Cognitive Modeling Research
Yes, some people still believe in symbolic models of cognition! My most recent work on cognitive modeling has the following thrusts:
1. Reinforcement Learning
We are in the process of adding reinforcement learning to Soar. This paper is the first publication on the work.
2. Episodic Memory
We are also adding an episodic memory to Soar.
We added activation to Soar's working memory. We built upon the original ACT activation scheme that was then implemented originally in Soar by Ron Chong. Our implementation is completely new and attempts to provide a very efficient implementation so that there is little overhead to use activation. We are initially using the activation for feature selection in the episodic learning work and expect to use it in the reinforcement learning work. Some day we might even use it for forgetting (but not yet).
4. Emotion Modeling
We are delving into adding a model of emotion to Soar. We initially have developed a framework for how emotion fits in Soar and have tried it on a simple task.
5. Inductive Learning
In this work, we've look at inductive learning of concepts using symbolic mechanisms, while still preserving many of the typicality and graded performance behaviors seen in humans. This work was done by Craig Miller for his thesis and was recently published:
The system that Craig Miller developed, SCA, has been used extensively in other Soar research projects with a simple demonstration developed by Doug Pearson.
6. Learning Dual Task Performance.
Our work on learning dual tasks, builds on the modeling of dual tasks by Kieras and Meyer in EPIC. Ron Chong, has integrated the sensor and motor system modules with Soar, creating EPIC-Soar, which we have used to model a specific task. In our modeling, we've been able to identify the specific bits of knowledge that must be learned to go from doing individual tasks to dual tasks. Our model learns one critical type of this knowledge, and our research will explore how the other types of knowledge can learn. This work is done by Ron Chong. We have a draft version of a paper on this topic:
As part of a course on symbolic cognitive modeling I taught with Randy Jones, our students explored the problem of how children represent and learn procedures for subtract. We built on the original work of VanLehn and Brown, although we moved the models into Soar. Randy and Peter Hastings created an environment in which students could modify existing subtraction procedures as well as develop their own. They could then experiment to discover which subtraction bugs were predicted by their model.
One of the goals of our research on computer generated forces was to have them be human like. A brief paper giving our philosophy on this and how it applied to simulated pilots was published in 1998 in the Cognitive Science Conference: Constraints on the Design of High-Level Model of Cognition, by Jones, R., and Laird, J.
General References to Research on Cognitive Modeling in Soar