ABSTRACT
In chapter 3 of this volume, Mitchell and Thrun focus on one of the key issues in the construction of an autonomous agent—how such an agent can learn to control its behavior through direct interaction with its environment. Although the chapter does not explicitly highlight specific hypotheses, it seems safe to extract from it three underlying hypotheses: one top-level and two subordinate. The actual expression of these hypotheses is mine, but the intent here is to capture the essence of what Mitchell and Thrun intended, and then to use these hypotheses to focus a discussion from the intertwined perspective of our experience with Allen Newell and Soar (which are often hard to disentangle). One thread that will be implicit through much of the material in this discussion is particularly identified with Newell's research philosophy: Working with an integrated architecture means looking deeply at its consequences for new phenomena or capabilities before hypothesizing new mechanisms to account for them, and in the process the architecture will often teach you what you couldn't have anticipated.
