Learning from Instruction in Soar

Huffman (1993) describes a Soar system (Instructo-Soar) that takes instruction from a teacher in a blocks world domain. The agent can learn completely new procedures, including information about the goal itself (i.e., what it means to have accomplished the goal) and how to perform the procedure.

As the agent receives instructions, it builds chunks that capture the details of the instruction; this is an example of episodic knowledge. Such a memory requires effortful reconstruction to be used. The chunked productions act as a recognitional memory when the semantic features that were associated with a particular instruction are placed in working memory. Thus, using the episodic memory is indirect, leading to the recognitional recall.

One of the results of the work on Instructo-Soar has been a deep analysis of the type superficially outlined above. For example, the 'Miller-Huffman' diagram shows the way the basic Soar architecture gives rise to entailments. The entailments then combine to form knowledge level capabilities (KLC) for the Instructo-Soar model. KLCs are capabilities that an agent may apply to many different tasks within the agent domain, that occur over a long-time period (i.e., minutes instead of immediate response), and which are chiefly characterized by the agent's knowledge. In the diagram, for instance, two KLCs include:

Interactive instruction requests which arise from the goal-based action of the agent and its ability to detect a lack of knowledge (determined by an impasse); and
Experiential, situated learning which arises from properties of the chunking mechanism.

This example shows that KLCs can be very specific to an agent's task (in the fist case) or much more general (all learning in Soar can be viewed as situated and experiential).

This diagram also points out an important detail about this analysis. Although the property list broke down the architectural entailments into categories determined by a particular architectural feature, many of these properties derive from more than one architectural mechanism as can be seen in the Miller-Huffman diagram. Thus, the breakdown of properties there is somewhat over-simplified; properties are actually derived from the synergistic interaction of several of the underlying architectural features and the primary one has been identified in the property list.

In addition to analysis of an existing system, the technique represented by the Miller-Huffman diagram is also useful for predicting behavior. A similar diagram was used to make predictions about Instructo-Soar's behavior. In this case, the interaction of KLCs predicted general effects such as learning transfer and psychological phenomena such as the Einstellung effect (the preservation of a skill no longer useful in a new or changed environment).

Return to the top of this architecture.

Go to a discussion of this capability for multiple architectures.

Current Location: Soar-Capabilities-Learning from Instruction

Go to NEXT page.