Igor Douven

The meta-inductive justification of induction is usually regarded as a social learning strategy. But, pre-theoretically, induction can be justified even for an isolated thinker, incapable of, or unwilling to engage in, any social interactions. This paper presents reinforcement learning as a natural meta-inductive strategy for such isolated thinkers. The meta-inductive reinforcement learner learns to choose an optimal prediction method among the available methods by applying trial-and-error learning. We use computer simulations to show that, under certain plausible assumptions, trial-and-error learning indeed leads the thinker to favor induction over non-inductive predictive methods. Under further plausible assumptions, the meta-inductive thinker is even seen to outperform all object-level methods. The results are shown to hold in both stationary and dynamic environments.

According to a recent proposal, natural concepts are those represented by the cells of an optimally partitioned similarity space, where optimality is defined in terms of various design criteria, including the criterion that natural concepts should be easily learnable. While computational studies have shown that, in contexts of individual learning, natural concepts are indeed more easily learned than nonnatural ones, the proposal also suggested that natural concepts should facilitate social learning, helping communities converge on shared meanings through interaction. We test this hypothesis using neural agent-based models to simulate how communities learn conceptual structures through social interaction. Our simulations demonstrate that natural concepts are learned more quickly and with higher accuracy than nonnatural ones in social contexts, thereby providing further support for the optimal partitioning account.

This paper investigates the reliability of the Suppositional Heuristic (SH) as a means of estimating the probabilities of conditionals, taking the Material Conditional Account (MCA) as background. Using theoretical analysis, simulations, and published data, we find that under ecologically realistic, relevance-laden conditions, conditional probabilities approxi-mate the probabilities of corresponding material conditionals and correlate strongly with them. However, it is also argued that while our findings offer a defensible best reply for the MCA, on a different reading, they rather favor relevance accounts, for which the observed dependence on relevance is built in rather than auxiliary.

This chapter discusses how critics of Inference to the Best Explanation (IBE) complain that the rule has never been properly articulated. Another common complaint is that whatever the precise formulation of the rule, we can do without it for it must be inferior to Bayes’ rule. It is here suggested that IBE is best thought of as a slogan that can be fleshed out in different ways, where different fleshings-out may have different merits and drawbacks, depending on the context of usage. Some reasons are given for believing that, in some contexts, versions of IBE outperform Bayes’ rule in all important respects.

Goldman (2001) asks how novices can trust putative experts when background knowledge is scarce. We develop a reinforcement-learning model, adapted from Barrett, Skyrms, and Mohseni (2019), in which trust arises from experience rather than prior expertise labels. Agents incrementally weight peers who outperform them. Using a large dataset of human probability judgments as inputs, we simulate communities that learn whom to defer to. Both a strictly individual-learning variant and a reputation-sharing variant yield performance-sensitive deference, the latter accelerating convergence. Our results offer an empirically grounded account of how communities identify and trust experts without blind deference.

This paper examines the question of when to terminate an inquiry. We are specifically interested in whether acceptance rules recently discussed in formal epistemology can be used for establishing when to stop inquiry by adopting an answer to the target question. Assuming that, in some of our inquiries, we value accuracy as well as speed, we use computer simulations to compare along these dimensions rules that we derive from the best known acceptance rules. It will be seen that the rules make different speed—accuracy trade-offs. We argue that, because different circumstances may ask for different speed—accuracy trade-offs, there is no single correct answer to the question of when an inquiry can be ended.

According to Inferentialism, a conditional is true precisely if there is a compelling argument for its consequent starting from its antecedent conjoined with contextually given background knowledge. As a direct consequence of this view, the truth of a conditional’s component parts does not guarantee the truth of the conditional, given that it is compatible with there not being an argument from the conditional’s antecedent to its consequent. Thus, Inferentialism invalidates the principle called Centering, according to which the joint truth of φ and ψ guarantees the truth of “If φ, ψ”. This distinguishes Inferentialism from all standard semantics of conditionals, all of which validate Centering. Here, we report the results from a study that registered (i) participants’ truth evaluations of a set of conditionals; (ii) their truth evaluations of the component parts of those conditionals; (iii) their judgements of the strength of the inferential connection between antecedent and consequent for each conditional; and (iv) their metacognitive confidence in the inference strength judgements. The data from (i)–(iii) allowed us to test Centering and to investigate whether possible violations of the principle can be explained in inferentialist terms. The data from (iv) allowed us to make some progress on testing predictions derived from Hypothetical Inferential Theory, the psychological face of Inferentialism. Our findings show that almost 60% of the relevant responses violated Centering; that participants obeyed Centering only when there was a strong inferential link between true antecedent and true consequent; and that the best models were those that involved inference strength and metacognitive confidence, while the Centering model was weak.

Recent work has argued for an ecological perspective on the rationality of updating, showing that non-Bayesian update rules can outperform Bayesian updating in certain environments. However, this work has left unaddressed the question of how to determine which rule to use without prior knowledge of environmental features—a challenge we term the “dependency problem.” We propose a solution that uses reinforcement learning, specifically a multi-armed bandit framework, to enable dynamic rule selection. Computer simulations indicate that the “adaptive updating” this approach results in is able to solve the dependency problem at low cost in that it performs competitively with fixed update rules across varied contexts. While adaptive updating does not universally outperform fixed rules, it offers a viable alternative when the optimal rule for a given context is unknown.

Linguists and philosophers commonly distinguish between semantics and pragmatics, where the former concerns the truth or falsity of linguistic items and the latter concerns aspects of the use of such items that may make them unassertable even when true. Common though the distinction is, there is an ongoing controversy about where exactly the line between semantics and pragmatics is to drawn. We report two experiments meant to investigate empirically whether there is any pre-theoretic distinction that might help settle the debate. The same experiments are meant to shed light on a related question, namely, whether pragmatic aspects of language use pertain only at the level of assertability and not at that of believability. Our results suggest that ordinary people do not reliably distinguish among truth, assertability, or believability. We argue that this has consequences for the methodology of experimental semantics and pragmatics.

Analogical reasoning is a form of non-deductive reasoning that gives special weight to similarity considerations. Here, we pursue an approach to formalizing this type of reasoning that was initiated by Carnap in posthumously published work. In it, Carnap abandoned his long-time project of trying to define inductive and analogical reasoning syntactically and introduced attribute spaces to model the meanings of predicates. While these spaces remain underdeveloped in Carnap’s late work, it is clear that what he envisioned is, or is close to, what are presently known as “conceptual spaces.” We use the conceptual spaces framework as it has been developed over the past two decades to make progress on formally representing analogical reasoning. This will also allow us to address the question of the normative status of analogical reasoning, which was raised by Carnap and others but which has hitherto remained unanswered. Finally, we work out some of the empirical content of our proposal and reanalyze a publicly available dataset to test it.

Similarity spaces are standardly constructed by collecting pairwise similarity judgments and subjecting those to a dimension-reduction technique such as multidimensional scaling or principal component analysis. While this approach can be effective, it has some known downsides, most notably, it tends to be costly and has limited generalizability. Recently, a number of authors have attempted to mitigate these issues through machine learning techniques. For instance, neural networks have been trained on human similarity judgments to infer the spatial representation of unseen stimuli. However, these newer methods are still costly and fail to generalize widely beyond their initial training sets. This paper proposes leveraging prebuilt semantic vector spaces as a cheap alternative to collecting similarity judgments. Our results suggest that some of those spaces can be used to approximate human similarity judgments at low cost and high speed.

In the context of the conceptual spaces framework, it has been argued that a natural concept is represented by a convex region in a similarity space. The convexity requirement has been defended on grounds of cognitive economy: among other benefits, concepts represented by convex regions have been said to be easily learnable, or more easily than concepts represented by nonconvex regions. There is some evidence that concepts in use are represented by regions that are convex, or at least almost so. There is so far no evidence that concepts represented by convex regions are more easily learnable than ones represented by regions that satisfy topological criteria that are somewhat less stringent, most notably that of connectedness. This note presents the outcomes from computational studies carried out on perceptual color space as well as on a shape space for representing various container objects, indicating that convexity indeed facilitates learning more than does connectedness. The studies use the training of neural nets as a model of human learning.

According to a recent proposal, natural concepts are represented in an optimally designed similarity space, adhering to principles a skilled engineer would use for creatures with our perceptual and cognitive capacities. One key principle is that natural concepts should be easily learnable. While evidence exists for parts of this optimal design proposal, there has been no direct evidence linking naturalness to learning until now. This article presents results from a computational study on perceptual color space, demonstrating that naturalness indeed facilitates learning.

We argue that concepts from the social sciences can be as natural as those from physics and chemistry, thereby answering in the positive the question of whether social metaphysics is or can be substantive. The argument takes as a starting point Douven & Gärdenfors’ (Mind & Language, 35, 313–334 2020) optimality account of natural concepts, according to which natural concepts are represented by the cells of an optimally partitioned similarity space. While the account applies straightforwardly to perceptual concepts, it does not apply as straightforwardly to more abstract concepts, such as many scientific concepts. We propose an extension of Douven and Gärdenfors’ optimality account that pertains to scientific concepts as well. The proposal reuses ideas from Lewis’ Best Systems Account of laws of nature, but rather than invoking natural concepts (as Lewis does) defines natural scientific concepts as those that figure in the or a best systematization of the totality of facts. On this proposal, there is nothing especially different between concepts from the natural sciences and those from the social sciences.

In the philosophical literature, the term “abduction” is used in two related but different senses. In both senses, the term refers to some form of explanatory reasoning. However, in the historically first sense, it refers to the place of explanatory reasoning in generating hypotheses, while in the sense in which it is used most frequently in the modern literature it refers to the place of explanatory reasoning in justifying hypotheses. In the latter sense, abduction is also often called “Inference to the Best Explanation.” This entry is exclusively concerned with abduction in the modern sense, although there is a supplement on abduction in the historical sense, which had its origin in the work of Charles Sanders Peirce—see the Supplement: Peirce on Abduction. See also the entry on scientific discovery, in particular the section on discovery as abduction. Most philosophers agree that abduction (in the sense of Inference to the Best Explanation) is a type of inference that is frequently employed, in some form or other, both in everyday and in scientific reasoning. However, the exact form as well as the normative status of abduction are still matters of controversy. This entry contrasts abduction with other types of inference; points at prominent uses of it, both in and outside philosophy; considers various more or less precise statements of it; discusses its normative status; and highlights possible connections between abduction and Bayesian confirmation theory.

According to a much discussed argument, reliabilism is defective for making knowledge too easy to come by. In a recent paper, Weisberg aims to show that this argument relies on a type of reasoning that is rejectable on independent grounds. We argue that the blanket rejection that Weisberg recommends of this type of reasoning is both unwarranted and unwelcome. Drawing on an older discussion in the philosophy of science, we show that placing some relatively modest restrictions on the said type of reasoning suffices to block the anti-reliabilist argument.

It is undeniable that inductive reasoning has brought us much good. At least since Hume, however, philosophers have wondered how to justify our reliance on induction. In important recent work, Schurz points out that philosophers have been wrongly assuming that justifying induction is tantamount to showing induction to be reliable. According to him, to justify our reliance on induction, it is enough to show that induction is optimal. His optimality approach consists of two steps: an analytic argument for meta-induction (that is, induction over inductive methods) and an application of meta-induction to empirical findings concerning the predictive accuracy of our actual inductive practices as compared to various non-inductive methods. The present article agrees with Schurz’s general conclusion but also argues that it leaves an important aspect of our inductive practices unaccounted for, to wit, that these practices have so far been highly successful; to satisfy the optimality requirement, induction only needs to be as successful as random guessing. To explain the success of induction, I turn to insights from social epistemology, also highlighted in Schurz’s work, and make these formally explicit in a framework that combines the Hegselmann–Krause model for opinion dynamics and Carnap’s system of λ rules. Within this framework, I use a standard evolutionary algorithm to show how communities of interacting agents may, through variation and selection, come to consist of highly successful inductive reasoners.