ABS VV&A Framework Study Phase II Pythagoras COIN – Application of the Validation Framework
description
Transcript of ABS VV&A Framework Study Phase II Pythagoras COIN – Application of the Validation Framework
ABS VV&A Framework Study Phase IIPythagoras COIN – Application of the Validation
Framework
Lisa Jean MoyaWernerAnderson, Inc.
Phase II Workshop 39 July 2008
7/9/2008
1
P-COIN Validation Briefing Wkshp3 Moya
Scenario
7/9/2008 P-COIN Validation Briefing Wkshp3 Moya
2
Tsun
ami o
ccur
sM
EU/M
EB o
rder
ed to
AO
A
MEU
sai
ls fo
r AO
A
MEB
sai
ls fo
r AO
A
MEB
arri
ves
AOA
MEU
arri
ves
AOA
PB
day
PB
-10
PB
-19
PB
-24
PB
-25
PB
-26
MEU
Fw
d m
oves
ash
ore
& es
t lia
ison
PB
-18
MEU
con
duct
s in
itial
sur
vey
Analysis Context• Can Pythagoras be used to model population dynamics?
• In a Disaster Relief/Humanitarian Assistance mission for the stated scenario, is it better to base the MAGTF ashore or afloat?
• Alternative selection drivers– Do no harm: create no increase in insurgency activity.– Improve the political situation: create an improvement in GOVT and
Pro-GOVT sectors.• Measures
– Box & Whisker plot comparisons of the percent of population by population segment in each insurgency sector at end state (18 months)
7/9/2008 P-COIN Validation Briefing Wkshp3 Moya
3
Conceptual Model of Civilian Population4
FARC Pro-FARC Neutral Pro-GoC GoC
Insurgency Behavior Orientation
CivilianPopulation
PopulationSegments
FARC = Revolutionary Armed Forces of Colombia
GoC = Govt of Colombia
Natural DriftSalienceInfluencing events
7/9/2008 P-COIN Validation Briefing Wkshp3 Moya
7/9/2008 P-COIN Validation Briefing Wkshp3 Moya
5
Catholic Church
Displaced Persons
Illicit Organizations
Military
Old Money
Police
Urban Middle Class
Urban Poor
FARC Pro-FARC Neutral Pro-GOVT GOVT
7/9/2008 P-COIN Validation Briefing Wkshp3 Moya
6
Catholic Church
Displaced Persons
Illicit Organizations
Military
Old Money
Police
Urban Middle Class
Urban Poor
FARC Pro-FARC Neutral Pro-GOVT GOVT
3.03
2.63
3.76
3.89
3.21
4.23
3.62
3.39
Areas of Interest
• Core– MAGTF influence on
Insurgency orientation• Cases
– MAGTF/No MAGTF– Ashore/Afloat
• Dynamic Influences– Natural Drift– Salience
• Background– Population segments
7/9/2008 P-COIN Validation Briefing Wkshp3 Moya
7
Areas of Interest
• Core– MAGTF influence on
Insurgency orientation• Cases
– MAGTF/No MAGTF– Ashore/Afloat
• Dynamic Influences– Natural Drift– Salience
• Background– Population segments
7/9/2008 P-COIN Validation Briefing Wkshp3 Moya
8
First order assessment
Agent Allocation
7/8/2008 WG29MoyaLisa 13
Segment FARC Pro-FARC Neutral Pro-GOVT GOVT Catholic Church 0 0 38 0 62 Displaced Persons 7 37 45 10 1 Illicit Organizations 0 39 28 18 15 Military 2 10 0 88 0 Old Money 0 2 4 63 31 Police 0 5 0 95 0 Urban Middle Class 4 4 62 10 20 Urban Poor 6 9 66 15 4 Total 19 106 243 299 133
Pythagoras-COIN Building Blocks
7/9/2008 P-COIN Validation Briefing Wkshp3 Moya
9
Natural Drift: Incremental
6/10/2008 WG29MoyaLisa 12
FARC Pro-FARC Neutral Pro-GOVT
GOVT
FARC 99.8% 0.1% 0.0% 0.0% 0.0%
Pro-FARC 0.4% 99.5% 0.2% 0.0% 0.0%
Neutral 0.0% 0.7% 98.7% 0.5% 0.0%
Pro-GOVT 0.0% 0.0% 0.9% 99.1% 0.1%
GOVT 0.0% 0.1% 0.0% 0.1% 99.7%
Attribute 1 Attribute 2 Attribute 3 Attribute 4 Attribute 5
+4 0 +2 0 0
Displaced Persons Pro-FARC Attribute Changes
Catholic Church
Displaced Persons
Illicit Organizations
Military
Old Money
Police
Urban Middle Class
Urban Poor
FARC Pro-FARC Neutral Pro-GOVT GOVT
Salience: Relative
6/10/2008 WG29MoyaLisa 13
Cath
olic
Chur
ch
Disp
lace
d Pe
rson
s
Illic
it O
rgs
Mili
tary
Old
M
oney
Polic
e
Urb
an
Mid
dle
Clas
s
Urb
an
Poor
Catholic Church 0.000 0.000 0.000 (0.078) 0.801 (0.078) 0.000 0.000Displaced Persons 0.706 (0.262) (0.052) (0.230) 0.178 (0.366) 0.480 (0.208)Illicit Organizations 0.207 0.000 0.425 (0.053) 0.420 (0.190) (0.467) 0.183Military 0.750 (0.351) (0.295) 0.492 0.262 0.685 0.000 (0.257)Old Money 0.474 (0.486) (0.713) 0.000 0.840 (0.506) 0.494 (0.250)Police 0.000 (0.843) (0.174) (0.039) (0.273) 0.767 0.000 (0.218)Urban Middle Class 0.816 (0.623) (0.387) 0.136 (0.269) (0.205) 0.392 0.132Urban Poor 0.202 (0.259) 0.101 0.225 (0.033) (0.134) 0.552 (0.079)
Catholic Church
Displaced Persons
Illicit Organizations
Military
Old Money
Police
Urban Middle Class
Urban Poor
FARC Pro-FARC Neutral Pro-GOVT GOVT
3.03
2.63
3.76
3.89
3.21
4.23
3.62
3.39
Influencing Events: Multiplier
6/10/2008 WG29MoyaLisa 15
Sea Based Shore Based
Segment Right Left Right Left Catholic Church 0.845 0.000 0.401 0.000 Displaced Persons 0.596 0.000 0.721 0.117 Illicit Organizations 0.000 0.397 0.000 0.447 Military 0.000 0.408 0.000 0.408 Old Money 0.631 0.000 0.631 0.000 Police 0.564 0.000 0.564 0.000 Urban Middle Class 0.780 0.000 0.184 0.210 Urban Poor 0.798 0.000 0.722 0.211
Salience as a Dynamic Influence
7/9/2008 P-COIN Validation Briefing Wkshp3 Moya
10
Cath
olic
Ch
urch
Disp
lace
d Pe
rson
s
Illic
it O
rgan
izatio
ns
Mili
tary
Old
Mon
ey
Polic
e
Urb
an M
iddl
e Cl
ass
Urb
an P
oor
Catholic Church 0.000 0.000 0.000 (0.078) 0.801 (0.078) 0.000 0.000 Displaced Persons 0.706 (0.262) (0.052) (0.230) 0.178 (0.366) 0.480 (0.208) Illicit Organizations 0.207 0.000 0.425 (0.053) 0.420 (0.190) (0.467) 0.183 Military 0.750 (0.351) (0.295) 0.492 0.262 0.685 0.000 (0.257) Old Money 0.474 (0.486) (0.713) 0.000 0.840 (0.506) 0.494 (0.250) Police 0.000 (0.843) (0.174) (0.039) (0.273) 0.767 0.000 (0.218) Urban Middle Class 0.816 (0.623) (0.387) 0.136 (0.269) (0.205) 0.392 0.132 Urban Poor 0.202 (0.259) 0.101 0.225 (0.033) (0.134) 0.552 (0.079)
Salience as a Dynamic Influence
7/9/2008 P-COIN Validation Briefing Wkshp3 Moya
11
Cath
olic
Ch
urch
Disp
lace
d Pe
rson
s
Illic
it O
rgan
izatio
ns
Mili
tary
Old
Mon
ey
Polic
e
Urb
an M
iddl
e Cl
ass
Urb
an P
oor
Catholic Church 0.000 0.000 0.000 (0.078) 0.801 (0.078) 0.000 0.000 Displaced Persons 0.706 (0.262) (0.052) (0.230) 0.178 (0.366) 0.480 (0.208) Illicit Organizations 0.207 0.000 0.425 (0.053) 0.420 (0.190) (0.467) 0.183 Military 0.750 (0.351) (0.295) 0.492 0.262 0.685 0.000 (0.257) Old Money 0.474 (0.486) (0.713) 0.000 0.840 (0.506) 0.494 (0.250) Police 0.000 (0.843) (0.174) (0.039) (0.273) 0.767 0.000 (0.218) Urban Middle Class 0.816 (0.623) (0.387) 0.136 (0.269) (0.205) 0.392 0.132 Urban Poor 0.202 (0.259) 0.101 0.225 (0.033) (0.134) 0.552 (0.079)
So "Initial" Values Displaced Persons Neutral "Initial"
Attribute 1 Attribute 2 Attribute 3 Attribute 4 Attribute 5 0 7 988 5 0
Catholic Church ProCOIN "Initial" 0 0 0 1000 0
Catholic Church Influence
- - - 71% - Displaced Persons Change
0 7 988 711 0 Final Values (Normalization)
0 4 579 417 0
As the simulation might progress …
7/9/2008 P-COIN Validation Briefing Wkshp3 Moya
12
So "Initial" Values Displaced Persons Neutral Agent
Attribute 1 Attribute 2 Attribute 3 Attribute 4 Attribute 5 0 7 988 5 0
Catholic Church ProCOIN "Initial" with attributes at a future timestep (currently Pro-FARC) 0 1000 0 0 0
Catholic Church Influence
- - - 71% -
Displaced Persons Change 0 7 988 1 0
Final Values (Normalization)
0 4 992 1 0
An extreme example to demonstrate the issue
Implications• Influence changers only applied w.r.t. initial state;
changes in orientation do not change the influence→ Dynamic effects of salience and natural drift are not
accounted for→ Secondary and tertiary effects of MAGTF arrival not
accounted for… Dampening on the insurgency orientation! Risk is that the simulation does not model the desired
population dynamics ! Risk is that the dynamics of the MAGTF arrival are not
adequately captured7/9/2008 P-COIN Validation Briefing Wkshp3 Moya
13
Data
7/9/2008 P-COIN Validation Briefing Wkshp3 Moya
14
"Humanitarian exchange." Wikipedia, The Free Encyclopedia. 3 Jul 2008, 10:26 UTC. Wikimedia Foundation, Inc. 9 Jul 2008 <http://en.wikipedia.org/w/index.php?title=Humanitarian_exchange&oldid=223273181>.
• Data imprecise– Mitigated by “tolerance” and
multiple runs• Data processing
– Need to verify that process results in expected directional & magnitude shifts
• Data is perishable– Natural drift data has an
embedded perishibility• Would an actual model use
require a “warm-up period” on the Markov Chain?
– Other influencing events might significantly change the data values
Uses of the Model
• Markov assumptions in referent descriptions– No long term effects– May need a “warm-up”
period• Outside the salience or along
with salience?
• Data precision– No exact results; results in
the distribution• Data perishability
– Need to collect new data after significant events
– Including MAGTF departure
• Were the dynamics captured … – Could add other influencing
events• Could add additional
dynamics
• Q: Can we apply the influencers more robustly?
• Q: Would changing our initial starting agents help?
7/9/2008 P-COIN Validation Briefing Wkshp3 Moya
15
Analysis Results
• Given in box & whiskers plots at end state with data table
• No statistical comparisons
• No “hard” description of better
• Point estimate in time (18 months)
7/9/2008 P-COIN Validation Briefing Wkshp3 Moya
16
Percent Population Pro-COIN, COIN
40
50
60
70
80
90
100
NoM
AG
TF
Ash
ore
Aflo
at
NoM
AGTF
Ash
ore
Aflo
at
NoM
AGTF
Ash
ore
Aflo
at
NoM
AG
TF
Asho
re
Aflo
at
NoM
AGTF
Ash
ore
Aflo
at
NoM
AGTF
Ash
ore
Aflo
at
NoM
AGTF
Ash
ore
Aflo
at
NoM
AGTF
Ash
ore
Aflo
at
Catholic Church Displaced Persons Illicit Organizations Military Old Money Police Urban MiddleClass
Urban Poor
Perc
ent P
opul
atio
n
Afloat has nearly equal or more Pro-COIN, COINPercent Population Pro-FARC, FARC
0
5
10
15
20
25
30
35
40
NoM
AGTF
Ash
ore
Aflo
at
NoM
AGTF
Ash
ore
Aflo
at
NoM
AGTF
Asho
re
Aflo
at
NoM
AGTF
Ash
ore
Aflo
at
NoM
AGTF
Ash
ore
Aflo
at
NoM
AGTF
Ash
ore
Aflo
at
NoM
AG
TF
Ash
ore
Aflo
at
NoM
AG
TF
Ash
ore
Aflo
at
Catholic Church Displaced Persons Illicit Organizations Military Old Money Police Urban Middle Class Urban Poor
Perc
ent P
opul
atio
n
Afloat has the same or fewer Pro-FARC, FARC
MAGTF Influence Data
• Expect: Sea better than shore – Urban Middle Class & Urban Poor “drive” result
• Catholic Church drive more right with Sea vs Shore• Displaced Persons “wash”?• Salience causes Urban Poor and Middle Class to be like Military;
Military to be like Catholic Church – in opposition to the direct MAGTF influence … what would we expect?
7/9/2008 P-COIN Validation Briefing Wkshp3 Moya
17
Sea Based Shore Based
Segment Right Left Right Left Catholic Church 0.845 0.000 0.401 0.000 Displaced Persons 0.596 0.000 0.721 0.117 Illicit Organizations 0.000 0.397 0.000 0.447 Military 0.000 0.408 0.000 0.408 Old Money 0.631 0.000 0.631 0.000 Police 0.564 0.000 0.564 0.000 Urban Middle Class 0.780 0.000 0.184 0.210 Urban Poor 0.798 0.000 0.722 0.211
Sea Based Shore Based
Segment Right Left Right Left Catholic Church 0.845 0.000 0.401 0.000 Displaced Persons 0.596 0.000 0.721 0.117 Illicit Organizations 0.000 0.397 0.000 0.447 Military 0.000 0.408 0.000 0.408 Old Money 0.631 0.000 0.631 0.000 Police 0.564 0.000 0.564 0.000 Urban Middle Class 0.780 0.000 0.184 0.210 Urban Poor 0.798 0.000 0.722 0.211
Cath
olic
Ch
urch
Disp
lace
d Pe
rson
s
Illic
it O
rgan
izatio
ns
Mili
tary
Old
Mon
ey
Polic
e
Urb
an M
iddl
e Cl
ass
Urb
an P
oor
Catholic Church 0.000 0.000 0.000 (0.078) 0.801 (0.078) 0.000 0.000 Displaced Persons 0.706 (0.262) (0.052) (0.230) 0.178 (0.366) 0.480 (0.208) Illicit Organizations 0.207 0.000 0.425 (0.053) 0.420 (0.190) (0.467) 0.183 Military 0.750 (0.351) (0.295) 0.492 0.262 0.685 0.000 (0.257) Old Money 0.474 (0.486) (0.713) 0.000 0.840 (0.506) 0.494 (0.250) Police 0.000 (0.843) (0.174) (0.039) (0.273) 0.767 0.000 (0.218) Urban Middle Class 0.816 (0.623) (0.387) 0.136 (0.269) (0.205) 0.392 0.132 Urban Poor 0.202 (0.259) 0.101 0.225 (0.033) (0.134) 0.552 (0.079)
Second Order Effects
7/9/2008 P-COIN Validation Briefing Wkshp3 Moya
18
What do we expect in the interactions
Earlier iteration (Military)
7/9/2008 P-COIN Validation Briefing Wkshp3 Moya
19
• Not clear from documentation what is being reported
• These multiple influences appear to be captured– Presuming no population
segment strays “too far” from initial state
– Except … Military does!
No MAGTF
AfloatAshore
0
10
20
30
40
50
60
70
80
90
100
0 2 4 6 8 10 12 14 16 18 20 22 24
Months
Popu
latio
n Se
gmen
t Per
cent
age
FARC Pro-FARC Neutral Pro-COIN COIN
0
10
20
30
40
50
60
70
80
90
100
0 2 4 6 8 10 12 14 16 18 20 22 24
Months
Popu
latio
n Se
gmen
t Per
cent
age
FARC Pro-FARC Neutral Pro-COIN COIN
20
Analysis Conclusions (IPR#5)• “Which COA is better?” cannot be answered with much
confidence• “What is the chance that ashore is better than afloat?” can be
answered with greater confidence– More pro-Government sentiment if Marines stay afloat– Lower pro-FARC sentiment if Marines stay afloat– Marine arrival has a polarizing effect (fewer neutrals)– Marine arrival in either case increases anti-Government sentiments of
the Illicit Organizations and the Military• Afloat seems to usually do less harm.
– There is no factor in our influence estimation that BOTH reduces the negative impact of Ashore AND increase the negative impact of Afloat
21
Analysis Conclusions (IPR#5) (cont)
• Because the current Markov chain will eventually return to the same steady state, regardless of MAGTF action, once the MAGTF leaves, we need to consider:– Does the MAGTF commander care about leaving a lasting
impression?– At what point in time do we measure ‘better’?– Pythagoras could change the final steady state as a
function of one or more population segments exceeding or falling below some target value. However, this data was not collected
Validation Conclusions1. The P-COIN simulation fails to capture the dynamic effects
intended in the conceptual model of the insurgency in Colombia provided to the P-COIN developer. That is, P-COIN does not capture the secondary and tertiary effects of the natural drift of population segments between insurgency sectors or the salience between population segments resulting from the influencing event of the MAGTF.
2. The data supporting the P-COIN model is perishable and of low precision. Care should be taken when using the data beyond its origination date; perhaps “warming-up” the Markov chains supporting the data used to build the P-COIN model. Further, the data cannot be deemed valid if an influencing event occurs that would cause the base data used in this simulation to change.
7/9/2008 P-COIN Validation Briefing Wkshp3 Moya
22
Validation Conclusions (cont)3. The P-COIN model should not be used to evaluate long term
effects on the population resulting from the influencing event of the MAGTF arrival.
4. This model and simulation cannot be deemed as predictive of the actual population distributions amongst insurgency sectors in the event that the scenario described in the scenario documentation actually occurs.
5. There is little risk in using the results of the analysis since the analysis does not advocate a change in current Marine Corps procedure. However, item 1 implies that P-COIN also provides little insight into the ashore or afloat question in its current implementation.
7/9/2008 P-COIN Validation Briefing Wkshp3 Moya
23
Recommend: Applying influencers more robustly
What Would Be Useful• Better documentation on the P-COIN instantiation• Time series data• A descriptive walk-thru of results charts (meaning &
implications)• Verification cases (isolated effects) to ensure dynamics have
expected direction (first derivative) and order of magnitude– Descriptions of why we believe it is correct
• Referent– Better explanations of expected resulting effects from data
values• Most had to be inferred• Order of magnitude differences unknown
• Expected interaction effects would be “spectacular”7/9/2008 P-COIN Validation Briefing Wkshp3 Moya
24
Can the Results Be Trusted?
• Without trusting the dynamics– Take caution but …– Recommendation is
innocuous• Under current political
circumstances– No … new data is required
• Can Pythagoras model population dynamics– Probably … more care is required in the instantiation
7/9/2008 P-COIN Validation Briefing Wkshp3 Moya
25
Impact
High
HighLow
Low
Likelihood of Failure
Unacceptable
Very High
High
Some
Acceptable
Low
Negligible
Risk Of Using The ABS
Impact
High
HighLow
Low
Likelihood of Failure
Unacceptable
Very High
High
Some
Acceptable
Low
Negligible
Risk Of Using The ABS
Levels of Validation Process
7/9/2008 P-COIN Validation Briefing Wkshp3 Moya
26
Adapted from Harmon & Youngblood (2005) p. 186
Subjective validation
Objective requirements
Objective results
Objective referent
Automated validation
Initial (level 0)
Subjective (level 1)
Complete(level 2)
Accurate(level 3)
Confident(level 4)
Automated(level 5)
Tier 0, “I have no idea.”
Tier 1, “It works; trust me.”
Tier 3, “It does the right things; its repns are complete enough.”
Tier 4, “For what it does; its repns are accurate enough.”
Tier 5, “I’m confident that this simn is valid.”
SME
SME
Indnt Observer
Indnt Observer
Formal Proof
None SME opinion Single source Multiple sources
Rigorously derived
Correlated with statistical estimates of
uncertaintiesReferent
None
Valid
ated
Analyz
ed
Conce
ptual
Model