David's Written Questions

Intelligent Systems Program

Comprehensive Examination

for David Joslin

Committee: Martha Pollack, Rich Thomason, Kurt VanLehn

This is an open book examination, so you may consult---and reference---any sources from the literature. Note, however, that you are not permitted to consult any person about this examination, with the exception that you may ask your committee members clarification questions. For Questions 1 and 2, please consult Rich Thomason; for Questions 3 and 4, consult Kurt VanLehn; and for Questions 5 and 6, consult Martha Pollack.

Each answer should be no more than 5 pages of single-spaced text in 12pt font, using 1" margins all around.

Describe the theory of action presented by R. Moore in the paper ``A formal theory of knowledge and action''. Evaluate it as a logical theory: how well motivated is it by intuitions about action and how well does it help to account for valid inference about action? Evaluate it from the standpoint of knowledge representation: could the theory play a useful role in computationally useful representations of action? In your answer, you may want to compare Moore's theory to other accounts of action.

All of the readings by philosophers on your philosophy of mind list discuss AI, but do so in a way that requires only the broadest and most superficial understanding of the subject. (Dennett alludes to cases in which he believes that contemporary results from AI have philosophical significance, but does not develop details in the paper you cite.) Can an argument be made for why a philosopher of mind might profit from keeping up to date with results in AI, and to thoroughly understand the details of some of the research? In your answer, try to develop a general account of how AI and philosophy of mind are related, and show how developments in AI might be significant for philosophy of mind.

Does learning change the problem space of the planners described in references [2], [3], and [4] of the section on learning in your reading list (i.e., Fikes et al.; Minton; and Minton et al.) If so, discuss the combinatorial implications. Are there some general principles of learning that you can state?

A typical AI system has three levels: (1) the host language (e.g., Lisp), (2) a representation language and interpreter (e.g., Prodigy, Klone, etc.) and (3) the domain-specific knowledge. Of course, the host language runs on some hardware, which may itself have multiple layers of interpretation. Newell is often seen as advocating a simple two-level view of mind: (1) a single fixed cognitive architecture constant across individuals and (adult) lifespan, (2) acquired knowledge. In other words, the cognitive architecture is like the mind's hardware (with all its levels of interpretation) and the acquired knowledge is like mind's software (with all its levels of interpretation). Critically examine this position. First, does it really reflect what Newell says in his book, "Unified Theories of Cognition"? Second, are there any cogent, plausible positions other than the two-level view? Third, is there any evidence that supports one of these positions at the expense of the others?

. Recently, there has been a great deal of work aimed at developing techniques to improve the efficiency of automatic plan generation. Describe these techniques, and discuss the relationships amongst them.

It has always been assumed that AI plan generation techniques would ultimately be applied to "real-world" problems. Given this, how can one justify the continued use of micro-worlds, like the blocks world, as a focus of planning research? Do you think that the use of the blocks world has fostered or hindered the progress of planning research? Comment on what you believe to be the prospects for applying AI plan generation techniques to real-world problems. To what extent has this already happened, and what hurdles must be overcome before it happens to a greater extent?

R. Michael Young < myoung+@pitt.edu>