about "abduction proper" i.e. the discovery or positing of relevant hypotheses.
abbrev. IBE: inference to the best explanation.
Two key challenges for IBE
- How should one explanation "be compared against another so that the best explanation can be identified" (412).
- In order to do this, we need to make IBE precise and have some measure for the quality of an explanation.
- "Is there any reason for thinking that the best explanation is likely to be true? Or to put it another way, does IBE track truth?" (412)
- "if IBE is to be accepted as a rational mode of inference, there must be some reason for thinking that it provides a good strategy for determining the truth." (412)
Lipton already pointed out that when considering what is the best explanation, this is not the same as what is the most likely explanation. Peter Lipton uses the distinction between the loveliest and the likeliest explanation.
Of course, a feature of the loveliest explanation should be its likeliness; an hypothesis with a very low posterior probability is almost certainly not the best explanation.
On page 142-3 Glass introduces the issue whether IBE is compatible with Bayesianism and discusses various positions. See the paper for those.
In what way does explanation play a role in the Bayesian process of updating beliefs, which is defined on (subjective) probabilities?
Even if Bayesianism is compatible with IBE, does it even define what explanation is best, and does it even need such a definition? (413)
The latter is clearly necessary for an account of IBE: "clearly if there is no account of 'best explanation', it cannot be IBE" (413).
Given the compatibilist position and assuming Bayesianism does have an account of "best", then it has to (with some varieties) be an account based on what explanation is most probable given the evidence.
But as pointed several times in this paper, this can only be done at the cost of making IBE trivial (reducing it to probability and not asking what explanatory features make an explanation the best). E.g.:
An obvious thought would be to simply define the best explanation to be the one that is most probable given the evidence. This approach would certainly establish the link between goodness of explanation and truth, but only at the expense of making IBE trivial. (412)
Quote from Peter Lipton (2004, p. 60):
We want a model of inductive inference to describe what principles we use to judge one inference more likely than another, so to say that we infer the likeliest explanation is not helpful
In other words, IBE intends to be a descriptive (right?) account of human inductive reasoning. Even though the benefit of the Bayesian approach is that it is truth tracking, it is really lacking in this regard. (In the discussion section Glass comes back to this distinction, saying that IBE is descriptive and Bayesianism normative (p. 425))
Goal and requirement of an IBE account
Goal:
The goal of IBE is to give an account of 'best explanation' that is conceptually distinct from 'most probable explanation' and yet show that the best explanation will often be the one that is most probable.
Glass (2007) identifies two components required for an account of IBE:
- "an account of what constitutes an explanation" (413)
- "a suitable methodology for comparing competing explanations" (413)
I would say in regard to point 1, the discours would move more in a direction of inductive reasoning models (e.g. deductive-nomological) and issues of relevance: what makes an explanation relevant? What features does a good explanation have? Perhaps also a bit more focus on abduction than IBE proper.
Point 2 is more specifically about ranking competing hypotheses, for which the probability-game is directly relevant.
This also means that a suitable method for comparing hypotheses, whether based on Bayes' theorem or on a coherence- or confirmation measure, that does not necessarily mean they provide a full account of IBE. For what is missing, consider the following point:
intensional vs hyperintensional:
One concern about using probability measures to compare explanations is that probability is generally considered to be an intensional notion and so logically equivalent propositions have the same probability, whereas this is not the case for explanation. Explanation is perhaps better understood as a hyperintensional notion since it depends not only on the logical content of the explanation, but on how it is presented. (414)
Important remark: what matters here is "not the absolute value which a measure yields for the goodness of an explanation, but the relative value when compared with competing explanations." (415)
Cf. this with the rationale between the Bayesian solutions to the Raven Paradox: a good explanation should be contrastive.
Discussed approaches
Note that some approaches do not indicate a total ordering over hypotheses, meaning that sometimes not all hypotheses can be ordered in terms of bettered or worse. See paper for details.
Bayesian (415):
- MPE: most probable explanation is the one with the higher posterior probability
- ML: maximum likelihood
- Fits the deductive-nomological idea that a good explanation makes the evidence more probable
- But is extremely sensitive to the prior, so not very suitable
Remarks (416):
- A Bayesian can always state they are simply only interested in probability, and not the best explanation. Likewise, they are not committed to saying MPE is also the best explanation.
- In fact, Bayesians are not committed to making a particular inference at all! It's primarily about updating degrees of belief.
Confirmation (416-417): see paper. I'm not super interested in these.
It is important to point out that the amount of confirmation is not the same as a posterior; instead "the degree to which E confirms H is a measure of how much more probable the evidence E makes the hypothesis H" (416)
Coherence (417-418):
- OCM: overlap coherence measure
- FCM: Fitelson coherence measure
- Shogenji coherence measure
Methodological considerations and experiment
MPE finds the true explanation often and is thus truth-tracking, but as an account of the best explanation it is trivial. A better account of IBE should nevertheless approximate MPE in order to be truth-tracking. (419)
The experiment shows that OCM approximates MPE with 99%. A second experiment shows that when the prior distribution is not certain, OCM on average even outperforms MPE.
Discussion
What is at stake is the "viability of IBE as an approach to inductive reasoning" (424)
A challenge for IBE is to define what is meant by 'best explanation' and show that there is a link between goodness of explanation when so defined and truth or high probability so that selecting the best explanation is a good strategy for acquiring truth.
OCM addresses the second requirement for an account of IBE (see above).
Claim:
This still still leaves a question concerning the importance of IBE as an inductive reasoning method distinct from Bayesianism. One approach to this question would be to think of IBE as a descriptive account of how humans actually reason, whereas Bayesianism is the normatively correct way that humans should reason. If so, the rationality of IBE depends on how well it tracks Bayesianism (my emphasis, 425)
Remaining issue: how to compare IBE and Bayesianism?
- "The focus of Bayesianism is on updating degrees of belief based on conditionalisation, but there is no reason to think that IBE must be formulated in a similar way as van Fraassen assumes in his critique of IBE. Rather than updating degrees of belief, IBE is concerned with selecting the hypothesis that provides the best explanation of the evidence and inferring its truth."
- Implying that in order to compare with Bayesianism, we need to cast it in a form like MPE.
- But if we do that, we must be explicit about this; Bayesianism may have nothing to do with explanations at all, or at least not committed to IBE.
Conclusion
The results show that of the other measures the OCM approach, which uses the overlap coherence measure to identify the 'best explanation', identifies the actual explanation almost as often as the MPE approach. Yet this account seems much more plausible as an account of 'best explanation' (427)
I think the evidence for the latter statement is still pending. Glass himself mentions that this is a open matter of psychologists.
His claims on the performance off OCM seem very warrented though. Great paper.
Refs:
Lipton 2004:
Glass 2007: