Introduction

Deductive reasoning is quite well understood, but the verdict is still out on a well-rounded theory of abductive reasoning. This is especially interesting because science often proceeds through abductive reasoning from observed evidence to a hypothesis that would best explain that evidence, if the hypothesis is indeed true. This type of reasoning is studied as "inference to the best explanation" (IBE) and is a better candidate for a theory of scientific inference than the more limited enumerative induction [@Harman1965]. A central tenet of a reasonably strong account of IBE is that "explanatory considerations are a guide to inference" [@Lipton2000, p.184, my emphasis], meaning that a hypothesis is not just inferred and then explains some evidence, but that it is inferred precisely because it explains the evidence.

Any complete theory of IBE should offer some methodology for choosing one hypothesis over another, but should also give an account of why the chosen explanation is in fact better [@Glass2012, p.413]. A main strength of a Bayesian approach to scientific inference is that it gives a numerically precise answer to the question how newly acquired evidence bears on our degrees of belief in a set of hypotheses, namely through the process of conditionalization. But although Bayesianism thereby offers a clear method for comparing hypotheses, it has no explicit account of what explanatory virtues good explanations have. Neither is it trivial where the explanatory considerations that are central to IBE play their part in the numerical computations of Bayesianism. Additionally, the Bayesian framework is not, strictly speaking, an account of inference. In the process of conditionalization, the Bayesian framework updates probabilities for the hypotheses under consideration, but at no point obliges one to make an inference [@Bird2017, p.109]. One may therefore simply insist that the Bayesian framework only provides the most probable hypothesis, rather than the one that provides the best explanation [@Glass2012, p.416]

It seems that those who wish to make IBE and Bayesianism compatible have an easy way out: the hypothesis that the Bayesian rational procedure indicates to be the most probable, given the evidence, is the best explanation. However, even though IBE and Bayesianism could be made compatible by interpreting "the best" as "the most probable," this would be at the cost of making the account of IBE trivial, as IBE then loses its ability to describe what explanatory virtues guide the inferences that scientist make [@Lipton2000; @Glass2012]. For example, Peter Lipton has argued that scientists do not straightforwardly choose the most probable explanation, but the most lovely one, i.e. explanations that add the most understanding [@Lipton2000]. This raises the question if there is a Bayesian account of IBE such that it does not make IBE trivial, but maintains its core "explanationist" idea. This in turn raises the question how the explanatory considerations of IBE can be integrated into Bayesian computations. This literature review is therefore guided by the following research question:

What are the current main proposals for making Bayesianism compatible with IBE, without making IBE trivial?

To that end, this literature review explores the various attempts that have been made to provide an "explanationist Bayesianism" under the working assumption that "the Bayesian and the explanationist should be friends" [@Lipton2001, 109-119; @Lipton2004a, 120]. Within the compatibilist perspective there is a variety of views on exactly how they should be friends [cf. @Ladyman2005]. The aim of this literature review is to provide an overview of the key positions in the literature arguing for compatibility between IBE and Bayesianism. Even though these positions will generally respond to arguments for incompatibility between IBE and Bayesianism, it is not the focus of this review to summarize those incompatibility arguments in detail. It is also not a goal of this review to discuss criticisms either on Bayesianism or IBE in their own right, even though there are plenty of issues. They will be treated only insofar as they are relevant for understanding compatibilist proposals.

Approach and outline

This review also identifies the main conceptual distinctions that are useful for mapping the arguments for the compatibility of IBE and Bayesianism. This section briefly outlines the conceptual distinctions that are used to structure the literature review.

For example, Henderson suggests a general division into "constraint-based compatibilism" and "emergent compatibilism" [@Henderson2017, p.249-250]. The constraint-based strategy investigates where explanatory considerations and virtues can play a role in determining the probabilities to be used in the Bayesian framework. In this sense, they constrain those probabilities. Possibilities include using explanatory considerations for likelihood judgments, "assessing the priors", "determining relevant evidence" and "guiding hypothesis construction" [@Lipton2004a, p.119].

Emergent compatibilism instead suggests an alternative strategy, namely to show how the Bayesian framework itself already favors better explanations based on inherent properties. For example, rather than fine-tuning priors based on explanatory considerations, one may point out that Bayesian model selection already penalizes more complex hypotheses (simplicity being an explanatory virtue) [@Henderson2017]. The amount of fine-tuning that is required to make the general Bayesian framework fit the given data, is arguably a native Bayesian measure of explanatory simplicity, where less fine-tuning indicates better explanations [@Henderson2014].

Another relevant distinction is between subjective and objective Bayesians. Subjective Bayesians commit to personalism - the interpretation of probabilities as subjective degrees of belief - but do not impose further requirements on the values of the probabilities. The influential view that IBE is a heuristic that people and scientists use to approximate the outcomes of Bayesian probabilistic reasoning, is compatible both with subjective and objective Bayesianism. Others point out that explanationists can reasonably make the normative claim that the inferences scientists make using IBE can lead to knowledge [@Bird2017, p.104-111]. But then compatibilism with IBE requires some form of objective, rather than subjective Bayesianism, which places stronger constraints on the probabilities that should be used in Bayesian conditionalization [@Weisberg2009]. For example, Poston argues that more explanatory theories should receive higher priors than ones that are not based on objective considerations, in order to avoid inductive scepticism [@Poston2014, p.174-179].

Yet another strategy may be to nevertheless redefine the focus of IBE, but without making it trivial. Cabrera for example argues that whereas Bayesianism assigns credences to hypotheses, IBE tells us when to accept a hypothesis [@Cabrera2017]. In that sense, IBE is compatible with Bayesianism because it is strictly supplementary.

We thus see that various conceptual distinctions are at play in the compatibilist debate. Arguments on the compatibility of Bayesianism with IBE depend on one's specific version of Bayesianism (e.g. subjective, objective, or even without personalism), as well as on one's account of IBE (e.g. that it merely approximates Bayesian results, or with the strong epistemic claim that IBE may lead to true knowledge). An additional factor may be one's understanding of compatibility: are we satisfied when both theories do not contradict each other, or do we minimally require a more tight and ideally synergetic integration?

In order to discuss compatibility between IBE and Bayesianism, we need a minimal description of a generally agreed upon core of both approaches. We thus start with discussing the basic accounts of IBE and Bayesianism as they come forward in the selected papers. Once we have some clarity on the content of both frameworks, we will shortly review the motivation behind attempts to make them compatible, before proceeding to the summaries of the identified positions. Practically all compatibility proposals fall in what Henderson called the 'constraint-based' approach, so this concept is too broad to specify the review structure. Instead, the papers will be organized by the concepts that emerged from the review as the main distinctive categories of compatibilist proposals, starting with a subjective version of the heuristic view, and then moving along a spectrum towards objective or even "super-objective" Bayesians. We will end with the "outsiders," notably Henderson's emergent compatibilism and Cabrera, because these are more agnostic about the particular strand of Bayesianism involved.

Background

Inference to the best explanation and explanatory virtues

The phrase "inference to the best explanation" was coined by Harman to indicate a type of non-deductive reasoning that includes, and is more general than, enumerative induction [@Harman1965]. Harman argued that many inferences that scientists make, as well as people in their daily lives, cannot straightforwardly be seen in terms of enumerative induction, whereas cases of enumerative induction can easily be framed in terms of IBE. When a scientist infers a hypothesis about quarks based on experimental results, this inference does not follow from the scientists regularly observing quarks and then making a claim about a regularity concerning the future observations of quarks. It is rather the case that the postulation of quarks is the best available explanation for the observed results, and that's why the existence of quarks is inferred. Lipton called this a "vertical inference" and points out that the existing models of non-deductive reasoning are too permissive, in the sense that they cannot explain, to stick with our example, why we infer quarks and not "blurbs", or quarks and "blurbs" [cf. @Lipton2000, p.186-187]. Compared to other theories, IBE is theoretically promising for two reasons.

Firstly, insofar as inference is guided by what would explain the observed phenomena best, not any hypothesis is permissible, but only the ones that actually explain the relevant phenomena. With a circularity that Lipton holds to be innocent, explanation can really be said to guide inference in IBE [@Lipton2000, p.186; @Lipton2001]. In this sense, IBE "is explanationism par excellence" [@Bird2017, p.98]. This requires, of course, an account of explanation. We will limit ourselves to a short overview of explanatory virtues below, since this review is not about what exactly constitutes an explanation.

Secondly, the fact that the "best" explanation is chosen already indicates the contrastive nature of IBE. IBE "discriminates between different hypotheses all of which would explain the evidence, since the model only sanctions an inference to the hypothesis which would best explain it" [@Lipton2000, p.187]. IBE is thus an account of why we choose one possible explanation over the others, rather than an absolute measure of inductive confirmation.

Proponents of inference to the best explanation look to the history of science for examples where scientists describe how they inferred some hypothesis because it had clear explanatory advantages over competing hypotheses. Charles Darwin, for example, came up with evolution theory because it explained a wider array of known facts better than the competing theory of special creationism, and Copernicus considered his theory to be more harmonic than the competing theory of Ptolemy [cf. @Henderson2017, p.250].

This type of inference is guided by so-called explanatory virtues, that reflect several methodological rules-of-thumb that scientists apply in practice to distinguish the explanatory quality of theories. The explanatory appeal of IBE thus comes from "its agreement with common sense and paradigm instances of good scientific inference" [@Weisberg2009, p.129]. A clear example is that scientists prefer a simpler theory over a more complex one, if both explain some observation to the same degree (cf. Ockham's Razor). A theory that successfully explains more types of phenomena (scope) is also preferable. A theory with the virtue of unification brings together phenomena that were previously not thought to be related. Other theories are better because they are able to provide detailed information about the mechanisms underlying some phenomenon [@Lipton2000, p.6].

Even though there is plenty of debate concerning the specifics of these explanatory virtues, at least there is a general agreement that these are indeed virtues and that they play a role in the practice of scientific reasoning. Henderson provides the following minimal description of what all IBE accounts share:

It is often suggested that a theory provides the best explanation if it possesses the optimal combination of explanatory virtues in relation to the phenomena, where explanatory virtues include considerations such as simplicity, unification, scope, and fruitfulness. [@Henderson2014, p.689-690]

We should further notice that within the literature terms such as 'abduction' and IBE can be easily confused. Bayesianism largely leaves aside abductive reasoning in the sense of generating new hypotheses, and specifically concerns IBE as the procedure of inferring the best out of a set of possible hypotheses.

With respect to the latter point, IBE may be taken both as a description of how scientists reason and as a model for justifying a given inference. We will focus on IBE as a descriptive account of reasoning, but even then we can distinguish purely descriptive and normative claims of IBE. In Harman's definition, performing IBE implies that "one infers, from the fact that a certain hypothesis would explain the evidence, to the truth of that hypothesis" [@Harman1965, p.89, my emphasis]. This is quite a strong claim that can be contested. For example, a very exotic explanation may explain some evidence very well, but may be very unlikely to be true. Modern accounts of IBE thus usually also require that the best explanation is sufficiently plausible. Nevertheless, strong accounts of IBE are still relevant. For example, in scientific practice one makes new inferences based on the results of previous inferences. To achieve compatibilism, we may thus ask how the Bayesian framework accommodates the use of acquired (background) knowledge for making new inferences. Bird in particular addresses the issue that within Bayesian conditionalization anything that we really know has P(E)=1. If we then involve (background) knowledge as evidence in Bayesian conditionalization, we end up with an extremely strict criterion where only that of which we are absolutely certain counts as knowledge. We should avoid that a Bayesian account of IBE enforces such a strict criterion, because then barely anything may be called knowledge, which amounts to a form of scepticism. In the probabilistic terms of Bayesianism, we would also have to commit to the absurd conclusion that for any piece of knowledge we would be willing to bet our lives on its truth (there is no chance it is wrong, after all).

Conversely, accounts of IBE describe how scientist infer a certain hypothesis to be true, while simultaneously not requiring it to be absolutely true. Or phrased differently, scientists are comfortable inferring a hypothesis even though other competing hypotheses may still have some plausibility as well [@Bird2017].

In any case, for papers on compatibilism it is worth it to consider whether IBE involves claims about the knowledge and truth of the inferred hypothesis, because the more normative the claims of IBE, the more we can expect IBE to conflict with the normative framework of Bayesianism. A normative account of IBE would claim that when scientists apply IBE, they reason as they ought to and have a legitimate claim to truthful knowledge. To resolve such a conflict between two normative frameworks, it is almost certain that extra constraints on Bayesianism are necessary to align it with the normative claims of IBE, i.e. it almost certainly requires a stronger form of objective Bayesianism [@Bird2017, cf. p.98]. We revisit this matter below.

Relevant versions of Bayesianism

This review often refers to "Bayesianism," which may mean very different things in different contexts. This section clarifies what the relevant notions of Bayesianism are with respect to IBE.

The absolute minimum that Bayesians agree on can be called probabilism, namely the idea that our evaluations of hypotheses can be expressed in terms of probabilities. This is almost always complemented with personalism, namely the idea that these probabilities express degrees of belief (also called credences). These are usually interpreted as a willingness to accept a set of bets on the truth of propositions. Insofar as the credences of some agent are rational (e.g. not be susceptible to Dutch books where the agent makes bets that guarantee a loss), the credences have to satisfy the axioms of probability theory. As a normative account of when inferences are rational, Bayesian personalism thus requires probabilism, which can be called a coherence postulate [@Henderson2014, p.690]. A last element that is common to all accounts of Bayesianism is some variety of conditionalization. Conditionalization describes a procedure for updating an agents credences if new evidence arises, in such a way that the updated credences are still coherent according to the axioms of probability theory. These three elements together can be called "standard Bayesianism" [@Bird2017, 97].

It should be noted that standard Bayesianism is clearly a normative account, because it prescribes how inference ought to proceed in order to be rational. This is relevant because this may make compatibility with more normative claims of IBE harder. It is not common to see Bayesianism as a descriptive account of how people actually think, so this position will be largely disregarded in what follows. The standard account of Bayesianism does however have what Bird calls a "descriptive kernel," namely the personalism that takes these probabilities to be descriptions of partial beliefs [@Bird2017, p.98]. A clear difference with more traditional philosophical epistemology is that the Bayesian probabilistic approach to belief allows partial beliefs to be rational, whereas philosophy traditionally deals with full beliefs. Henderson points to a similar difference that is at play between IBE and Bayesianism. The Bayesian talks of degrees of belief whereas "IBE tells us what hypothesis to select for belief (or perhaps acceptance)" [@Henderson2014, p. 689]. In order for the compatibilism debate to make any sense, it should at least be allowed that IBE and Bayesianism concern similar epistemic attitudes. In other words, Henderson points out that if one absolutely insists that IBE is "wedded to belief rather than degrees of belief", then there may be nothing to discuss [@Henderson2014, p.689].

The most common form of conditionalization, called strict conditionalization, is done by application of Bayes' theorem.

$$P_{new}(h) = P(h|e) = \frac{P(e|h)P(h)}{P(e)}$$

Where $P(e|h)$ is the "likelihood" of a hypothesis, $P(h)$ the prior probability of the hypothesis, $P(e)$ the prior probability of the evidence, and $P(h|e)$ the posterior probability, i.e. the credence in hypothesis h after taking into account evidence e. The result will be the updated new belief in hypothesis h, $P_{new}(h)$. We will consider later how these components may correlate with explanatory considerations. Note that $P(e|h)$ reads as the probability of e given hypothesis h, but that we speak of the likelihood of the hypothesis instead, because a hypothesis is more likely if it assigns a high probability to the observed evidence [@Easwaran2011a, cf. p.322].

There are other forms of conditionalization, such as Jeffrey conditionalization [@Jeffrey1965]. Whereas strict conditionalization assumes that the evidence is unambigiously observed, Jeffrey conditionalization allows updating beliefs when there is uncertainty in the observation of evidence.

A last relevant distinction is between subjective and objective Bayesians. For subjective Bayesians the "standard Bayesianism" from above summarizes the constraints on credences in order for them to be rational. This version of Bayesianism is called "subjective" because it allows rational inter-subjective disagreements if agents start with different prior credences [@Weisberg2009, p.127]. (Obviously, people may agree in practice in any account, but subjective Bayesian justifies subjectivism and thus involves a form of relativism). Objective Bayesians instead put more severe restrictions on the probabilities that can be used in conditionalization for the resulting credences to still count as rational. For example, the principle of indifference states that we should equally distribute degrees of belief over the available hypotheses if we do not have relevant evidence with a bearing on those hypotheses. Jaynes' Maximum Entropy Principle prescribes that a rational prior distribution should account for the given facts and only the given facts, and otherwise maximize the entropy of the prior distribution [e.g. @Fiorillo2012]. A strong account of objective Bayesianism does not justify inter-subjective disagreement insofar as the available background beliefs (or knowledge) admit only one "rationally acceptible" probability function [@Henderson2014, p.690]. In practice subjective Bayesians also allow some constraints on probabilities and that objective Bayesians allow multiple rationally acceptable probability functions as well.

Motivation for compatibilism

Most papers on the compatibilism between IBE and Bayesianism start by pointing out that IBE and Bayesianism "are our two most prominent theories of scientific inference" [@Weisberg2009, p.125]. If both of these theories can be successfully applied to known cases of scientific inference, then this suggests that they should be broadly compatible [@Henderson2017, p.250].

Henderson points out that studies of scientific inference can be roughly divided in those that focus on the confirmatory aspects of scientific theory, whereas others consider the more explanatory aspects. However, the specific relation between confirmatory and explanatory aspects remains largely unclear. Bringing them together, we should ask how on the one hand "scientific theory evaluation should be driven primarily by how well confirmed various rival theories are in relation to the evidence" whereas on the other hand "scientific arguments for one theory over another often appeal more or less explicitly to explanatory advantages of that theory" [@Henderson2017, p.249-250].

The Bayesian framework natively expresses confirmation with respect to evidence and IBE natively expresses the importance of those explanatory advantages. Their compatibilism would therefore shed light on the connection between these two aspects and would be a push towards a more unified theory of scientific inference.

If IBE and Bayesianism are indeed compatible, we would ideally end up with two mutually reinforced theories. So what's there to gain for Bayesianism and what's to gain for IBE?

From the perspective of IBE, "part of the purpose of giving a Bayesian model of IBE is to elucidate the epistemic significance of explanatory considerations by connecting them to an established account of evidential support" [@Henderson2017, p.256-257]. IBE provides an intuitive and relatable account how scientists reason, but at the same time explanatory virtues are notoriously imprecise. If the relevance of explanatory virtues could somehow successfully be expressed in the precise, mathematical terms of the Bayesian framework, this would be strong support for IBE.

Conversely, the Bayesian framework is indeed very precise, but its mathematical precision comes at a cost. For example, where do the prior probabilities come from? Many proposals for compatibilism use the explanationism of IBE to "fill in the gaps" of Bayesianism, for example by using explanatory considerations for determining priors [@Weisberg2009, p.136]. The Bayesian framework provides a coherent approach for rationally updating (degrees of) beliefs. But, as Bird points out, the Bayesian framework does not answer several surrounding epistemological questions, such as at what point it is rational to count a proposition as evidence. In short, given rational preliminaries, Bayesian conditionalization guarantees a rational outcome. But the Bayesian framework cannot itself justify its preliminaries [@Bird2017, p.107]. Poston additionally points out that explanationist may help Bayesians "bring on board the concepts of law, cause, and explanation" [@Poston2014, p.167]. Explanationism may aid Bayesianism on these points.

Challenges to compatibilism

Arguments for compatibilism generally respond to well-known arguments against compatibilism. We have already identified that nobody cares about compatibility at all costs; compatibility is rather desired to be meaningful in the sense that none of the two approaches is made trivial, and ideally also in the sense that the strengths of both accounts benefit the other. The first main challenge is thus non-triviality (usually for IBE, because Bayesianism is more specified).

Almost without exception, the selected papers feel obliged to address Van Fraassen's "boosting" argument against compatibility. Several authors summarize the general gist of this argument without diving too much in details [e.g. @Poston2014, p.168]. If IBE is to be made compatible with Bayesianism, then Van Fraassen holds that 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.

Theoretically speaking, compatibilism could instead also be achieved by introducing a more qualitative notion of belief into Bayesianism, especially for those like Weisberg who hold that "IBE is a claim about what we ought to believe" [@Weisberg2009, p.127]. Weisberg however also points out that such an approach has been largely rejected by Bayesians and in any case, it is not the common approach. So we grant Van Fraassen that indeed IBE has to be cast in terms of degrees of belief.

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, if IBE is not irrelevant, then it would be an incoherent account of inference and for example susceptible to Dutch books where agents would irrationally engage in bets they think are fair, but that they are guaranteed to lose money on [@Poston2014, p.168-169].

The heuristic answer (Lipton, Poston)

A main proposal for making IBE and Bayesianism compatible is to claim that "inference to the best explanation (IBE) is a way of realizing Bayesian inference" [@Poston2014, p.167]. This position is commonly referred to as the heuristic view, because all versions of this view consider IBE to be an approximate realization of Bayesian inference, which is in turn acknowledged as a rational and normative account of inductive inference.

Poston identifies two main types of this view [@Poston2014, p.167]: Firstly, there is a weak heuristic view where explanatory considerations are merely "crutches for those who are not smart enough to do probability theory on the fly." In this weak view, the heuristic is a handy aid but is in principle dispensible if we can find the right probabilities to plug into the Bayesian computations. Secondly, one may argue for the stronger view that IBE as a heuristic is required to correctly realize Bayesian inference. Here one rather wishes to show that "explanatory considerations are required for good (i.e., induction-friendly) probability functions" [@Poston2014, p.168].

Lipton could be seen as a main representative of the first weaker heuristic view in the sense that he does not explicitly argues that IBE as a heuristic is required [@Poston2014, cf. 171-172]. Poston is a main representative of the second stronger heuristic view and already employs a relatively moderate form of objective Bayesianism.

Lipton and the loveliest explanation

Lipton has been a main defender of IBE and many other compatibilists positions refer to and argue with his position. In the second edition of his book Inference to the Best Explanation, Lipton dedicated a new chapter to the topic of Bayesian abduction, which is not surprising due to the popularity of the Bayesian framework and its proximity to the concerns of IBE [@Lipton2004a]. Lipton compares possible strategies for countering Van Fraassen's boosting argument that states that Bayesianism is rational and coherent, whereas IBE is incoherent because it deviates from Bayesian conditionalization.

One may argue that Bayesianism is not entirely right, and point to canonical problems such as the problem of old evidence or the problem of new theories [@Lipton2004a, 105; @Easwaran2011a]. One may also point out that Bayesianism fails to account for the ways people actually reason, including all their fallacies and deviations from the Bayesian rational results. So even though Bayesianism may be a correct normative account of inference, this opens the door for IBE as a descriptive account of how people actually reason [@Lipton2004a, 105]. Another way out may be to admit that indeed Bayesianism places constraints on combinations of degrees of beliefs, but that this does not imply a commitment to any particular belief. What Lipton means is that, when we let explanation guide our inferences with IBE and we end up with a "boosted" posterior, this does not necessarily mean that we violated Bayes' theorem. Instead, we may preserve the consistency that Bayes' theorem prescribes by giving that explanatory hypothesis a higher prior. So in this response, explanatory considerations guide our inferences, but this does not necessarily have to conflict with the probabilistic consistency of Bayes' theorem [@Lipton2004a, p.106]. This is already a form of compatibilism in the sense that IBE and Bayesianism may exist next to each other. We established, however, that a good account of compatibilism should show how both theories relate and ideally complement each other. Lipton's own proposal indeed considers IBE and Bayesianism to be complementary:

Bayesian conditionalization can indeed be an engine of inference, but it is run in part on explanationist tracks. That is, explanatory considerations may play an important role in the actual mechanism by which inquirers 'realize' Bayesian reasoning. [@Lipton2004a, p.107]

In other words, IBE is "a heuristic we need because we are not very good at making the probabilistic calculations directly" [@Lipton2004a, 107].

Lipton is not after compatibilism with a "thin" notion of IBE, where "best" equals "most probable" and where IBE only provides minor additions where Bayesianism is silent (e.g. the discovery of hypotheses). Instead, Lipton wants to make Bayesianism compatible with his "thick" notion of the most lovely explanation. Lipton's main thesis is that scientists infer the most lovely explanation, which is the explanation that adds the most understanding, and that in doing so they are ipso facto generally also choosing the most likely explanation [@Lipton2000, p.6; @Lipton2004a, 107].

This heuristic answer circumvents Van Fraassen's argument that IBE must be wrong because Bayesianism is right: IBE is now a way of doing Bayesianism! However, this comes with the great burden of showing that being guided by the explanatory considerations of IBE indeed approximates or tracks the normative probabilistic outcomes of Bayesianism.

An obvious challenge to Lipton's view are the many cases where people reach conclusions based on explanatory considerations that do not agree with Bayesian outcomes. A clear example is the case of Linda the Bank Teller [@Lipton2004a, cf. p.110]. Examples like these may be taken as an argument for incompatibility. If IBE is descriptively right in describing how people actually reason (including their mistakes), and Bayesianism descriptively wrong, what's the common ground between the two?

Lipton is more optimistic than for example Tversky and Kahneman, who more or less conclude that we are bad Bayesians by relying on heuristics. Lipton instead points out that 1) fallacies are atypical, and if our heuristics would never track likeliness it becomes hard to imagine how we deal with our complex environment on a daily basis; 2) that we can always point out the fallacies to those making them, after which the correctness of the Bayesian solution can be recognized [@Lipton2004a, p.111].

Lipton admits that his notion of loveliness does not neatly map onto any one element of Bayes' theorem. However, because we cannot follow the Bayesian recipe in our daily lives, he suggests three ways in which explanatory consideration may "lubricate the Bayesian mechanism" [@Lipton2004a, p.114-116].

Firstly, explanatory considerations may help in determining the likelihood, which is needed to move from prior to posterior. Perhaps loveliness is an indication of the likelihood $P(E|H)$, which would in turn influence the posterior probability of a hypothesis. In any case, loveliness does not exclusively map onto likelihood, as lovely explanations also need to have sufficient (prior) plausibility.

Secondly, explanatory considerations may help in determining the priors. If loveliness indeed plays a role in conditionalization, then it also has a role in determining priors "in large part because today's priors are usually yesterday's posteriors" [@Lipton2004a, p.115]. The same holds for the prior over the evidence, insofar as this may be the result of a previous conditionalization. The prior over the evidence can be interpreted as how surprising it would be to observe evidence E, given the current beliefs. Additionally, explanatory virtues such as simplicity may directly play a role in determining the prior probability of hypotheses.

Thirdly, explanatory considerations may play a role in determining what evidence is relevant to take into consideration. Bayesianism is silent in this regard. In IBE, we may instead "come to see that a datum is epistemically relevant to a hypothesis precisely by seeing that the hypothesis would explain it" [@Lipton2004a, p.116].

To summarize, Lipton claims that our heuristics do not replace Bayesian reasoning, but in the general case help us realize its constraints, which we do ultimately recognize as normative for our reasoning. As a heuristic for Bayesianism, lovely explanations should be a guide for likely explanations, but Lipton put this forward as a line of research more than that he has conclusively shown this. Noticing that reasoning fallacies are an exception is barely a satisfying argument and we can easily show that we do not need fallacies for loveliness and likeliness to diverge. If we have a sick patient, a doctor may provide the extremely likely explanation: "the patient is ill". In Lipton's view you would think such an explanation receives a low likelihood because it is very unattractive, since it adds no understanding whatsoever. It is however a highly likely explanation not because it's lovely, but because it is so unspecific. Lipton has however provided a great kickstarter for the compatibility debate and many proposals work out one of his suggestions in more detail.

Poston's stronger heuristic view against inductive scepticism

Poston's heuristic view addresses Van Fraassen's "boosting" argument by denying that explanatory hypotheses receive a probability boost. Instead, in his strong heuristic view "explanatory considerations are reflected in the probability function itself" [@Poston2014, p.169].

More specifically, the idea is that several explanatory considerations of IBE are correlated with the prior probability and likelihood from Bayes' rule.

The difference with Lipton's view is quite subtle. Lipton argues that IBE is an approximate way of doing Bayesian computations, but leaves open the theoretical possibility that Bayesianism would suffice, if our practical (cognitive) limitations were removed. Poston instead argues, on theoretical grounds, that Bayesianism needs IBE, because otherwise it will suffer from "inductive skepticism" in cases where that is not appropriate [@Poston2014, p.174].

As a case study, Poston provides an illustrative coin flip example [@Poston2014, p.174-175]. Imagine we have a coin with a perfect bias for heads and a second fair coin. We pick one of the two at random and set out to do an experiment where we flip it ten times. Now imagine that our evidence is that the first seven successive results are heads. The best explanation for this evidence is obviously that we randomly selected the biased coin (H1), but in principle it could have been sheer luck that a fair coin (H2) came up heads seven times in a row.

Now, Bayesian updating supports inductivism in the sense that after each coin flip where the coin came up heads, the confidence in H1 increases, i.e. we would assign a higher posterior credence after each successive heads. In contrast, a sceptical position would maintain that each successive throw is independent of the previous one and that the probability of seven consecutive heads is fixed at $0.5^7$. Another way of expressing this scepticism is by noting that "each sequence of outcomes is treated as equiprobable" [@Poston2014, p.177]. Note that this is one possible application of the objective Bayesian principle of indifference: each possible outcome (i.e. sequence) receives equal prior probability. However, a well known issue with the principle of indifference is that it is inconsistent, because the same evidence can receive different probabilities depending on how you define hypothesis space, as exemplified by Bertrand's Paradox [@Easwaran2011a, p.326-327]. In the coin example, if we instead equally distribute prior credences over the proportion of heads rather than the specific sequence, we have inductivism [cf. @Poston2014, p.177].

Bird gives a similar example, but now we know the coin is biased, we just don't know if it's towards heads or tails. If our hypothesis space consists of each possible outcome (i.e. sequence), we assign each hypothesis 1/1024 credence. The probabilities of various proportions of heads and tails can be computed by adding up the chances of relevant sequences. Then we have $P(10 heads)=0.001$. If instead, we make hypothesis about the fairness of the coin in terms of proportion (0, 0.1, 0.2, 0.3 ... 1), each hypothesis has prior credence 1/11. Then we have $P(10 heads)=0.09$ [@Bird2017, p.101-102]. The same principle yields different credences for the same event and is thus not so indifferent after all.

The crux of Poston's argument is that when "we think that there is a physical process that yields a non-epistemic, objective chance" then we should go with the proportion interpretation, "because the relevant non-epistemic, objective chance is explanatory with respect to the proportions but not the sequences." [@Poston2014, 178]. This means that according to Poston, it is an explanatory consideration that determines when inductivism or scepticism is appropriate. For the correct choice of priors, in this objective Bayesian approach, we thus need explanatory considerations extraneous to Bayesianism to allow for non-sceptical inductive inferences.

Compared to Lipton's heuristic view, we still see that IBE serves as a heuristic for the actual Bayesian computation. The difference is that according to Poston, "inductive inferences are justified only if there are explanatory appeals to objective chances in the background" and, more generally, if they are "part of a good explanatory system which beats relevant competitors" [@Poston2014, p.180-181]. Poston cites a clear example from McGrew [cf. @Poston2014, p.179]: if you encounter someone in a poker game who has amazing hands each time for half an hour straight, you strongly suspect he's a cheat beyond reasonable doubt, even though you are not able to precisely assess your degrees of belief either for $P(E)$ or $P(E|H)$.

Poston thus proposes a heuristic account for objective Bayesianism where explanatory considerations impose constraints on the prior such that a explanatory outcome is the result, even when we do not fill in the precise probabilities.

The objective Bayesian's answer (Weisberg)

Is Van Fraassen's boosting argument neutralized that easily? Weisberg argues that Van Fraassen does show a severe conflict, specifically between IBE and what he calls "Subjectivist Conditionalization", i.e. what we called strict conditionalization without any further constraints. We saw that Lipton's heuristic view assumes that loveliness tracks likeliness, but this is a big claim that requires more support. Weisberg is more pessimistic on this front by accepting that "IBE tracks explanatory goodness or beliefs about explanatory goodness, while Subjectivist Conditionalization tracks prior conditional degree of belief" [@Weisberg2009, p.142].

I think Weisberg chose this somewhat uncommon phrasing "prior conditional degree of belief" to emphasize that, given the same evidence to two agents with different prior degrees of belief, conditionalization will yield two different results that are nevertheless rational. That is, subjective conditionalization allows for rational inter-subjective disagreement [@Weisberg2009, cf. 127]. The general tenet of Weisberg's argument is that we should take Van Fraassen's argument more seriously, but should then realize it is specifically a problem of subjective Bayesianism.

The heuristic view on compatibility is unsatisfactory to Weisberg because it entails a demotion of IBE, in at least two senses: IBE becomes inapplicable in cases where we desire it to be applicable, and IBE suffers from a subjectivism where it is permissible that explanatory goodness and likeliness diverge.

When subjective conditionalization is the sole source of normative claims, like in the subjective heuristic view, IBE has no role to play where Bayesianism is silent. This contradicts at least Lipton's hope that IBE may supplement Bayesianism at points where it is silent, for example in the context of discovering hypotheses [cf. @Lipton2004a, p.116-117]. In order to supplement Bayesianism, Lipton should also allow IBE to be applicable in cases where Bayesianism is silent, for example when relevant prior beliefs are lacking. This in turn implies that IBE is not just a heuristic for approximating the results of subjectivist conditionalization, and has at least some normative force of its own [@Weisberg2009, p.134]. Bird likewise points out that Lipton seems to commit to normative claims of IBE without knowing it [@Bird2017, p. 105]. This type of argument emerged as a common approach of objective Bayesians: they are defending the more normative claims of IBE against the heuristic approach to compatibility.

The fact that the point of subjectivism is problematic to Weisberg also shows that he holds a more normative view of IBE where IBE not only describes how we actually think, as was the case in the subjective heuristic view, but also how we ought to think. This is a serious point to defend: the main appeal of IBE is that known paradigmatic historical examples of great scientific inferences were guided by explanatory considerations, and moreover that IBE generally agrees with common sense [@Weisberg2009, p.129].

An objective Bayesian approach is then promising for several reasons. The success of "compatibilism depends crucially on a tight correlation between an explanation's superiority and its prior conditional degree of belief" [@Weisberg2009, p.130]. If the heuristic approach cannot easily bring together loveliness and likeliness, we need to impose further constraints on the prior and likelihood in order to ensure that they are in agreement. By having stronger constraints guided by explanatory considerations, IBE has a larger part to play. If objective considerations play a role in determining the prior (rather than it being strictly subjective degrees of belief), the possibility of rational inter-subjective disagreement is also mitigated (for those who consider this to be a problem). This provides more grounds to call cases where Bayesian conditionalization diverges from explanatory considerations bad inferences.

For example, when we find our house with the front door open, with a broken lock, a mess in the living room, and valuables missing, then the hypothesis with explanatory superiority is that a burglar broke in and stole your valuables. We may however also hypothesize that one burglar broke the lock, only to find another burglar that entered through the window. As they fight over the valuables, a policeman shows up but fails to catch them. Ashamed of his failure, he decides to take the valuables himself and disappears. We all agree that the first explanation is better, even if we encounter someone who holds the strong prior belief that the second hypothesis is more likely. And yet, when we disagree with that person, the subjective Bayesian has to call this a rational inter-subjective disagreement, because the difference is explained by the person's high prior subjective belief in the bad explanation. The answer that such excessive situations are uncommon (a la Lipton) is not very satisfying, because it does not exclude that this situation is possible. The point is that we must constrain such unreasonable probabilities in order to align Bayesian probabilities with the explanatory goodness of IBE.

The objective Bayesian's solution is quite straightforward: priors are not just any subjective degrees of belief, but can be called reasonable or unreasonable by more objective standards, using methods such as the principle of indifference and entropy maximization, that take into account what we know about all the relevant facts. Defining prior distributions as conditioned upon a shared background of the total available evidence mitigates the downsides of subjectivism. Explanationism then has a normative role to play in determining the prior, so that the posterior cannot be dominated by unreasonable priors but instead is sensitive to how the given "evidence bears on a hypothesis" [@Weisberg2009, p.137]. Compatibility between IBE and Bayesianism can thus be integrated into the objective Bayesian's project of constraining the prior distribution. Weisberg's additional requirement for genuine compatibilism then is that for a set of background assumptions B, we should have $P(H|E \land B) > P(H'|E \land B)$ iff H is a better explanation (in IBE terms) against the background assumptions and the evidence.

Unfortunately, despite the strength of Weisberg's theoretical arguments, more concrete applications that satisfy this requirement are missing and remain future work. A challenge that is particularly hard to overcome to this objective approach is that objective principles, such as the principle of indifference, do not necessarily prescribe a unique probability distribution, as for example Bertrand's Paradox shows. This is because the outcomes that you evenly distribute credence over are partly determined by how you formulate the problem at hand.

Note that Poston is a proponent of the heuristic view and a proponent of (objective) constraints on priors. Weisberg would likely argue Poston took half measures. Based on what we've reviewed, Weisberg's argument would firstly be that the explanatory constraints are indeed necessary, but that he needs to drop subjective conditionalization altogether to avoid Van Fraassen's criticism. Secondly, he would tell Poston to drop the heuristic view unless he wants to deny IBE its normative appeal. But denying IBE's normative appeal is not desirable, as it would be unclear on what basis IBE needs to constrain Bayesianism for compatibility.

The 'super-objective' Bayesian (Bird)

We have seen that Weisberg defends a stronger account of IBE, where we allow it to also have normative power. This required a move towards objective Bayesianism according to him. Bird's position on compatibility is interesting because he arguably defends an even stronger claim of IBE, to the point where his Bayesianism needs to be "super-objective" according to his own words. His criticism on the heuristic view largely follows along the same lines as Weisberg's, so it will not be repeated here. Bird also argues that the heuristic view may satisfy the subjective Bayesian but not the objective one. But interestingly, when everything points to Bird defending a more typical objective Bayesian account, he then proceeds to argue that even objective Bayesianism is not strict enough for compatibility with IBE. This is not because the objective Bayesian would be dissatisfied; Bird's point is rather that even on the objective account, proponents of IBE would not be satisfied! This is because Bird holds that the "explanationist thinks that on some occasions IBE can give us knowledge" [@Bird2017, p.104].

We thus see the reappearance of Harman's original account of IBE where the best explanation is held to be true once inferred. We saw that Henderson raised the following issue, but left it aside: is IBE committed to full beliefs and thus incompatible with Bayesian partial beliefs? This issue reappears in Bird's position, because if indeed IBE want to maintain the possibility of reaching full belief, knowledge, even truth, then "the explanationist cannot regard standard Bayesianism as a complete normative epistemology" [@Bird2017, p.104]. If IBE has this strong normative role, priors and likelihoods that are guided by explanatory considerations should count as better than those that are not. And Bird sharply distinguishes "knowledge" from "highly probable". But why are "regular" accounts of objective Bayesianism not constrained enough?

In order to account for knowledge with strict conditionalization, Bird argues we would have to include it in the evidence term with $P(E)=1$, for example when we have acquired knowledge through IBE and build on it for new inferences. This has an awkward consequence if we maintain personalism: a degree of belief of 1 in proposition 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" [@Bird2017, p.105]. However, Bird rightly points out that we would not be willing to make such bets on most knowledge we have about contingent propositions (and reasonably so). Bayesian conditionalization then seems to lead to the sceptical conclusion that we apparently do not know most things that we call knowledge after all, as they then cannot serve as evidence in conditionalization. This is also a problem for the Bayesian, as this would make strict conditionalization largely inapplicable [@Bird2017, p.105-106].

An obvious counter would be to suggest Jeffrey conditionalization, which avoids situations where $P(E)=1$. But this does not address Bird's main argument. Bird requires knowledge to have $P(E)=1$ as an evidence term. If we go along with that, then Jeffrey conditionalization is not a relevant updating rule at all, because it maintains uncertainty in the evidence term, meaning that competing hypotheses that are inconsistent with what we know to be true, always remain an epistemic possibility with $p>0$. If we really know something to be true, this should however exclude the alternatives, so Bird does not consider Jeffrey conditionalization as a viable option to include knowledge in Bayesianism [@Bird2017, p.108].

Objective Bayesianism is however also not the answer insofar as it is still committed to personalism. Bird sees a way forward by using strict conditionalization but dropping personalism. Instead, Bird reinterprets Bayesian probabilities as objective plausibilities, rather than subjective credences.

Objective Bayesianism also still suffers from another issue with respect to IBE and knowledge. In real-life scenarios IBE may select amongst multiple hypotheses that are compatible with the evidence. Even the objective Bayesian approach will assign credences to multiple hypotheses and thus assign a credence $p<1$ to the hypothesis that IBE infers to be true [@Bird2017, p. 109-111]. Hence, we again have the issue that Bayesian explanationist inference cannot lead to knowledge, unlike IBE in Bird's view.

To solve this issue, the a priori constraints of objective Bayesianism concerning the distribution of prior probabilities are not sufficient, but need to be complemented with explanatory considerations that can at best have de facto reliability. This, combined with interpreting probabilities as "plausabilities" instead of credences, is what Bird calls super-objective Bayesianism [@Bird2017, p.112]. Knowledge can unproblematically have a plausibility of 1 without the problematic semantics of credences in terms of bets.

The question then is how IBE helps filling in the elements of Bayes theorem, i.e. how do scientists then assess plausibility? Analogical reasoning is one interesting example for assessing the prior plausibility of a hypothesis independently from its relation to evidence [@Bird2017, p. 113-114]. The constraint that priors are explanatorily plausible does indeed seem to prevent the implausible prior in Weisberg's burglary example. The plausibility of the likelihood and its increase after considering evidence could in turn be assessed by comparison with known successful paradigmatic theories. This however, makes the assumption that those theories are already truth-conductive and that similarity with those theories also implies truthfulness [@Bird2017, p.114-115].

I would say these assumptions are strong and lead to circularity. And yet, if you drop them it becomes unclear to me how "objective" Bird's super-objective Bayesianism is. Even if we speak of plausibilities, it is rather unclear why this does not also retain a subjective element insofar as something appears to be plausible to someone. I also think another (mildly) relativistic element is maintained, because the truth-conductiveness of IBE then relies largely on scientific consensus about paradigm instances of scientific inference.

Whereas Bird claims that proponents of IBE do not have to be satisfied with the objective Bayesian's answer, I would counter that I suspect many Bayesians will not be happy with Bird's approach. All other authors reviewed here consider personalism to be part of an absolutely minimal account of Bayesianism, so they will probably wonder what's left of Bayesianism in Bird's proposal. Because the issues Bird raises dominantly depend on his strong definition of IBE, I think further discussion should focus on whether this account is perhaps too strong.

Reconceptualizing IBE (Cabrera)

As we have seen, one of the requirements for a compatibilist position is to ensure we do not achieve compatibility by making IBE trivial. However, one recent out-of-the-box proposal from Cabrera fleshes out a way to reformulate IBE such that it is compatible with Bayesianism, without making IBE trivial (that's the claim, at least).

Cabrera's main observation is that some of IBE's explanatory virtues have informational rather than confirmatory content, meaning that they are indications that a hypothesis may add understanding, without necessarily making that hypothesis more probable.

For example, one of the explanatory virtues that Lipton mentions is "precision," which means that a hypothesis $H_1$ that is more specific is a better explanation than hypothesis $H_2$ that is more vague, other things being equal. If loveliness implies likeliness, we should have $Pr(H_1|E) > Pr(H_2|E)$. Now have $H_1$ be "A and B differ by amount X" and $H_2$ be "A and B differ by some amount." Because $H_1$ is more precise it should be the better explanation and thus also more likely, according to Lipton. But $H_1$ is a specific case of $H_2$, so by definition we have $Pr(H_1) \leq Pr(H_2)$ [@Cabrera2017, p.1255]. At the same time, $H_1$ can still be a better hypothesis to investigate, because it is more informative. Science does not only care about finding the most probable hypotheses, since many probable hypotheses are trivial [@Cabrera2017, p.1265]. Cabrera shows this distinction between information and confirmation for several other explanatory virtues.

A main problem for Lipton's view is that he has to convincingly show that loveliness guides likeliness. The objective Bayesians try to align loveliness and likeliness by imposing more constraints on Bayesianism. But if many explanatory virtues do not even play any confirmatory role, this effort will lead to an incomplete account at best.

One approach could be to only retain the explanatory virtues that have some confirmational import, but then we would arguably fail the requirement that IBE should not be made trivial. IBE would be "a kind of un-precisified Bayesianism" [@Cabrera2017, p.1261]. Cabrera's solution is instead to keep IBE intact, but to reconceptualize IBE's nature as a theory of hypothesis acceptance rather than inference [@Cabrera2017, p.1261-1262].

Even if some explanatory virtues do not make a hypothesis more likely, they are still signposts of good theories that merit potential acceptance, in the sense that they merit committing research to them. So, "informational virtues are still virtues" [@Cabrera2017, p.1261]. Acceptance in this context does not imply the full belief in the truth of a hypothesis, but that a hypothesis is worth pursuing (for which plausibility is one factor, but not the only one) [@Cabrera2017, p.1265]. Cabrera defends IBE's appeal as a theory of acceptance in this specific sense. Interestingly, this happens to be an area where Bayesianism is silent.

So if "Bayesianism is silent on the issue of acceptance, and probable truth does not exhaust the conditions under which a hypothesis is worthy of commitment, there is some substantive, interesting role for the explanatory virtues to play in our account of scientific practice" [@Cabrera2017, 1269].

To conclude, Cabrera sees IBE as compatible with Bayesianism because it is strictly supplementary to Bayesianism. I would point out that supplementary is not the same as complementary: it is clear how Bayesians benefit from IBE, but less clear how IBE benefits from Bayesians. Nevertheless, IBE does have an independent and substantive role to play in scientific practice under Cabrera's view. It is, however, clear to me that it will not be substantive enough for those objective Bayesians that are wedded to IBE as a normative account of scientific inference, in particular Bird who insists that IBE can lead to true knowledge.

The emergent approach (Hitchcock, Henderson)

Hitchcock

We have seen that Lipton hopes to establish that loveliness is a guide to likeliness. For example, if a theory is more simple than another (an explanatory virtue) we should also generally judge this theory more likely. But Hitchcock asks how such a claim may be justified, because this seems to imply the a priori belief (and I would say, metaphysical belief) that the world is in fact simple. Hitchcock deems Lipton's response that this problem is another version of Hume's problem of induction to be unsatisfying. Instead, Hitchcock points out that a more down to earth justification is readily available: "Hypotheses that are simpler in the sense of having fewer free parameters tend to fare better when they are used to make predictions about new data" [@Hitchcock2007, p.5]. In other words, the mathematical theory of model selection, as well as ample accompanying practical experience, show that simpler theories are less likely to overfit. This does not assume that the world affords lovely explanations nor does it solve the problem of induction. It only assumes the data is being drawn from a "stable underlying probability distribution" [@Hitchcock2007, p.6]. The crux: if this explains our preference for simple hypotheses "then our inductive preferences do not commit us to assigning higher prior probabilities to simpler hypotheses" (cf. Van Fraassen's "boosting") because our preference is strictly instrumental [@Hitchcock2007, p.6].

I take Hitchcock to describe the preliminaries of the project that Henderson develops, namely an account of compatibilism solely in pragmatic Bayesian terms, called "emergent compatibilism".

Henderson's emergent compatibilism

The previous proposals are all various shades of what Henderson calls "constraint-based" compatibilism and they all have their particular strengths and weaknesses. Henderson puts forward yet another approach to compatibilism, called "emergent compatibilism," that avoids several weaknesses of both subjective and objective Bayesian compatibilists. Henderson explores in what sense Bayesians can accept constraints on their probabilities strictly on their own terms (so non-IBE considerations), in such a way that they end up favouring the theories that are more explanatory. If this is successful, a new form of compatibilism is provided where IBE emerges from Bayesianism [@Henderson2014, p.698-699].

Henderson starts by distinguishing IBE on specific hypotheses and IBE on general theories, and treats them separately, which none of the other proposals did so explicitly. In most of the discussed examples, such as the examples with dice, we directly compare two hypotheses that are fully specified, as opposed to a general theory with free parameters. In the specific case, Henderson's approach to IBE is very straightforward. She argues that some hypothesis h is a better explanation than h' when it makes the data more expected than h'. So in her view, the most lovely explanation makes the data more understandable by rendering it as expected, better so than a competing hypothesis [@Henderson2014, p.700]. But she clearly deviates from other compatibilists by claiming that this does "need not be seen as governed by explanatory considerations" [@Henderson2014, p.700]. Instead, in order to "determine which hypothesis provides the best explanation, we combine the judgments about the expectedness of the outcome, given the hypotheses, with judgments about how plausible the hypotheses are" [@Henderson2014, p.700]. If IBE requires no extra explanatory considerations, the above statement is directly compatible with what Bayes' theorem does. This approach may, I think, be susceptible to the criticism that it uses a too thin version of IBE.

Henderson's more interesting claim is however that the disproportionate focus on specific hypotheses has obscured how Bayesianism may account for the more interesting aspects of IBE.

Real scientific theories, as has been pointed out in the philosophy of science (e.g. Lakatos or Kuhn), contain several general core claims that are often supplemented with auxiliary hypothesis in order to account for phenomena. General theories thus have certain free parameters that allow them to be fitted in order to be applicable and make predictions in a specific situation [@Henderson2014, 702]. A theory with many free parameters may succeed in making a wide array of phenomena highly expected by adding several auxiliary hypotheses. A proponent of IBE would obviously prefer a simpler theory that makes less (ad-hoc) assumptions. Henderson proposal is ...

that in such situations, one theory provides a lovelier explanation than another the more the core claims of the theory by themselves allow the phenomenon to be explained and the less the explanation is reliant on auxiliary hypotheses [@Henderson2014, p.702].

The paradigmatic example of Copernicus versus Ptolemy shows how intuitive that proposal is. Henderson summarizes that "the Copernican theory provides the lovelier explanation because accepting its core claim that the sun is at the centre of the planetary orbits allows the phenomenon of retrograde motion to be explained with less dependence on auxiliary hypotheses" [@Henderson2014, p.707].

Several explanatory virtues follow from this proposal quite naturally. For example, it is widely accepted that a theory with fewer auxiliary hypotheses is a simpler and better theory. The virtue of unification also follows when many different phenomena can be explained by the core of the theory, rather than by auxiliary hypotheses. There is a natural trade-off between the two for determining if a theory has the right scope. Explaining more phenomena increases a theory's unification until it starts relying too much on auxiliary hypotheses [@Henderson2014, p.707-708].

Henderson's main claim is that Bayesianism can capture this account of IBE solely on its own terms. IBE's considerations about the initial plausibility are taken into account in the prior. But Henderson's most substantive claim is that the likelihood term in Bayes' theorem takes into account the 'loveliness' of explanations in IBE. Recall that Lipton denied that we could identify loveliness with any specific element of Bayes' theorem. I think Henderson would reply that the loveliness component only becomes clear when considering general theories, whereas Lipton focuses on highly specific examples where we choose between two structured hypotheses.

The likelihood of a general theory is not determined by the likelihood of its best fitting model, but rather is the average likelihood over all its possible models. So given a general theory T, given data or evidence D, and $\theta$ being the parameters of T, we have:

$$Pr(D|T) = \int{}^{}Pr(D|\theta,T)Pr(\theta|T)\delta\theta $$

Again, note that the likelihood of the theory is expressed in terms of how expected it makes the data. Because we are interested in the likelihood of the general theory, we marginalize out the auxiliary hypotheses. In a follow-up paper Henderson makes it more explicit that she is assuming a hierarchical Bayesian model instead of one large hypothesis space, where frameworks occupy a higher level than fitted theories [@Henderson2017, p.251].

Of course, this likelihood is also influenced by the prior over the parameters $\theta$, but often Bayesians make the "natural" assumption that it's most reasonable to equally distribute credences over the parameter space. This small point should not be overlooked: it means that Henderson does not use explicit explanatory considerations for constraining the prior, but follows common Bayesian practice. It is an open point for discussion when we may call a prior "natural" and if we can do this solely in Bayesian terms. Ted Poston for example responds that the "naturalness" of a prior would have to come from explanatory considerations (recall, Bayesianism needs IBE). In that sense, Poston argues that his own position is compatible with Henderson's emergent compatibilism [@Poston2014, p.173-174].

Assuming "natural" priors, it can be shown that Bayesians would favor Copernicus' theory for the same reasons as explanationists, because Copernicus' theory relies less on free parameters to fit the data and instead has a better core theory. Having more free parameters is penalized in the likelihood of the general theory. If core claims account for the data, then all fitted models have a good fit with the data as well, so the prior credence is distributed over models that fit well on average. If, on the other hand, fitting the data heavily relies on a few particular parameter settings of the general theory (i.e. auxiliary hypotheses), then all the possible models that do not fit the data bring down the average likelihood of the general theory [@Henderson2014, p.709-710]. In this way, the explanatory virtue of simplicity may be said to emerge from Bayesianism. Bayesianism would provide the quantitative underpinnings of explanatory features, that IBE describes qualitatively on a higher level of description [@Henderson2014, p.711].

An advantage of Henderson's view is that Bayesian model selection theory is well established. Whereas many compatibilist positions discuss toy examples, Henderson is able to provide a convincing and in-depth case study of theory selection, alluding to explanatory considerations but not relying on them [@Henderson2017]. One may counter that Henderson has a weak account of IBE and retrofits "loveliness" to fit Bayesian terms. In a way, her proposal slightly resembles Lipton's in that it dominantly sees IBE as a qualitative description of what Bayesianism shows normatively. However, because Henderson does not use IBE to constrain any probabilities, she does not need IBE to be normative to justify those constraints [cf. @Henderson2014, p.711].

Conclusion

This literature review addresses the research question: "What are the current main proposals for making Bayesianism compatible with IBE, without making IBE trivial?"

Even though there is a variety of concerns and proposed solutions at play in the reviewed papers, we can identify some trends. Almost all proposals turned to objective Bayesianism for compatibility with IBE. This is not very surprising, because most proposals hold that explanatory considerations should constrain the probabilities used in Bayesianism in order to align them with explanatory IBE outcomes. This in turn clearly assumes that IBE needs to have normative status, otherwise the imposed constraints would be arbitrary. We have also seen that the more normative the account of IBE is, the more Bayesianism needs to be constrained, up to the point where perhaps not even objective Bayesianism suffices anymore. An interesting takeaway here is that it would be worthwhile to study how normative IBE should be taken to be, because this largely determines the specific type of constraint-based compatibilism.

On the other end of the spectrum, the only subjective Bayesian account we reviewed is Lipton's. His account also fits the relation between the normativity of IBE and the "objectivity" of the Bayesian account, because criticasters notice that Lipton's account demotes IBE to a derivative position with respect to the normative account Bayesianism. Lipton optimistically assumes, more or less, that loveliness guides likelihood, but this remains too vague. On what grounds can we be convinced that they align? When we start constraining Bayesian probabilities from a subjective Bayesian's perspective, there are many different ways to achieve alignment with IBE, but unless we have a normative account of IBE the chosen approach will almost certainly be arbitrary and ad-hoc. Hence, we saw the massive turn to objective Bayesianism, even for others that hold the heuristic view (e.g. Poston).

I did find that proposals based on objective Bayesianism lacked convincing and detailed case studies that deal with real-world examples. Despite all theoretical critique, Lipton's view does have an intuitive appeal for describing historical cases of scientific inference.

Henderson's position is more agnostic about whether a subjective of objective Bayesian account should be used. In both cases, her point is clear: we can account for loveliness with Bayesianism without explicitly using explanatory considerations as constraints. Her argument does however assume that we use "natural" priors and mentions a uniform prior distribution, following the principle of indifference as an example. In this respect, her proposal still seems slightly biased towards some strand of objective Bayesianism. On the other hand, objective Bayesianism is not strictly required because she is not committed to a normative account of IBE.

The fact that Henderson does not impose any explanatory constraint on Bayesianism highlights an interesting contrast with the other constraint-based accounts. If IBE emerges from Bayesianism then they are compatible, whereas it can often seem as if constraint-based solutions are trying their best to make them compatible. But then we can also ask: at what cost?

In my assessment, Bird's type of proposal is too strong if it requires us to drop core elements of Bayesianism. From this review it did however become clear that a form of objective Bayesianism is a more likely candidate. I think Poston's and Henderson's proposals can be allies and that future work should try to combine both accounts. Henderson requires the assumption of natural priors, and Poston's account precisely offers a way to select appropriate priors by taking into account explanatory considerations. Henderson's account is also compatible with Poston's moderate strand of objective Bayesianism. However, if Cabrera is right that many explanatory virtues do not have confirmational import, then they cannot "emerge" from the confirmational framework of Bayesianism. Future work on emergent compatibilism should therefore show that reducing IBE to its confirmational aspects is not too harmful.

Reflection on learning objectives

This project had the following learning objectives:

  • Explain the core ideas of IBE and the relevant aspects of Bayesianism
  • Differentiate the main academic positions arguing for the compatibilism of IBE and Bayesianism
  • Summarize the main arguments for these respective positions
  • Evaluate the merits of and challenges for these respective positions and arguments.

In my opinion I succeeded in reaching these learning objectives. I have considered and explained IBE and Bayesianism in their own rights based on relevant literature. With respect to compatibility positions, I can now distinguish the relevant IBE and Bayesianism variants that are involved. I've been able to identify, summarize, and classify a variety of compatibility positions along two axes that largely run parallel: how normative IBE is, and how objective the Bayesian account. Special attention has been paid to the justification of these positions and the type of argument used for support. I have indicated how these positions address challenges raised by other positions or by Van Fraassen's boosting argument.

One point of feedback on the initial project plan was that it would be desirable to provide my own critical assessment. I have therefore included critical remarks in the paper summaries, and in my conclusion I argued for what I think is a viable route for future research.

The most positive aspect of the project, for me personally, was that it gave me a deeper understanding of Bayesianism. In the AI bachelor we get used to applying Bayes' theorem to solve exercises. But this is a form of application with little attention to the interpretation of the used probabilities. For example, I never sufficiently realized that strict conditionalization is not the only way to go. Neither was I aware that various forms of objective Bayesianism are possible. The matter of compatibilism with IBE is particularly interesting in this regard, because it challenges the Bayesian framework to really justify the used probabilities and their interpretation.

The most negative aspect of the project is that I struggled to keep up with my intended planning. For the majority of my other courses I also did project work, and I found it difficult to parallelize the work for this project. The reason for that is that I require longer deeper focus for reading papers and taking detailed notes, whereas I have no problem doing coding for one project for about an hour, then coding on another for the next. My other projects also were collaborations with others, which made them harder to compartmentalize. An additional factor that I struggled with is that I found it harder to maintain deep focus for close reading while working from home due to COVID-19.

These are some personal points of improvement concerning the process, but all in all I'm very pleased with what I have learned through this project.