The Computational Basis of Interactive Skill

Paul P. Maglio
Doctor of Philosophy in Cognitive Science
University of California, San Diego, 1995

Abstract

My claim is that it can be computationally efficient to incorporate redundant actions into skilled behavior. My studies of how people improve at playing the videogame Tetris show that part of getting better means getting faster, which is what standard theories of skill learning predict. However, my data also reveal that sometimes getting better involves doing more backtracking in the task environment rather than doing less. This finding opposes standard views of expertise in which increases in skill are driven by improvements internal to the agent, with external changes reflecting better honed internal reasoning, motor, or recognition processes. By implementing a series of computer models, I discovered that even for a skilled perception model of expertise, backtracking is adaptive because it can help constrain the problem that needs to be solved. In particular, I argue that the perceptual computation required for Tetris is more efficiently done by serial search than by fully parallel pattern recognition. If skilled perception is completely parallel, then there is no way to use external action to facilitate processing because processing happens instantly. But because serial search is in fact more efficient, external actions can play a role in skilled recognition. Hence, I provide a computational reason for skilled players' redundant interactions with the Tetris game.

If you are interested in the details, you can read the document. Download each chapter individually. Below are a series of postscript files---I don't think I'll take the time to turn it into html. The whole thing is 301 pages long, and the total space needed is about 6 megabytes.