Christian Torsell

I.J. Good's ``On the Principle of Total Evidence" (1967) looms large in decision theory and Bayesian epistemology. Good proves that in Savage's (1954) decision theory, a coherent agent always prefers to collect, rather than ignore, free evidence. It is now well known that Good's result was prefigured in an unpublished note by Frank Ramsey (Skyrms 2006). The present paper highlights another early forerunner to Good's argument, appearing in Janina Hosiasson's ``Why do We Prefer Probabilities Relative to Many Data?" (1931), that has been neglected in the literature. Section 1 reviews Good's argument and the problem it was meant to resolve; call this the \textit{value of evidence problem}. Section 2 offers a brief history of the value of evidence problem and provides biographical background to contextualize Hosiasson's contribution. Section 3 explicates the central argument of Hosiasson's paper and considers its relationship to Good's (1967).

Humans and many animals are capable of learning and learning how to learn better. We are concerned here with one way that reinforcement learners might learn how to learn better. In an experiment described by Harlow in (Psychol Rev 56:51–65, 1949) a group of rhesus monkeys learn a new way of learning in the context of a specific type of problem. We will consider how such agents might coevolve a new learning dynamics and new attendant saliences. To this end, we propose a self-tuning dynamics that illustrates one way that a reinforcement learner might acquire forms of learning that are well-suited to context-specific problems.

An agent is inequity averse if their preferences reflect a distaste for material inequality. Inequity aversion is of both scientific and philosophical interest. In behavioral economics, it is central to popular explanations of “anomalous” behavior in resource-division games. In social and political philosophy, it is of interest because it embodies an egalitarian approach to distributive bargaining. We use the indirect evolutionary approach, an extension of standard evolutionary game theory, to consider whether inequity aversion can survive competition with selfishness under natural dynamics like social learning and Darwinian evolution. In particular, we study evolution in populations playing a variant of the Nash demand game, a popular model of bargaining. Unlike related extant work, we assume that individuals play the base game repeatedly within their lifetimes, learning as they do. The central result is that, under a broad range of assumptions about the evolutionary dynamics and the initial population state, inequity aversion goes to fixation in populations comprising both inequity averse and selfish agents. Incorporating learning turns out to be crucial to the evolutionary result. The model thus supports two conclusions. First, egalitarian attitudes can survive and proliferate in competition with selfish ones. Second, in applying the indirect evolutionary approach, integrating intragenerational learning can influence the evolutionary dynamics in important ways.

A persistent problem in the philosophy of induction concerns the variety of analogies among experiences that might guide one’s learning. Given that any observational record will exhibit a multiplicity of regularities, which ones should we attend to in forming empirical beliefs and expectations? Call this the projectibility problem. Nelson Goodman (Fact, fiction, and forecast. 1983) dramatized the problem with his “new riddle of induction,” and David Lewis (Convention: A philosophical study. 1969) recognized the challenges it raises for the possibility of establishing conventions by precedent. Learners in nature face a version of the projectibility problem inasmuch as they must learn which regularities in their environments are useful guides for action and prediction in order to survive. The present paper considers one method by which this problem might be solved: trial-and-error learning. The strategy is to begin with an experiment from comparative psychology in which subjects face a version of the projectibility problem, then to develop a concrete model of a learner that can solve that problem. In the model, the dispositions of three separate modules, or subagents, responsible for the learner’s attentional and behavioral responses coevolve under a simple reinforcement learning dynamics. On simulation, the model reliably learns to attend to the practically relevant regularities in its environment.