Instance-based and rule-based learning

Citation

Taatgen, Niels A. and Wallach, Dieter. “Whether Skill Acquisition is Rule or Instance Based is determined by the Structure of the Task.” Cognitive Science Quarterly: 2.2 (2002). 163 -204.

Summary

The traditional view of skill acquisition as “a gradual transition from behavior based on declarative rules in the form of examples and instruction towards general knowledge represented by procedural rules” is challenged by instance theory. While the latter seems to explain the inability of experts to verbalize rules, research on the directional asymmetry of rules seems to support the traditional importance of rules.

ACT-R’s instance-based architecture is based on 2 key arguments:

  • the strategy to use or the memory to retrieve is based on which has the highest expected gain (optimization)
  • declarative memory is activated (and filtered) by environmental demands and past experience (which has been encoded as production rules by procedural memory)

At the symbolic level of ACT-R, procedural rules are applied to declarative chunks which store information in a proposition; chunks are either new (perceptions) or created internally by prior knowledge/experience. Each rule contains a condition- and an action-part, and declarative items are pattern-matched to the condition and applied in the action. The subsymbolic level of ACT-R deals with the choice of which rule to apply according to Bayes’ Theorem (increases base-level activation each time it is retrieved and decays over time).

The article seems to imply that if instance-based learning fails (because the problem is too time-consuming–no discernible pattern), learners will attempt to derive some sort of rule. The authors argue that instance-based learning works best when the relationships between variables is very difficult; rule-based learning (simplified cases) is more successful with a large number of cases with obvious relationships.  The authors then test this hypothesis with two detailed experiments.

The results:

  • Previous research showed no learning through observation without direct rules (explicit relationships)
  • Subsequent research, indicated that exploratory participants did better than observers, but that observers could better verbalize and construct a causal model.
  • This research shows evidence of learning by observing even without rules; and participants seem better able to answer questions about old systems than new. Both results support instance-based learning.

Response

ACT-R seems to offer tremendous explanatory power. However, instance theory seems related to the concept of expert (tacit?) knowledge:

  • know when to apply
  • gets better over time (more instances)
  • gets better with use
  • additive for community

Because production rules that propose new declarative rules are not accounted for in the ACT-R architecture, this “missing link” may be the elusive transfer element; the authors propose partial-matching in ACT-R retrieval as a solution.

In instance theory, encoding and retrieval seem more closely linked to temporal/spatial reality than to attention as the authors claim; however, the view of memory as evolving from algorithmic processing to memory-based processing succinctly describes expert knowledge.

The recommendations for design strategies seem profound:

  • Instance-based learning takes over from rule-based over time
  • Creating declarative rules is the most important (first) step
  • Analogy works best at the start but declines quickly
  • Declarative rule works well at start and persists
  • Instance continues to improve over time to become best
Advertisements

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out /  Change )

Google+ photo

You are commenting using your Google+ account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s

%d bloggers like this: