Untangling Computational Intuitions about Naturalistic Decision Making

1
Untangling Computational Intuitions about
Naturalistic Decision Making

• Walter Warwick
• Micro Analysis and Design, Inc.

2
Overview

• The old tangle
• Our approach
• What it means to model a naturalistic decision
  computationally in general
• What it actually looks like in practice
• A new tangle

3
The Old Tangle

• It is often argued that a careful line must be
  drawn between the attempt to accomplish with
  machines the same tasks that humans perform, and
  the attempt to simulate the process humans
  actually use to accomplish these tasks. The
  program discussed in this report maximally
  confuses the two approaches
• Newell and Simon, 1963

4
The Old Tangle

• In 1976 I read Dreyfuss book What Computers
  Cant Do and realized that his critique of the
  Artificial Intelligence position was also a
  critique of the information-processing account of
  cognition and expertise.
• Klein, 1998

5
The Old Tangle

• Cognitive models are often predicated on the
  metaphor of the Human Information Processor
• If nothing else, cognitive models are symbolic
  (even the connectionists ones)
• The roles of the computer qua metaphor for
  theoretical model and qua modeling artifact are
  often maximally confused
• Naturalistic theories reject computational
  accounts of cognition
• So how do you build a computational model of a
  naturalistic process?

6
Our Approach

• Initial Spin wheels, argue and explain away
  whatever whiffs of a rule-based implementation we
  could
• Plan B Recognize computer simulations as tools
  and focus instead on identifying the
  naturalistic dimensions to support principled
  (i.e., human-like) variability in synthetic
  behaviors (CGFs in particular)

7
Naturalistic Dimensions (ala the RPD)

• Experience (LTM, reinforcement criteria,
  sentinel events)
• Judgments of significance / recognizing relevant
  cues (cue weights)
• Differences that make a difference (cue
  discriminations)
• Expectancy evaluation (diagnostic reasoning via
  CBN)
• Process level effects (recall precision, variable
  certainty thresholds, workload induced
  limitations on WM)

8
(No Transcript)
9
(No Transcript)
10
CGF Applications

• Meso-level behaviors in the IUSS
• Tactical adjustments to course (posture, speed
  route)
• Obstacle avoidance
• Exploiting cover and concealment
• Sensitive to stressors

11
CGF Applications

• Actions on Contact in JointSAF
• Capture the notion of a force multiplier
  naturalistically

12
CGF Applications

• Building clearing in DISAF
• Which door, Which team, How to enter,
  Fire-permissions
• Reactive model for seemingly intentional behavior

13
Lead Balloons

• Naturalistic decision making isnt algorithmic
• Whats a probability calculus doing in a model of
  decision making?
• These new algorithms arent naturalistic
• Why dont we use this approach to improve pattern
  matching algorithms?
• Why model decision making at all?
• The OJ Defense Humans dont really make
  decisions, but if they did, why should we take a
  naturalistic approach when we already have a long
  normative tradition in place?

14
Lead Balloons (cont.)

• Important to distinguish between computational
  models of cognition and models of cognition as a
  computational process
• Likewise, its important to recognize that AI and
  cognitive modeling are distinct endeavors
  (maximally confusing the two often leads to
  maximal confusion)
• The assumption that humans really do make
  decisions we can model is certainly no stronger
  than the assumption that cognition is symbolic
• No useful alternative in either case
• Naturalistic view suggests new sources of
  variability for HBRs

15
The New Tangle

• In the context of CGFs, the really pressing
  questions are of a different sort
• Top-down versus bottom-up
• Integrative cognitive architecture versus modular
  approaches (e.g., a model of single decision
  embedded via a client-server architecture)
• Where do the physical-models stop and the
  cognitive models start (e.g., DIs in ModSAF as
  lightly armored tanks)?
• How can we make CGFs more accessible to cognitive
  modelers?
• CGFs as testbeds rather than just as the target
  applications