Title: INSIGHT, COMPLEXITY, AND VALIDITY IN HUMAN BEHAVIOR REPRESENTATION 11 Sep 06
1INSIGHT, COMPLEXITY, AND VALIDITY IN HUMAN
BEHAVIOR REPRESENTATION11 Sep 06
- Kevin A. Gluck, PhD
- Senior Research Psychologist
- Human Effectiveness Directorate
- Air Force Research Laboratory
2Acknowledgments
- Thanks to Christy Caballero (research assistant)
and Deborah Russell (graduate student intern) for
their contributions to portions of this
presentation - I will borrow lightly from Bob Fosters BRIMS
2003 Keynote Address - I will borrow heavily from the theorizing of Herb
Simon
3A Call for Insight
Achieving an ahh..ha in behavior
representations why DoD is interested
Slide adapted from Dr. Bob Fosters keynote
address at the 2003 Conference on Behavior
Representation in Modeling and Simulation (BRIMS)
in Scottsdale, Arizona.
4Alternative InterpretationsThe Locus of the
Insight
5Overview
- Insight
- Insight as problem solving
- Complexity
- Provides both challenge and guidance
- Validation
- Crucial component of the process
- Propose a change to standard practice
6Sample Insight Problems
- Verbal
- One morning a woman's earring fell into a cup
that was filled with coffee, yet her earring did
not get wet. How could this be?
From http//www.indiana.edu/bobweb/Handout/insigh
tproblems.doc
7Insight as Problem Solving
- Newell Simon (1972) Human Problem Solving
- All goal-directed activity is problem solving
- Solving problems involves traversing a problem
space
Problem Space
Initial State
Intermediate States
Goal State
operators
8A Process Theory for Insight
The Mutilated Checkerboard Problem (Kaplan
Simon, 1990)
Cover the remaining 62 squares with 31 dominoes.
Each domino covers two adjacent
squares. Or Prove logically why such a covering
is impossible.
9Conclusions from Insight Research
- The key to achieving insight is to arrive at an
appropriate representation (have to be in an
appropriate problem space, with all necessary
operators). - Typically do not have a sense of progress
toward the goal until you arrive at the solution. - An important heuristic in insight problems
- Attend to the features of the problem that remain
invariant
10The Connection to Complexity
- We will achieve HBRs capable of having insight
when they become an appropriate (necessary and
sufficient) replication of the human cognitive
system - This goal (computational replicates of humans)
involves nested search across multiple, infinite,
interacting problem spaces - The space of all possible HBRs
- The space of all possible insight contexts
- The use of heuristics constrains search and
allows progress in very large (or infinite)
problem spaces - The complexity of this goal is crippling, unless
we find a way to draw heuristics out of the
complexity
11Complexity
- The Architecture of Complexity Hierarchic
Systems - Chapter 8 in Herb Simons book The Sciences of
the Artificial
4 Key Points
Complex systems are nearly always
hierarchic Hierarchy facilitates evolution Near
decomposability Nearly decomposable, hierarchic
structures facilitate comprehension
12A Vector to Validity
- HBRs are complex systems that are evolving
- Modular, hierarchic designs will increase the
pace of that evolution - Evolution is a process that evaluates the
goodness of a system for some purpose
Validity
Must have the evaluation function of validation
in order to make decisions about which
sub-features of the HBRs should remain stable and
which should change.
13A Call for Validation
WHAT WE NEED
- Affordability
- Adaptability
- Robustness
- Scalability
- Composability
- Interoperability -- Jointness
- Common research testbeds
- Validity---validity---validity!!
Slide adapted from Dr. Bob Fosters keynote
address at the 2003 Conference on Behavior
Representation in Modeling and Simulation (BRIMS)
in Scottsdale, Arizona.
14DoD Position on Validation
- DoD Instruction 5000.61 (May 13, 2003)
- 4.5 The DoD Components shall establish VVA
policies and procedures for models and
simulations they develop, use, or manage. - 6.1. Verification and validation (VV) shall be
- 6.1.1. Incorporated into the development and
life-cycle management processes of all MS.
15DMSO Position on Validation
- From DMSOs VVA Recommended Practices Guide
- http//vva.dmso.mil/Default.htm
- Why is VVA performed?
- To determine whether a model or simulation or
federation should be used in a given situation,
its credibility should be established by
evaluating fitness for the intended use. - The decision to use the simulation will depend
on the simulations capabilities and correctness,
the accuracy of its results, and its usability in
the specified application.
16The Situation
- Despite appropriate policy guidance there is
little or no standard practice (or expectation,
apparently!) in our research community for
validating HBRs - Often no validity evaluation at all
- When there is, it is often reported in vague,
qualitative terms - Results were very satisfying actual BRIMS
paper quote, with no supporting quantitative data
17The EvidenceData from BRIMS 2005
What sort of papers were presented/published at
BRIMS 2005?
Papers
- Tool/Capability
- Cognitive Model
- Position/Review
- Human Beh. Repr.
13 12 9 6
Without the position/review papers, thats 31
papers in the sample. Of those 31, how many
report any form of validity evaluation at all?
12
18More Data from BRIMS 2005
When validity is evaluated, how is it evaluated?
Type
- Application Validity
- Construct Validity
- SME Face Validity
- Developer Face Validity
6 2 3 1
How many of these involved a quantitative
evaluation of validity?
5
19A Little Closer to HomeData from Spring 2006 SIW
- 7 Papers related to HBR
- 029 Aha Christman testbed for analysis of SAF
planning behaviors - 045 Goerger et al agents for traffic flow
analysis - 055 McKenzie Piland crowd federate GUI
- 073 Nanda Weeks path prediction of dynamic
entities - 076 Tegner, Huang, Pakucs visual rep comm
standards for HBRs - 085 Weisel et al crowd modeling
- 127 Rowe et al improving human interfaces
- (but one of these is a position paper)
Of the 6 papers on HBR-related models, tools, or
methods, how many report any form of validity
evaluation at all?
2
20Why This is a Problem
- We need objective, quantitative evaluations of
validity in order to establish credibility, track
ST progress, and increase the probability that
our HBRs and associated tools are evolving in a
direction that actually increases their utility. - How can we take ourselves seriously if were
investing the effort to establish our credibility
(through quantitative evaluation of validity) in
less than 1/3rd of our published research
efforts? - Why should anyone take us seriously?
21You Want a Standard?
I propose a new standard practice
- ALL government contracts and grants for
developing cognitive models, human behavior
representations, or new CM/HBR-related tools and
capabilities must include an explicitly defined
and budgeted plan for objective, quantitative
validation. - Validation must be a technical deliverable, even
in the early stages of the acquisition process
22Validation Must Be Done
- It is not enough to simply implement code and
describe it. - Step 1 in any new HBR RD effort should be to
identify or collect the empirical
data/phenomena/functionality we are trying to
replicate, understand, or achieve - Provides the foundation for assessing accuracy
and correctness, thereby establishing credibility - Excuses such as its expensive/difficult/time-con
suming are cop-outs that reduce credibility,
slow progress, and are not sufficient
justification for avoiding validation
23Implications
- If we want HBR systems that are understandable
and that evolve as quickly as possible in
demonstrably useful directions, we should - Design them hierarchically
- Validate the sub-components at each level, so we
know where we have stable utility and where we
need adaptation/mutation - The probability that we will achieve the goal of
HBRs that have a general capacity for achieving
insights of their own, any time in the next 50
years, depends on our adopting appropriate design
and search heuristics such as these.
24Thank You