The Algorithmic Mind: A Study of Inference in Action Thomas Icard Abstract: What is the nature of human inference? How does it work, why does it work that way, and how might we like it to work? I advance a framework for answering these questions in tandem, with a rich interplay between normative and descriptive considerations. Specifically, I explore a view of inference based on the idea of probabilistic sampling, which is supported by behavioral psychological data and appears to be neurally plausible, and which also engenders a philosophically novel and appealing view of what grounds subjective probability. I then discuss this view in the context of the Bayesian program in cognitive science, proposing a methodology of boundedly rational analysis, which particularly exemplifies the normative/descriptive interplay. By taking resource bounds seriously, we can improve and augment the more standard rational analysis strategy. This helps us focus efforts to understand how minds in fact infer, and in turn allows sharpening normative questions about how minds ought to infer. Against this background I explore the phenomenon of metareasoning, which arises naturally when discussing resource-limited but representationally sophisticated agents, but which has not been explored in the context of probabilistic approaches to the mind. I propose an analysis of metareasoning in terms of the value of information, and explore the consequences of this view for how we should think about inference. The focus of this dissertation is on implemented (or at least implementable) models of agents, and the role of inference in guiding and driving intelligent action for real, resource-bounded agents. Consequently, a number of the suggestions and claims made are supported by simulation studies.