Abstract
Two experiments found that people can adapt routine procedural behavior to changing circumstances quite readily. An ACT-R model of the two experiments is the first to understand complete human performance of a routine procedure—including error, error recovery, and adaptation to procedure change—across multiple between-subjects conditions. Results from the behavioral and modeling studies favor an account of human routine procedural memory that uses discrete, hierarchically-organized goals and action representations that are split between procedural and declarative memories. The modular nature of this account of action selection was crucial for generating three behaviors of interest: 1) The types and amounts of errors observed in the behavioral studies, 2) Selecting error recovery actions, and 3) Adapting old procedural knowledge to new but very similar procedures. Additionally, the account included some prepackaged routines for handling error recovery as well as interpretation of instructions and their application to task context. The modular nature of action representation enabled those behaviors because it allowed for sufficient abstraction of action representations to evoke behaviors in novel contexts. These capabilities are fundamentally beyond the realm of accounts of action selection that rely on holistic, distributed representation of learned step co-occurrence associations.
The Dissertation [PDF, 1.6 MB]
Lisp code to run the human experiments [zip archive]
Training Manuals
Post-Experiment Survey
Model Code [zip archive]
Works cited by my dissertation
Defense Slides [PDF]
Last modified 2010.11.18
