Reading guide: Gentle Intro to Soar

Background

Soar is a cousin of ACT-R. Both started at Carnegie Mellon, and had the same ancestor systems. Soar is used more for high-level modeling of expert behavior, and ACT-R is used more for low-level psychological models.

Things to watch out for

• Intro section gives the background and underlying rationale.
• Section 2 motivates using an architecture with lots of constraints, as opposed to developing special-purpose models for separate cognitive tasks.
• Sec. 3: basic assumptions of the approach. Look for similarities and differences with ACT-R.
• Sec 4 is a bit abstract. It describes how the "thinker" can separate out relevant from irrelevant knowledge while doing a task. Fig. 3 shows a set of possible ways that a certain situation could play out over time, based on the choices that the actor/agent makes. Fig. 4 gives another view of the same idea. Here, an "operator" (What's that?) moves from one state node (What's that) to another, starting at the initial state, and ending (hopefully) at the goal state.
• Sec. 5 shows with the baseball example how the knowledge for the situation is represented in Soar. It must be specific enough to allow the agent to come up with a solution, but it must be general enough so that it fits within the framework of the architecture (which is supposed to support all of cognition).
• Sec. 6 shows the interaction of the major components of Soar. Perception makes information about the real world available for short-term (working) memory. (But no word about attention until later.) Long-term memory is accessed to find applicable knowledge to help the agent solve its goals. When an operator is chosen, the agent acts, moving its body (somehow) to create some effect in the world. (How does this relate to ACT-R?)
• Sec. 7 sets up the approach to learning in Soar. It happens automatically when the agent does not immediately know what to do in a certain situation. This section describes how it knows that it doesn't know.
• In Sec. 8, the learning approaches are described. Approaches?? for the first 20 or so years of Soar's existence, it had only only learning mechanism, Chunking (Sec. 8.1). Recently the Soar folks have realized that there's got to be more, and added three others. Compare these to the learning described in ACT-R.
• Sec. 9 gives brings everything together in the baseball example.
• Sec. 10 gives a (very helpful) review.
• Sec. 11 talks about generalizing from one example to something that proposes to be applicable to all of cognition. And it gives many examples of Soar models for a variety of tasks.