Scientific method, Symbolic models - Intro to Cognitive Science: Cognitive Modeling - Fall 2009

Contents

Slides

Overview of Today's Class [1/5, 2009/09/28]

• Project 1 postmortem?
• More CogSci on the radio
• Continue reflection exercise which we skipped last time.
• Discuss Anderson paper
• Prep for Repair Theory paper

Reflection exercise [2/5, 2009/09/28]

Much of language can be described as rule-like.

Small group discussion: Groups of 4

1. Think of a cognitive task besides language -- one of your areas of "expertise" would be good.
2. Speculate on how that task could be "modeled" with a rule-based system.

Anderson (ACT-R) discussion [3/5, 2009/09/28]

Your issues?

Working the figures

ACT-R [4/5, 2009/09/28]

Example: Learning past tense of english verbs.

How it works:

• Set of "training words" with frequencies.
• Two rules for retrieving. Same conditions, different actions.
• If retrieval works directly, say it.
• If another word is chosen, use "analogy": either regularize or do not inflect.

Sample code from ACT-R tutorial

• "Hears" verbs (reg and irreg) in same freq distrib as kids (from CHILDES database). Top 6 (out of 31K): 12000 x have, 4300 x do, 2300 x make, 1500 x get, 1000 x use, 900 x look.
• First: request retrieval. Two production rules with same conditions. One asks for a retrieval of inflection for given word, the other asks for inflection of any word.
• Second: two rules for processing result, one for success -- notice what was retrieved, one for failure -- give up.
• Finally: if exact word was retrieved, use that result, reg or irreg. If another word retrieved, follow its example, either +ed or not.
• Learns strength of rules, strength of chunks in memory.
• Irregs are more frequent. More regs than irregs.

Some sample runs

Next time [5/5, 2009/09/28]

• Brown and VanLehn paper: Hairiness warning! But it should be the last one. Describes a model of subtraction, and how it goes wrong. Focus on:
• Difference between good and bad subtraction procedures.
• Description and detailed example with GAO graph.
• What is a "bug"? What is a "repair"?
• Some answers:
• Bugs are systematic errors.
• Assumes that there is a complete, ideal solution to the task, but some part or parts may be missing. If so, might hit impasse, need a repair.
• GAO: Generalized And-Or graph. Rule-like with some extra machinery.
• Project 2 description and groups: Basically, experiment with an application that simulates Brown and VanLehn's model