Validation Methodology for Agent

Validation Methodology for
Agent-Based Simulations Workshop
Perspectives on
Agent-Based Simulation and VV&A
Dr. Bob Sheldon
Joint and External Analysis Branch
Operations Analysis Division
Marine Corps Combat Development Command
01 May 2007
Overview
• VV&A and Agent-Based Simulation (ABS)
thoughts from Dr. George Akst, Senior
Analyst, Marine Corps Combat Development
Command (MCCDC)
• MORS historical perspectives on VV&A and
ABS
• Personal reflections
Perspectives from Dr. Akst
• It’s the data, stupid!
 How do you come up with data for parameter Z = x.x %?
• Especially a problem for Irregular Warfare (IW)
 Sometimes, model developers who are structuring algorithms
don’t worry about data & assume data can be developed after the
fact
• Dr. Kirk Yost data triage – consider data sources when building
models
– Generally accepted (produced regularly by some believable
source)
– Semi-valid (reasonable information derived from various sources)
– Judgment and knobs
 If you start with meaningless data, and execute a design of
experiments with 210 runs (just because you can), then you will
have 210 useless results
• To be useful, ABS need to provide more than just
simplistic insights
 ABS should go beyond being an automated tool that regurgitates
SME intuition
More Dilbert Data
Perspectives from Dr. Akst
• Two ends of the spectrum
 Engineering-level model: should very closely predict how
system would operate in the real world
 Campaign-level model: measure relative differences that
changes to forces, tactics, or equipment have on the outcome
• Trying to literally match a combat model’s results with
some other set of results (real world, experiment, or
another model) is not realistic
• What validation is:
 Failure to invalidate after concerted effort
 Ascertaining that results are “plausible” – no obvious logic
flaws and results are “reasonable” and “relatively consistent”
with past modeling results
From “Musings on Verification, Validation, and Accreditation (VV&A) of
Analytical Combat Simulations” Phalanx, September 2006
MORS Meetings on VV&A
•
•
•
•
Simulation Validation (SIMVAL), October 1990
SIMVAL II, April 1992
SIMVAL '94, September 1994
Simulation Validation tutorial, MORSS & ALMC, 1995
(Pete Knepell)
• SIMVAL '99: Making VV&A Effective and Affordable,
January 1999
• Evolving Validation Topics in MORS
 Descriptive validity, Structural validity, Predictive validity
 Structural validation, Output validation
 Conceptual Model validation, Data validation, and Output
validation
MORS Meetings on ABS
• New Techniques: A Better Understanding of
their Application to Analysis, November 2002
 Included 1-day tutorial on Agent-Based Models
• Agent-Based Models and Other Analytic
Tools in Support of Stability Operations,
October 2005
• Plus substantial coverage in MORSS working
groups, e.g., WG 31 – Computing Advances
in Military OR and WG 32 - Social Science
Methods
Personal Reflections
• How to validate (or invalidate) counterintuitive results (e.g., Surprise)
 Clay Thomas “Analysis either verifies your
intuition or educates your intuition.”
• Simple visualization helps validation
 Gantt chart example for sortie generation
100 T/O - Enrt
150
150
101
102
150
STK
150
2
T/O - Enrt
150
150
150
140
2
150
Enrt-F FARP Rearm & Refuel
Enrt-S STK
120
110
110
110
110
110
110
110
110
100
90
80
2
2
2
2
2
STK
Enrt-F FARP Rearm & Refuel
Alert FARP
150
150
140
130
120
110
110
110
110
110
110
110
110
2
2
2
2
T/O - Enrt
STK
Enrt-F FARP Rearm & Refuel
150
150
150
150
150
140
130
120
110
110
110
110
2
2
2
2
130
2
70
2
60
2
110
110
110
110
 Provide visualization that SMEs understand
Personal Reflections (Cont’d)
• Ready access to source code helps
 Example: Effect of (0,1) parameter
• Good mathematical
documentation a
plus
1
0.8
0.6
Y
0.4
0.2
0
0
0.2
0.4
0.6
X
0.8
1
Personal Reflections (Cont’d)
• Comparing counter-intuitive results to
“intuitive” results: a case study
• At a Project Albert workshop, the agent-based
model Socrates gave counter-intuitive results
 Simulation attrition results varied over 3 phases
with 2 breakpoints
 When I fit a Lanchester linear model to the results,
the regions where the fit was “bad” corresponded to
the counter-intuitive results
 Drill-down investigation explained these anomalies
 Mysterious results were due to scenario data &
tuning parameters
“Comparing the Results of a Nonlinear Agent-Based Model to
Lanchester’s Linear Model” Maneuver Warfare Science 2002
Questions?
Juan Muñoz, Five Seated Figures, 1996
Hirshhorn Museum and Sculpture Garden