IMPRINT/ACT-R: integration of a task network modeling architecture with a cognitive architecture and its application to human error modeling

摘要

This paper describes ongoing efforts to integrate IMPRINT (IMproved Performance Research INtegration Tool), a task network modeling architecture with ACT-R (Adaptive Character of Thought -Rational), a hybrid cognitive architecture. IMPRINT consists of a set of automated aids to conduct human performance analyses built on top of the Micro Saint task network modeling environment. ACT-R combines a goal-directed production system with a subsymbolic activation calculus that tunes itself to the structure of the environment using Bayesian learning mechanisms. Because ACT-R and IMPRINT were targeted at different behavioral levels, they perfectly complement each other. IMPRINT is focused on the task level, how high-level functions break down into smaller-scale tasks and the logic by which those tasks follow each other to accomplish those functions. ACT-R is targeted at the atomic level of thought, the individual cognitive, perceptual and motor acts that take place at the subsecond level. Goals in ACT-R correspond directly to tasks in IMPRINT, providing a natural integration level. Certain tasks in an IMPRINT task network can be implemented as ACT-R models, combining the cognitive accuracy of a cognitive architecture with the tractability and ease of design of task networks. A hybrid IMPRINT/ACT-R model works as follows. The IMPRINT model specifies the network of tasks and includes the definition of how higher-order functions are decomposed into tasks and the logic by which these tasks are composed together. For certain tasks, IMPRINT sends to ACT-R over a Component Object Model (COM) link the state of variables providing a detailed description of that task. ACT-R then creates a goal corresponding to that task, with the components of the goal set to the description of the task. The ACTR model for that goal is then run, producing detailed cognitive predictions including latency of the run, whether an error occurred, etc. Those results are then passed back over the same COM link to IMPRINT which uses them as parameters of the task to advance the task network model. We describe an application of this hybrid modeling to the prediction of human errors that lead to runway incursions. Finally, we discuss future extensions of our work, including the use of a standardized High Level Architecture (HLA) link to handle communications between IMPRINT and ACT-R and the extension of the task parameters exchanged to include workload predictions.