Leader-Follower Studies in PMFserv A Performance Moderator Function

Leader-Follower Studies in PMFserv: A Performance Moderator Function (PMF) Unification Architecture Barry G. Silverman, Ph. D Roy Eidelson, Ph. D, Tony Smith, Ph. D Evan Sandhaus, Gnana Bharathy, Benjamin Nye Electrical & Systems Engineering Dept. Inst. For Research in Cognitive Science (IRCS) Center for Human Modeling & Simulation University of Pennsylvania Philadelphia, PA 19104 barryg@seas. upenn. edu © Barry G Silverman, 2005

Outline • AFOSR Simulation Dashboard Project – Leader-Follower Dynamics/What-Ifs • PMFserv Model of Human Behavior • Case Studies – – – Leader. Sim and Athena’s Prism Iraqi Insurgents & Crowds Intifadah (Suicide – mission awry) Somalia (Black Hawk Looters/Crowds) SE Asia – (Tipping to Insurgency) • Lessons Learned © Barry G. Silverman, 2006

Editable* Archetype Group Roles, Assets, Strategies Leader Type*: Elected Tribal Dictator Religious B: eg, secular* Enemy’s Fringe Fs Enemy Loyal Fs Enemy Leader(s) C: eg, fundamentalist* Speech Acts Physical/Political Acts Greivance Membership Real* & Perceived Assets/Resources • Goods • Law/Mil • Popular Support Prospective Followers (and their “leaders”) Sub. Tasks/ Missions* SPIN - Get (buy) support for/against group and action FORM PACT with another group to become more powerful (more of each tank) against C • Contract Terms Recruit Pay (ongoing Militaristic Attack on C, spoils to A, Brag Leader(s) Loyal Followers ‘Rival’ Leaders Fringe Followers A: e. g. , theocrat* Greivance Membership Real* & Perceived Assets/Resources • Goods • Law/Mil • Popular Support Grand Strategy Category* Build, Recon, Move, Attack, Assassinate, . . Block Goods Take Jobs Deny Infrastr Economic War on C, spoils to A, Brag Improve Life for B, Brag Defend Economically: Protect/Secure/Defend, Whine About C Defend Militarily Protect/Secure/Defend, Whine About C Group Values*: Secular Theocrat Fundamental Autocrat Anarchy Negotiate Offer Haggle Abide Violate Give Goods Provide Jobs Build Infrastr Ban Trade, Boycott, Limit Jobs-In. Group Fortfy border, Patrol, Intel, No Privacy

PMFserv’s Unified Architecture for Cognition (Breaking Stovepipes Between Sub-Fields) Stimuli Response Perception Module Expression Biology Module/Stress T - BR = E [ S P * U(st, at) ] t=1 + Personality, Culture, Emotion be free help others support terrorist protect children hide terrorist distract guards crowd together Cognitive be independent sacrifice life survive protect terrorist run for cover Social Module, Relations, Trust block guards vision Memory www. seas. upenn, edu/~barryg/HBMR © Barry G. Silverman, 2006

SJT distal PMFserv Stimuli Gibson Affordance SJT proximal Biology/Stress Response Observations about Other Agents Perception Physiology Updates Simulated World Time Pressure Action Choices Afforded CCT Physiolo- Energy main fatigue Inte. Conflict event gy Tanks Tank stressor Theory grated • sleep tanks time Stress i. Stress (Janis. Mann) Status Coping Style Game DT Alternate Prosp Decision Theories Heur ELM Decisionmaking BR EV SJT central Value Trees (GSP, Bayes) Negative Emotions (event stress) GSP Cog. App. Node (OCC) Fail/ Succeed Updates SEU Emotions (11 pairs) Subj. Relationship Tanks Utility (Damasio) Intention Management Relationship Parameter Levels SEU Candidate Action Current World State: GSP Leaf Node Affordance Updates Emotions (11 pairs) Action Choice (physical, speech) Coping Style • nutrition • injury Expression Relationship Parameter Levels • Alignment • Credibility/Trust • Objectification • Valence Personality, Culture , Affect Identity Nested Repertoires Intent • Demography ionality • Social. Group Proc’g • Role Social Module Memory LEGEND: Implements Literature Interprets Literature New PMF Validated Profiling Instrument Available (Hermann, Eidelson, Hofstede/House, Nf. C) BR -- Best Response CCT – Cognitive Continuum Theory(Hammond) DT – Decision Theory (Keeney, Raiffa) ELM – Elaboration Likelihood Model (Petty) EV -- Expected Value Game-Game Theory (von Neumann, Nash) GSP – Goals, Standards, Preferences Heur – Heuristics, Rules, Biases (Simon, Klein, Slovic) OCC – Ortony, Clore, Collins Prosp –Prospect Theory(Kahneman & Tversky) SEU – Subjective Expected Utility(Edwards, Wright) SJT – Social Judgement Theory(Brunswik, Hammond)

PMFserv Incorporates Personality, Cultural Values, Utility (common math framework – subjective expected utility) 11 pairs of emotions internally-derived utility U = S Ix(sk)/11 GSP Trees (Bayesian-weighted) • Preferences - longer term hopes • Standards - means acceptable in self and others • Goals - short term needs and actions to reach Prefs

GSP Trees – Leader Hermann TRAIT 1. belief in control 2. need for power 3. conc. complex 4. self-confidence 5. task vs. relations 6. distrust 7. In-group bias © Barry G. Silverman, 2006

GSP Trees – Villager © Barry G. Silverman, 2006

Many Parameters (qi) in PMFserv • Physiology/Biology – – – Nourishment Muscle Energy Injury Levels Sleep Need Adrenaline Others (open to user) • Stress/Coping Style – – Time Pressure Event Stress Effective Fatigue Decision Style (5 levels) • Adherence • Vigilance • Panic • Emotions (11 pairs) – Joy/Despair, Fear/Hope, etc. © Barry G. Silverman, 2006 • GSP Value Trees (10 E 2 nodes) – Long Term Preferences (by Resource and Territory) – Standards (Norms, Doctrine) – Short Term Goals (Maslow-type) • Relationship Parameters – – Alignment Level (Ally-Foe, 5 levels) Group Affiliation (6 -10 groups) Valence/Cogn. Unit/Agent-Object Trust (by Resource and Player) • Decision Parameters – – Utility and Cost (continuous) Action Choices (10 s to 100 s) Discount Factors (risk-prone/averse) World State • Perception & Modeling of Others – N*(GSPs+Relations+Actions)

Gallery of Some Past PMFserv Agent Studies Asymmetric Plots (Culture/Emotions) § Recreate Black Hawk Down: Four types of Somalians § § § Crowd Behavior Emergence (Bio-Affect-Values-Panic-Riot) • WTO Talks in Seattle -- Protesting/rioting crowds: Males (employed/unempl. ), females, instigators • Rioting/looting crowds at police station (impact of chanting upon crowd behavior) • Soccer Hooligans (Manchester United Supporters) • Scale up to 2000 agents in Sony Open. Steer Political Agents for RPGs • Nested intentionality, speech acts, reputation management • World leaders in diplomatic strategy role playing game • Third Crusade Leaders (Saladin, Emir, Richard, Philip, etc. ) Women/Kids, Civilian Males, Militia, Clan Leaders Intifadah dynamics – cell leader, suicide-bomber, Mayor, populace reactions Grade B Movie - Al Qaeda & Iraqi Insurgency, SE Asia, Elsewhere

Scope of Leader. Sim Prototype Territories (3) Resources (3) Actions (5) Total (3 x 3 x 5 xpayment levels) x no. of plies x N leaders (3) References www. seas. upenn. edu/~barryg/HBMR. html • . Silverman, BG, Rees, R. , Toth, J, et al. , (2005, Jan). “Athena’s Prism – A Diplomatic Strategy Role Playing Game for Generating Ideas and Exploring Alternatives”, 1 st Internat’l Conf on Intel Anal • Silverman, B. G. , Johns, M. , Bharathy, G. (2004, August). “Agent. Based Simulation of Leaders. ” ACASA/UPenn, Tech Report. • Silverman, B. G. , Johns, M. , et al. (2002, May). “Constructing Virtual Asymmetric Opponents from Data and Models in the Literature. " 11 th BRIMS, SISO. © Barry G. Silverman, 2006 Scale. Up Territories (10) Resources (10) Actions (70) Total (10 x 70 xpayment levels) x no. of plies x N leaders (10)

Agents Form Beliefs about the GSP Trees of other Agents (Static Model: ‘Mirroring and Stereotyping) © Barry G. Silverman, 2006

Prototype Leader. Sim Results Nash Equilib: 2 winners in conflictual world Rare 3 in endgame. Yellow specialized away from Red and Blue. Using threats, Yellow turns Red and Blue against one another Y’s power is curtailed early. Y then uses treaties to negotiate peace.

Utility vs. Time Impression Error vs. Time Power vs. Time Preference Achievement vs. Time

Athena’s Prism strategic role playing game • Useful to rapidly mockup conflict scenarios and test what-i • PMFserv can simulate world leaders if not enough humans © Barry G Silverman, 2005

Big. Wig (late ’ 06)

Iraqi Crowd Demonstration Scenario: PMFserv drives Crowds in Big. World© Squad leader Community leader Crowd Agitator Partners: DARPA, DSO BBN Total Immersion © Barry G. Silverman, 2006

Demonstration • Exercise 2 paths through the scenario – Naïve participant – Experienced participant • Interact with 2 crowd variations – Moderate crowd – Extreme crowd • Observations – Crowd animations – PE internal states • Focus on time-critical decision making © Barry G. Silverman, 2006 Sample outcomes (crowd states) Moderate crowd Extreme crowd Naïve player Hostile Violent Suspicious Misled Experienced Trusting player Satisfied Hostile Suspicious

Follower Got Most Utility From Killing Himself and Civilians at Bank Churc Bank Military Facility. City Hall Checkpoint © Barry G. Silverman, 2006 Partners: DMSO, GM Terrorist agent Sporting Event

Leader Wanted Bank to be Robbed Church Bank Military Facility City Hall Checkpoint © Barry G. Silverman, 2006 Terrorist agent Sporting Event

SOMALIA Suicide Bomber (PMFserv) Militiaman with Female Shields (PMFserv) Civilian Chopper Looters (PMFserv) Civilian Crowd (PMFserv) Helicopter Crash Site Militia Unit W/ Leader (AI Implant) Partners: DMSO ONR ICT Soar Suicide Bomber (PMFserv) Militiaman with Female Shields (PMFserv) Start

PMFserv drives all crowds at Mike Durant’s Crashed Helicopter in Bakara Market, Somalia • Observers • Women • Looters • Militia • Suicidists © Barry G. Silverman, 2006

PMFserv Ramp Up Testing Leader Type*: Elected Tribal Dictator Religious Greivance Membership Real* & Perceived Assets/Resources • Goods • Law/Mil • Popular Support Leader(s) Loyal Followers ‘Rival’ Leaders Fringe Followers A: Budhist/TRT Party* Greivance Membership Real* & Perceived Assets/Resources • Goods • Law/Mil • Popular Support Prospective Followers (and their “leaders”) B: Muslim* Enemy’s Fringe Fs Enemy Loyal Fs Enemy Leader(s) C: eg, fundamentalist* TRT Leader Model Tahksin Shinawatra Prime Minister Budhist Majority Patani Moderates • Sultan • Parents (Villagers) • Unempl College Grad Bersatu or BRN Insurgents: • Leader Model • Follower GSPs • Jemaah Islamiyah Group Values*: Secular Theocrat Fundamental Autocrat Anarchy © Barry G. Silverman, 2006 Partner: LM/ATL

Evidence of Escalation of Violence Transformation of society: people occupying higher grievance states increases. Actual Event Data Simulated Villager Feelings Training Data Set Tsunami (relief forces) © Barry G. Silverman, 2006

Snapshots of Key Simulation Events • In the beginning, … the villagers do not have significant reaction to the Country Leader. • The leader initiates an overt reaction of discrimination as it happened in early 2004. • Conservative villagers disagree, while radicals oppose the moves. © Barry G. Silverman, 2006

Sensitivity Analyses of Follower Outcomes vs. +/- 15% mean Leader Personality Standard

AFOSR Simulation Dashboard Status • Find principled way to explore the simulation space – White Paper on possibilities for conflict, non-conflict • Understand how model parameters (qi) influence outcomes – Manually Testing – Ho: P(Conflict | qi ) > Threshold OR < Limit – Parameter elasticities (e. g. , regression estimators) • Create wizard for Policy Analysts (dashboard) – XML-RPC interface to Monte Carlo shareware – Adding a module for design of experiments with PMFserv’s leader and follower parameters

Monte Carlo Integration & Interface Subscribable Input Server for PMFServ For (param, val) in Input. Stream: set. Property() Execute PMFServ. Model Loads Tokens Interrupt Runs Publishing Client Receives Output from PMFServ Publish Results SIMLAB 2. 2 Pre-Processor Module Generates Input Stream SWIG Interface PMFServ Integration Layer (Publish & Subscribe ) Executor Module • Starts external PMFServ • Sends Input. Stream • Receives Output. Stream • Halts PMFServ • Sends Token for Loading Intermediate Results (Sequential Sampling) Post-Processor Module Design of Experiments Dashboard Parameter Experiments Input Interface (XL) • Param. Name, Central. Value, Std. Devn, Distribution • DOE Specification • Inspection, Tracing Flows in the Model, Back-Chaining (large parameter set to be reduced) – as done in villager • Statistical screening or direct experiment using Latin Hypercube, Morris Random Walk, etc. Model Output UI Interface (XL)

Summary • Leader. Sim – Rapidly mockup realworld scenarios and play out how policies & action choices lead to alternative effects and ways to influence leaders • Human Behavior (PMFserv) – Compose leaders and followers. Open the agenda to research on parameters across many human behavior disciplines (biology/stress, values/personality/emotion, culture/groups, trust/reputation, decisions/gaming) • Sensitivity Studies – Find principled ways to explore the space of possible outcomes, to avoid conflict states, and to understand the elasticities of behavior parameters as interventions are attempted