Humean learning (how to learn)

Philosophical Studies https://doi.org/10.1007/s11098-023-02086-3 Humean learning (how to learn) Jeffrey A. Barrett1 Accepted: 18 November 2023 © The Author(s) 2023 Abstract David Hume’s skeptical solution to the problem of induction was grounded in his belief that we learn by means of custom . We consider here how a form of reinforcement learning like custom may allow an agent to learn how to learn in other ways as well. Specifically, an agent may learn by simple reinforcement to adopt new forms of learning that work better than simple reinforcement in the context of specific tasks . We will consider how such a bootstrapping process may lead to a system that includes trial-and-error forms of learning like win-stay/lose-shift, probe and adjust, and simple reinforcement itself together with higher-rationality inferential tools. Keywords Humean learning · The problem of induction · Hume’s skeptical solution · Learning how to learn · Pragmatism 1 Introduction David Hume was skeptical regarding our ability to rationally justify beliefs concerning matters of fact, but he held that we nevertheless routinely learn matters of fact by means of custom. This sort of instinctive learning might be understood as a form of reinforcement where an agent’s dispositions to act are strengthened on success and perhaps weakened on failure in action. While Hume was right to suppose that humans, like other animals, very often learn by means of reinforcement, we also learn in other ways. We consider here how an agent may be led by simple reinforcement to adopt new forms of inductive learning. By such means she may evolve a learning system that includes trial-and-error forms of learning like win-stay/lose-shift, probe and adjust, and simple reinforcement itself together with higher-rationality inferential tools.1 1 See Erev, I., Roth, A. E. (1996), Fudenberg and Levine (1998), Erev and Roth (1998), Bereby-Meyer and Erev (1998), Barrett and Zollman (2009), Skyrms (2010), Huttegger (2017), Cochran and Barrett (2021, * Jeffrey A. Barrett 1 University of California Irvine, Irvine, USA 13 Vol.:(0123456789) J. A. Barrett While simple reinforcement may lead an agent to adopt forms of learning that are more sophisticated or better-suited to the inferential tasks she faces, there is no magic. Even when she is led to adopt a learning rule that has been highly reliable for a particular purpose, it may fail to work well in the future.2 That said, simple reinforcement provides a reliable way of tracking which learning rules have worked well, and insofar as reinforcement on success is in fact psychologically efficacious in tuning our dispositions, this will lead one to evolve more sophisticated forms of learning regardless of whether one is rationally justified in doing so. This will serve the agent well in future action should those rules continue to work. 2 Humean learning Hume believed that we can never have rational justification for our expectations or beliefs regarding matters of fact.3 While one may observe that an event of type A has always been followed by an event of type B, constant conjunction fails to entail any necessary connection. No sequence of past conjunctions, no matter how extensive, provides any reason whatsoever for concluding even that the occurrence of A makes the occurrence of B more likely. To get something like this, one would need to assume that what has happened in the past is a reliable guide to what will happen in the future, but such an assumption begs the question. Even if the past has in some ways been a reliable guide to the future in the past, it need not be in the future. As a result, our experience provides no ultimate justification for any beliefs at all regarding future events. And since the same line of argument applies to conclusions regarding causal relations generally, only by means of which Hume argued can one learn matters of fact, one can have no ultimate justification for believing any matter of fact (1975, 25–39). This poses an immediate problem for rational action. Inasmuch as one cannot infer anything concerning the future from the past, Hume held that one can never have any rational justification for acting one way rather than another. That said, there is an important sense in which he was not at all skeptical regarding the expected efficacy of his actions or his judgments regarding matters of fact more generally. Understanding the position requires some care. Footnote 1 (continued) 2022), Barrett and Gabriel (2022), and Barrett (2023) for descriptions and discussions of a great many alternative forms of learning. Each has potential virtues and vices depending on the learning problem at hand and the resources available to the learner. See Barrett (2023) for an extended discussion of learning how to learn and reflections on how various basic and task-specific forms of learning might self-assemble. 2 In this regard, note that any particular learning algorithm R, no matter how subtle or sophisticated it may be, may routinely fail to provide successful predictions. Consider a world where whenever one learns by R to expect E on the basis of one’s evidence so far ¬E occurs. See Putnam (1963) for a more elaborate version of this argument. 3 He argued for this in both A Treatise of Human Nature (1739–40) and An Enquiry Concerning Human Understanding (1748). Here we will follow the argument of the latter. Regarding learning, we follow his natural propensity account grounded in custom. 13 Humean learning (how to learn) Hume explicitly recognized that he, like everyone else, was in fact firmly committed to a rich collection of beliefs regarding future events and matters of fact. Further, he found that he remained committed to these beliefs even when he knew that he possessed no ultimate justification for believing them. As a result, he was perfectly comfortable using beliefs that he had formed in the context of experience to guide even his most important actions (1975, 42). Hume held that beliefs regarding matters of fact, and expectations regarding the future in particular, were produced from experience by means of custom or habit. Custom, in the sense in which he used the term, is a principle of our psychological nature that acts to produce and adjust propensities when presented with experience. Hume explained that “wherever the repetition of any particular act or operation produces a propensity to renew the same act or operation, without being impelled by any reasoning or process of the understanding ... this propensity is the effect of Custom” (1975, 43). In other words, we learn just as animals do who “by the proper application of rewards and punishments, may be taught any course of action.” The upshot is that, rather than being an activity grounded in reason, the ability to engage in empirical inquiry is one that “we possess in common with beasts” and “is nothing but a species of instinct or mechanical power, that acts in us unknown to ours (...truncated)