Production Models of Cognition

1
Production Models of Cognition

• CS/ISYE/PSYC 7790 Cognitive Modeling
• Fall 2003

2
Administrivia

• Lab 1 due this Wednesday, in class
• Lab 2 will use ACT-R

3
Outline

• What constitutes a production system?
• How can production systems be mapped onto
  psychological phenomena?
• Productions as mediating states

4
What is a production system

• Problem-solving architecture
• Classical problem solver
• Rule-based reasoners
• Expert systems
• Part
• Memory
• Set of productions
• Solves problems by working through a problem space

5
Example plan Travel to Brasilia hotel

• Ask wife to drive to Midtown station
• Drop off at station
• Take subway to Atlanta airport
• Check in
• Get on plane
• Take plane to Sao Paulo
• Take another plane to Brasilia
• Take taxi to hotel
• Check in at hotel
• Go to room
• Sleep off effects of long trip

6
Production systems work on a problem space

• A set of states
• Declarative knowledge
• Start state
• Goal state
• A set of productions
• Rules or operations or productions that
  indicate how the existing state can be changed
• Specified abstractly using variables

7
Example of states

• Declarative knowledge states
• I own a car
• I do not own two cars
• Starting state
• Im on campus at Georgia Tech
• Goal state
• Im at a hotel in Brasilia

8
Productions

• Each production contains
• Preconditions
• What states must be true for this production to
  fire?
• Add rules
• What new states must become true if this
  production fires?
• Delete rules
• What old states become false if this production
  fires?
• For the last two, there are some variations
  (e.g., modification rules)

9
Example Checking in at airport

• Preconditions
• Must have goal of checking in
• Must be at airport
• Must be at ticket counter
• Must have passport
• Must have checkable luggage
• Must be at least one hour before flight
• Add rules
• Checked onto flight
• Have boarding pass
• Have stubs for checked luggage
• Delete rules
• No longer have goal of checking in
• No longer have checked bagged

10
Other aspects of production systems

• Goal stack is used to keep track of goals and
  subgoals
• Productions are selected in parallel
• Leads to conflicts when there are multiple
  applicable productions
• Example Taking the MARTA or driving to the
  airport?
• Various mechanisms are used for conflict
  resolution for productions
• Which production gets me closest to my goal?
• Which production did I use more recently?
• Segmentation of knowledge by domain
• Distinction between procedural and declarative
  memory
• Declarative memory is accessible and slow
• Procedural memory is inaccessible and very fast

11
Outline

• What constitutes a production system?
• How can production systems be mapped onto
  psychological phenomena?
• Productions as mediating states

12
Production systems are good at modeling
particular aspects of cognition

• Focus on problem-solving tasks
• Depends on how productions themselves may fail,
  or may fail to fire

13
How human problem-solving errors may be modeled
by production systems

• Bad productions
• Always subtracting the smaller numeral in a
  multicolumn subtraction problem
• Over- or under-specific preconditions
• Leads to over-application of production, or
  failure to apply production when it could be used
• Bad heuristics for choosing productions
• Even if a better production exists,
  conflict-resolution strategy may favor inferior
  production
• Water job problem and persistence
• Problem with working memory or goal stack
• Both have been used to model problems with
  problem-solving in Tower of Hanoi

14
Modeling learning in production systems

• (Least to most risky)
• Strengthening tendency to use certain productions
  (tuning)
• Composing productions into aggregate productions
• Chunking in SOAR
• Mutating existing productions
• Removing preconditions
• Example Dont need luggage to check into airport
• Creating new productions by analogy
• Often works better for sets of productions

15
Outline

• What constitutes a production system?
• How can production systems be mapped onto
  psychological phenomena?
• Productions as mediating states

16
Productions as mediating states

• Enduring
• Productions endure
• Production instantiations may not
• Discrete and composable
• These belong together productions are
  definitely both
• To model learning and errors, productions must be
  both
• Model learning by strengthening one rule over
  another

17
Productions as mediating states

• Abstract
• Productions abstract away from particular
  instances
• Yet, at the same time, they show many of the
  characteristics of concrete items
• Wont easily transfer
• Not accessible to consciousness
• Another interesting difference
• Situated cognition concrete knowledge
  foundational, abstracted eventually
• ACT-R declarative knowledge fundamental,
  eventually compiled into procedural knowledge
• Rule-driven

18
Differences between production systems and
connectionist networks

• Production systems have no intrinsic notion of
  similarity like that of connectionist networks
• Conflict resolution is handled differently
• Wired into connectionist networks
• For production systems, tends to involve
  conflicts between applicable productions

19
Next time ACT-R

• Overview of ACT-R system
• We begin our lab on ACT-R