UC Irvine

Issues of projectibility are central to the philosophy of induction. The basic problem goes like this. Inductive learning involves projecting regularities in past experience onto unobserved cases. But any body of experience can be described as exhibiting any number of regularities, and depending on which of these we project, induction will yield different beliefs and predictions. A principled procedure for identifying regularities that provide good guidance for belief and prediction—i.e., projectible ones—has proven elusive. Nevertheless, the impressive track record of human and animal learning suggests that nature has found ways to deliver us assumptions of projectibility that track stable, practically relevant regularities in many contexts. This dissertation uses computational models to show how simple natural adaptive processes might accomplish this critical task. Specifically, the goal is to explain how such assumptions might emerge from trial-and-error learning in kinds of problems that arise frequently in nature. Chapter 1 supplies philosophical background and motivation. Chapter 2 considers projectibility in the context of signaling, building on a game theoretic model due to Lewis (1969). Chapter 3 turns to discrimination learning, modeling the emergence of assumptions of projectibility in a setting based on classic primate learning experiments conducted by Harry Harlow. Chapter 4 concerns the special challenges of projectibility posed by task switching problems. Chapter 5 concludes.