A Comprehensive Computational Model of PRIMs Theory for Task-Independent Procedural Learning

Abstract

The human ability to reason about and learn practically any task has been studied for countless years, but to date we still do not truly understand how human learning is task-independent at the computational level. Researchers have theorized that we can account for many human cognitive behaviors if we combine a task-independent set of primitive procedures with a robust, general learning mechanism that compiles them into cognitive skills for various tasks. The PRIMs theory of procedure learning and transfer is a cognitive architecture theory of human learning that shows how a task-independent set of primitive procedures can support learning in any task that is also supported by the underlying architecture. However, its published architecture implementation, Actransfer, focuses on modeling transfer and does not specify all of the computational details of PRIMs theory.

This thesis presents a computationally comprehensive cognitive architecture model of PRIMs theory that I call the PROPs system. I comprehensively define each of the processing steps that PRIMs theory requires and implement these in an agent model using the Soar cognitive architecture. I do this through a methodology for incrementally refining a cognitive architecture model. I use this methodology to extend PRIMs theory and unify it with three-phase learning theory from human performance research, task set theory from psychology and neuroscience, and Soar theory from cognitive architecture research. This achieves several improvements in the model’s ability to replicate human learning behavior.

Among the contributions of this work, I introduce a novel form of primitive processing that explains the origins of the primitive procedures of PRIMs theory and supports procedural learning in an unbounded, dynamic working memory space. I show that this improves the model’s ability to match human power-law learning. I also extend Soar cognitive architecture theory with gradual procedural learning in a manner consistent with Soar’s existing theory and introduce a novel computational approach by which a cognitive architecture model can learn to guide automatic long-term declarative memory retrievals based on working memory contents. I finally introduce a novel computational approach by which a model can guide deliberate retrievals through choice-based decision making.

In my evaluation of the PROPs system, I identify ways in which PRIMs theory for procedural learning might be further unified with neuroscience theory to broaden the model to include declarative learning. I also identify boundaries where PRIMs models can or cannot currently account for types of human cognitive processing when the models are constrained to be fully task-independent and consistent with the surrounding cognitive architecture. This reveals a path for future cognitive architecture research and development.