Building a Cognitive Menu Within a CGF

1
Building a Cognitive Menu Within a CGF

• Walter Warwick
• 29 June 2004

2
Overview

• The problem
• The approach
• A new problem
• Why bother?

3
The Problem

• The behavioral realism of the synthetic entities
  that populate CGFs is lacking
• rigid, brittle and buggy
• impacts the use of MS in both analysis and
  training
• divorced from theory
• No easy avenue for remediation
• no one wants you to mess with their VVd library
  and no one wants to start over
• lots of one-off efforts to hook individual
  behaviors

4
One Approach to the Problem

• Identify the human behaviors within the CGF
• Expose selected behaviors to external control via
  a client server architecture
• following sensors and effectors (SA and decision)
  paradigm
• Develop external models using traditional HSI
  modeling tools (e.g., IMPRINT, Micro Saint) and
  more detailed cognitive architectures (e.g.,
  ACT-R)
• Make the approach user friendly by providing a
  GUI to the middleware layer

5
(No Transcript)
6
Menu GUI
7
Advantages

• Scalable architecture
• First level of disentanglement between systems
  models and behavioral models (not everything is a
  tank)
• Minimizes hacking within the CGF code
• Provides the human performance modeler with a
  familiar set of MS tools
• Make the client side application transparent to
  the external model

8
An Example
9
New Problems

• In this case, we have an external model
  controlling the high-level, tactical behaviors
  (e.g., which door, entry mode, fire permissions)
• the external model reflects all the advantages
  alluded to earlier
• SAF controls low-level, primitive behaviors
  (e.g., route planning, rate of movement, posture,
  formation)
• Plausible behavior at the tactical level, lousy
  behavior at the primitive level

10
Several Issues

• Some behaviors can be modeled in relative
  isolation (e.g., which door), some cant (e.g.,
  actions on contact, IFF)
• hardly ever clear a-priori how a given behavior
  will work out
• do we tweak the conceptual model or the
  architectural constraints within the CGF?
• Clearly, we cannot fix our attention at the
  tactical level, to the exclusion of lower levels
• How much behavioral realism is actually rooted in
  good primitive behaviors (a la Brooks)?
• How do we avoid regressing toward robotics?

11
Several More Issues

• Primitive is in the eye of the beholder
• Platform dependent
• Model dependent
• The line between cognitive modeling and CGF
  behavior engineering often gets very blurry
• E.g., would (should) an aggregation/arbitration
  module be cognitive plausible?
• What does any of this tell us about cognition?

12
Why Bother?

• A cognitive menu allows us to frame such
  questions empirically
• eliminate some of the practical obstacles to
  model development
• support bakeoffs
• More manageable scale rather than take top down
  approach (e.g., TacAir SOAR), start with small
  set of behaviors and attributes and see what you
  can do
• Incremental progress by reaching for low-hanging
  fruit
• Draw a line in the sand and whether it needs to
  be crossed