Plausibilistic coherence

Why should coherence be an epistemic desideratum? One response is that coherence is truth-conducive: mutually coherent propositions are more likely to be true, ceteris paribus, than mutually incoherent ones. But some sets of propositions are more coherent, while others are less so. How could coherence be measured? Probabilistic measures of coherence exist; some are identical to probabilistic measures of confirmation, while others are extensions of such measures. Probabilistic measures of coherence are fine when applicable, but many situations are so information-poor that the requisite probabilities cannot be determined. To measure coherence in these cognitively impoverished situations, this article proposes that the discussion be broadened to include plausibilistic measures of coherence. It shows how plausibilistic measures of coherence can be defined using plausibilistic measures of confirmation. It then illustrates how plausibilisic coherence can be measured in situations where probabilistic coherence cannot be determined. The coherence values obtained through the use of plausibilistic measures are often, though not always, comparable. The article also shows that coherence can be instantiated on different levels, one of which permits connections to inductive strength and deductive validity.     

Against Probabilistic Measures of Coherence

It is shown that the probabilistic theories of coherence proposed up to now produce a number of counter-intuitive results. The last section provides some reasons for believing that no probabilistic measure will ever be able to adequately capture coherence. First, there can be no function whose arguments are nothing but tuples of probabilities, and which assigns different values to pairs of propositions {A, B} and {A, C} if A implies both B and C, or their negations, and if P(B)=P(C). But such sets may indeed differ in their degree of coherence. Second, coherence is sensitive to explanatory relations between the propositions in question. Explanation, however, can hardly be captured solely in terms of probability.     

Reliability conducive measures of coherence

A measure of coherence is said to be truth conducive if and only if a higher degree of coherence (as measured) results in a higher likelihood of truth. Recent impossibility results strongly indicate that there are no (non-trivial) probabilistic coherence measures that are truth conducive. Indeed, this holds even if truth conduciveness is understood in a weak ceteris paribus sense (Bovens & Hartmann, 2003, Bayesian epistemology. New York, Oxford: Oxford University Press; Olsson, 2005, Against coherence: Truth probability and justification. Oxford: Oxford University Press). This raises the problem of how coherence could nonetheless be an epistemically important property. Our proposal is that coherence may be linked in a certain way to reliability. We define a measure of coherence to be reliability conducive if and only if a higher degree of coherence (as measured) results in a higher probability that the information sources are reliable. Restricting ourselves to the most basic case, we investigate which coherence measures in the literature are reliability conducive. It turns out that, while a number of measures fail to be reliability conducive, except possibly in a trivial and uninteresting sense, Shogenji’s measure and several measures generated by Douven and Meijs’s recipe are notable exceptions to this rule.     

Coherence measures and inference to the best explanation

This paper considers an application of work on probabilistic measures of coherence to inference to the best explanation (IBE). Rather than considering information reported from different sources, as is usually the case when discussing coherence measures, the approach adopted here is to use a coherence measure to rank competing explanations in terms of their coherence with a piece of evidence. By adopting such an approach IBE can be made more precise and so a major objection to this mode of reasoning can be addressed. Advantages of the coherence-based approach are pointed out by comparing it with several other ways to characterize ‘best explanation’ and showing that it takes into account their insights while overcoming some of their problems. The consequences of adopting this approach for IBE are discussed in the context of recent discussions about the relationship between IBE and Bayesianism.     

Probabilistic measures of coherence: from adequacy constraints towards pluralism

The debate on probabilistic measures of coherence flourishes for about 15 years now. Initiated by papers that have been published around the turn of the millennium, many different proposals have since then been put forward. This contribution is partly devoted to a reassessment of extant coherence measures. Focusing on a small number of reasonable adequacy constraints I show that (i) there can be no coherence measure that satisfies all constraints, and that (ii) subsets of these adequacy constraints motivate two different classes of coherence measures. These classes do not coincide with the common distinction between coherence as mutual support and coherence as relative set-theoretic overlap. Finally, I put forward arguments to the effect that for each such class of coherence measures there is an outstanding measure that outperforms all other extant proposals. One of these measures has recently been put forward in the literature, while the other one is based on a novel probabilistic measure of confirmation.     

Why There Cannot be a Single Probabilistic Measure of Coherence

Bayesian Coherence Theory of Justification or, for short, Bayesian Coherentism, is characterized by two theses, viz. (i) that our degree of confidence in the content of a set of propositions is positively affected by the coherence of the set, and (ii) that coherence can be characterized in probabilistic terms. There has been a longstanding question of how to construct a measure of coherence. We will show that Bayesian Coherentism cannot rest on a single measure of coherence, but requires a vector whose components exhaustively characterize the coherence properties of the set. Our degree of confidence in the content of the information set is a function of the reliability of the sources and the components of the coherence vector. The components of this coherence vector are weakly but not strongly separable, which blocks the construction of a single coherence measure.     

A weak symmetry condition for probabilistic measures of confirmation

This paper presents a symmetry condition for probabilistic measures of confirmation which is weaker than commutativity symmetry, disconfirmation commutativity symmetry but also antisymmetry. It is based on the idea that for any value a probabilistic measure of confirmation can assign there is a corresponding case where degrees of confirmation are symmetric. It is shown that a number of prominent confirmation measures such as Carnap’s difference function, Rescher’s measure of confirmation, Gaifman’s confirmation rate and Mortimer’s inverted difference function do not satisfy this condition and instead exhibit a previously unnoticed and rather puzzling behavior in certain cases of disconfirmation. This behavior also carries over to probabilistic measures of information change, causal strength, explanatory power and coherence.     

Measuring Uncertainty

Two types of measures of probabilistic uncertainty are introduced and investigated. Dispersion measures report how diffused the agent’s second-order probability distribution is over the range of first-order probabilities. Robustness measures reflect the extent to which the agent’s assessment of the prior (objective) probability of an event is perturbed by information about whether or not the event actually took place. The properties of both types of measures are investigated. The most obvious type of robustness measure is shown to coincide with one of the major candidates for a dispersion measure, the mean square deviation measure.     

The Impossibility of Coherence

There is an emerging consensus in the literature on probabilistic coherence that such coherence cannot be truth conducive unless the information sources providing the cohering information are individually credible and collectively independent. Furthermore, coherence can at best be truth conducive in a ceteris paribus sense. Bovens and Hartmann have argued that there cannot be any measure of coherence that is truth conducive even in this very weak sense. In this paper, I give an alternative impossibility proof. I provide a relatively detailed comparison of the two results, which turn out to be logically unrelated, and argue that my result answers a question raised by Bovens and Hartmann’s study. Finally, I discuss the epistemological ramifications of these findings and try to make plausible that a shift to an explanatory framework such as Thagard’s is unlikely to turn the impossibility into a possibility.     

A corrective to Bovens and Hartmann’s measure of coherence

Bovens and Hartmann (Bayesian Epistemology, Oxford: Oxford University Press, 2003) propose to analyze coherence as a confidence-boosting property. On the basis of this idea, they construct a new probabilistic theory of coherence. In this paper, I will attempt to show that the resulting measure of coherence clashes with some of the intuitions that motivate it. Also, I will try to show that this clash is not due to the view on coherence as a confidence-boosting property or to the general features of the model that Bovens and Hartmann use to analyze coherence. It will turn out that there is at least one other measure that is similarly based on the concept of a confidence-boosting property, but does not have the same counterintuitive results.     

Coherence,striking agreement,and reliability

Striving for a probabilistic explication of coherence, scholars proposed a distinction between agreement and striking agreement. In this paper I argue that only the former should be considered a genuine concept of coherence. In a second step the relation between coherence and reliability is assessed. I show that it is possible to concur with common intuitions regarding the impact of coherence on reliability in various types of witness scenarios by means of an agreement measure of coherence. Highlighting the need to separate the impact of coherence and specificity on reliability it is finally shown that a recently proposed vindication of the Shogenji measure qua measure of coherence vanishes.     

Measuring coherence

This paper aims to contribute to our understanding of the notion of coherence by explicating in probabilistic terms, step by step, what seem to be our most basic intuitions about that notion, to wit, that coherence is a matter of hanging or fitting together, and that coherence is a matter of degree. A qualitative theory of coherence will serve as a stepping stone to formulate a set of quantitative measures of coherence, each of which seems to capture well the aforementioned intuitions. Subsequently it will be argued that one of those measures does better than the others in light of some more specific intuitions about coherence. This measure will be defended against two seemingly obvious objections.     

A Logic For Inductive Probabilistic Reasoning

Inductive probabilistic reasoning is understood as the application of inference patterns that use statistical background information to assign (subjective) probabilities to single events. The simplest such inference pattern is direct inference: from “70% of As are Bs” and “a is an A” infer that a is a B with probability 0.7. Direct inference is generalized by Jeffrey’s rule and the principle of cross-entropy minimization. To adequately formalize inductive probabilistic reasoning is an interesting topic for artificial intelligence, as an autonomous system acting in a complex environment may have to base its actions on a probabilistic model of its environment, and the probabilities needed to form this model can often be obtained by combining statistical background information with particular observations made, i.e., by inductive probabilistic reasoning. In this paper a formal framework for inductive probabilistic reasoning is developed: syntactically it consists of an extension of the language of first-order predicate logic that allows to express statements about both statistical and subjective probabilities. Semantics for this representation language are developed that give rise to two distinct entailment relations: a relation ⊨ that models strict, probabilistically valid, inferences, and a relation that models inductive probabilistic inferences. The inductive entailment relation is obtained by implementing cross-entropy minimization in a preferred model semantics. A main objective of our approach is to ensure that for both entailment relations complete proof systems exist. This is achieved by allowing probability distributions in our semantic models that use non-standard probability values. A number of results are presented that show that in several important aspects the resulting logic behaves just like a logic based on real-valued probabilities alone.     

Why does coherence appear truth-conducive?

This paper aims to reconcile (i) the intuitively plausible view that a higher degree of coherence among independent pieces of evidence makes the hypothesis they support more probable, and (ii) the negative results in Bayesian epistemology to the effect that there is no probabilistic measure of coherence such that a higher degree of coherence among independent pieces of evidence makes the hypothesis they support more probable. I consider a simple model in which the negative result appears in a stark form: the prior probability of the hypothesis and the individual vertical relations between each piece of evidence and the hypothesis completely determine the conditional probability of the hypothesis given the total evidence, leaving no room for the lateral relation (such as coherence) among the pieces of evidence to play any role. Despite this negative result, the model also reveals that a higher degree of coherence is indirectly associated with a higher conditional probability of the hypothesis because a higher degree of coherence indicates stronger individual supports. This analysis explains why coherence appears truth-conducive but in such a way that it defeats the idea of coherentism since the lateral relation (such as coherence) plays no independent role in the confirmation of the hypothesis. An earlier version of this paper was presented at the workshop Coherence: Interpreting the Impossibility Results held in Lund, Sweden. I would like to thank the participants of the workshop, especially Erik J. Olsson who subsequently sent me written comments. I would also like to thank Ken Akiba for comments on precursors of this paper     

Inference to the best explanation: does it track truth?

In the form of inference known as inference to the best explanation there are various ways to characterise what is meant by the best explanation. This paper considers a number of such characterisations including several based on confirmation measures and several based on coherence measures. The goal is to find a measure which adequately captures what is meant by ‘best’ and which also yields the truth with a high degree of probability. Computer simulations are used to show that the overlap coherence measure achieves this goal, enabling the true explanation to be identified almost as often as an approach which simply selects the most probable explanation. Further advantages to this approach are also considered in the case where there is uncertainty in the prior probability distribution.     

EXPLAINING THE LIMITS OF OLSSON'S IMPOSSIBILITY RESULT

In his groundbreaking book, Against Coherence (2005), Erik Olsson presents an ingenious impossibility theorem that appears to show that there is no informative relationship between probabilistic measures of coherence and higher likelihood of truth. Although Olsson's result provides an important insight into probabilistic models of epistemological coherence, the scope of his negative result is more limited than generally appreciated. The key issue is the role conditional independence conditions play within the witness testimony model Olsson uses to establish his result. Olsson maintains that his witness model yields charitable ceteris paribus conditions for any theory of probabilistic coherence. Not so. In fact, Olsson's model, like Bayesian witness models in general, selects a peculiar class of models that are in no way representative of the range of options available to coherence theorists. Recent positive results suggest that there is a way to develop a formal theory of coherence after all. Further, although Bayesian witness models are not conducive to the truth, they are conducive to reliability.     

A General Measure of Ecological Behavior1

Measurement of ecological behavior across different domains has been troublesome. The present paper argues that the lack of agreement in measuring general ecological behavior may be due to the measurement approach that is commonly used. An ecological behavior measure should be grounded on a probabilistic measurement approach that takes the important features of ecological behavior into consideration. Such a measure was developed in a survey study of 445 members of 2 Swiss transportation associations. Three types of ecological behavior measures were included: a general measure, 3 multiple-item measures, and 3 single-item measures. Results are controlled for social desirability effects. Reliability, internal consistency, and validity scores indicate that a probabilistic measurement approach can measure general ecological behavior accurately and unidimensionally.     

Discovering syntactic deep structure via Bayesian statistics

In the Bayesian framework, a language learner should seek a grammar that explains observed data well and is also a priori probable. This paper proposes such a measure of prior probability. Indeed it develops a full statistical framework for lexicalized syntax. The learner's job is to discover the system of probabilistic transformations (often called lexical redundancy rules) that underlies the patterns of regular and irregular syntactic constructions listed in the lexicon. Specifically, the learner discovers what transformations apply in the language, how often they apply, and in what contexts. It considers simpler systems of transformations to be more probable a priori. Experiments show that the learned transformations are more effective than previous statistical models at predicting the probabilities of lexical entries, especially those for which the learner had no direct evidence.     

An implicit enumeration method for an exact test of weighted kappa

The kappa coefficient is one of the most widely used measures for evaluating the agreement between two raters asked to assign N objects to one of K nominal categories. Weighted versions of kappa enable partial credit to be awarded for near agreement, most notably in the case of ordinal categories. An exact significance test for weighted kappa can be conducted by enumerating all rater agreement tables with the same fixed marginal frequencies as the observed table, and accumulating the probabilities for all tables that produce a weighted kappa index that is greater than or equal to the observed measure. Unfortunately, complete enumeration of all tables is computationally unwieldy for modest values of N and K. We present an implicit enumeration algorithm for conducting an exact test of weighted kappa, which can be applied to tables of non‐trivial size. The algorithm is particularly efficient for ‘good’ to ‘excellent’ values of weighted kappa that typically have very small p‐values. Therefore, our method is beneficial for situations where resampling tests are of limited value because the number of trials needed to estimate the p‐value tends to be large.     

Dynamic Update with Probabilities

Current dynamic-epistemic logics model different types of information change in multi-agent scenarios. We generalize these logics to a probabilistic setting, obtaining a calculus for multi-agent update with three natural slots: prior probability on states, occurrence probabilities in the relevant process taking place, and observation probabilities of events. To match this update mechanism, we present a complete dynamic logic of information change with a probabilistic character. The completeness proof follows a compositional methodology that applies to a much larger class of dynamic-probabilistic logics as well. Finally, we discuss how our basic update rule can be parameterized for different update policies, or learning methods.