The Nexus Cognitive Agent Social Models

1. The Nexus Cognitive Agent Social Models Deborah V. Duong, Ph.D.

2. Nexus • A project of OSD/PA&E Simulation and Analysis Center • Nexus is a set of Cognitive Agent Social Models • Cognitive • Perception is modeled • Agent • Individuals act autonomously according to perceptions of utility • Social • Agent perceptions and actions form social structures • Model • A system forms • The system represents generally accepted social phenomena • Nexus can be instantiated to a particular scenario for analysis • Two models presently exist in Nexus • Nexus Schema Learner • Group Level Agents the model Coherence in Social Psychology • Nexus Network Learner • Individual Level Agents that model Resources flowing through Social Role Networks

3. Composition Needed in Social Models • To analyze a social environment in quick turnaround fashion • Instead of having the world in a single simulation, choose several small, general, building block models from a library to build a scenario • Different social theories may be switched in and out • Necessary for objective treatment of conflicting theories • Models need to resolve conflicts • One way to resolve them is Adaptive Simulation • Models adapt to each other, and adapt to the data • To make a coherent picture of the social environment

4. Nexus Models are Made for Composition • Generality • Theory is generally accepted and universally applicable • Nexus Schema Learner • Coherence and Cognitive Dissonance Theory • Nexus Network Learner • Social Role Networks • Resource flows based on utility • Both model Interpretive Social Science • Easily Instantiated to Particular Scenarios • Nexus Schema Model • Used in 3 different studies of IW, both Kinetic and Non-Kinetic • Nexus Network Learner • Used to model corruption in the Africa Study • Could be used to model the flow of funds through Terrorist Networks • Flexibility • Nexus Network Learner easily ingests and adapts to data in a general manner • Real world data, initialization and live connection • Data from other models in a composition • Conflict Resolution of composed models through Adaptive Simulation

5. The Social Theory Behind Nexus Interpretive Social Science

6. Scientific Rigor in Social Science • Thin Description Vs. Thick Description • Thin Description: methods that find patterns in behavior without referring to the meanings of behaviors to the actors • Objective • Behavioralism, statistical techniques, Etic anthropology, Social Network Analysis • Criticized for ignoring crux of the issue • Thick Description: methods that look at the context-dependent meanings which motivate actors • Subjective • Interpretive Social Science, Emic anthropology, verbal analyses • Radical Version: No generalizations can be made • Criticized for lack of rigor • Computational Social Science • A new tool for rigorous thick descriptions • Objective descriptions of subjective phenomena • Meaning and context modeled • Relevant generalizations made and patterns found NEXUS models Interpretative Social Science

7. Introduction to Interpretive Social Science • People’s thoughts matter • “Perception is reality” – Kilcullen • Meaning is context-dependent • “Anything can mean anything” – Goffman • Social structure comes from people’s interpretations of their situations • To change social structure, you must change subjective interpretations • Agents may be quite pragmatic in their interpretations • Interpretative Social Science exists across all of the social sciences. Nexus uses: • Institutional Economics • Symbolic Interactionist Sociology • Narrative Paradigm in Communication Theory

8. Interpretive Social Science Used in Nexus • From communications theory: The Narrative Paradigm (Fisher) • People think based on Narratives, or stories with characters in roles, that have historical continuity • People think based on Narrative coherence, with the belief that persons behave consistently. • For example, if they believe Islamic extremist rhetoric, they believe Bush caused 9/11, despite evidence to the contrary • People think based on Narrative fidelity, looking to see if the explanations fit their identity and their values. • For example, they may have a historical consciousness and a national identity, in which they play a respectable role • From economics: The New Institutional Economics (NIE) (North) • Institutions (Social and Legal Norms and Rules) underlie economic activity and constitute economic incentive structures • Institutions come from the efforts of agents to understand their environment, so as to reduce uncertainty, given their limited perception • When some uncertainties are reduced, others arise, causing economic change • To find the leverages to corruption, NIE would look at actor’s definition of their environment, and how this changes incentives and thus institutions • From sociology: Symbolic Interactionism (Mead) • Roles and Role Relations (such as in trade roles and trade relations) are learned, created during the display and interpretation of signs (such as gender, ethnicity, and other demographic characteristics) • Institutions (rules of trade) are commonly accepted interpretations of symbols, that start out as a subjective perception and engrained in society as an objective rule.

9. Interpretive Social Science is needed for modeling Irregular Warfare and Corruption • Irregular Warfare analysis is concerned with subjective concepts such as “legitimacy,” “will,” and public support for the actions of governments and insurgents • In information operations, the interpretation of actions is critical • People don’t accept messages just because they see them over and over: interpretation is important • However, In most simulations, meaning is given, and interpretation is assumed away • To find levers to corruption, one must recreate dynamic institutional change • To understand institutional change is to understand changes of norms, roles and role relations, the subjects of symbolic interactionism • Symbolic interactionism can generate new rules for Kinship, Social, and Patron- Client Role relationships • The New Institutional Economics can depict changes in incentive space as a result of new role relationships, laws, resource rents, and aid, needed for understanding corruption.

10. All Nexus Agents are Cognitive Agents • IW study: How does a population interpret an action? • Agency and Cognition are needed to simulate the interpretation of a situation • IW study: How do we change social structure? • Cognition is needed for agents to learn new norms, and in doing so develop new social structures that combat problems like corruption • Kilcullen emphasizes the importance of coevolution for IW • Coevolution implements feedback in agent-based simulations • Coevolution requires learning, so agents must be cognitive

11. Scientific Rigor in Agent Based Modeling • We are Analyzing, not Describing • We want to avoid putting the answer to the question in the question • Results can not follow directly/obviously from assumptions • Assumptions should be hard-to-argue, at crux of the problem • Primordial soups preferred: • Few assumptions to start • Many known-to-exist patterns emerge • Ockham’s razor makes it likely to be true • If it is general, it can be applied in new situations without adjustment: separation of testing set from training set • Emergence in agent based simulations have good epistemology • If the model is incorrect, you can tell • You cant just add another node and over-fit the data • You have to explore the reason for a pattern, because you didn’t tell it to have that pattern in the first place • Some methods are transparent because their structure is predetermined by the model, not computed from assumptions • We are walking out the implications of the assumptions • Cognitive Agents that use methods of soft computation do not presume categories or group perceptions according to pre-conceived notions: perception is emergent • We may have approximate or faulty data • Cognitive Agents that use methods of soft computation are robust with respect to data

12. Review of Nexus’ Algorithms

13. Nexus Cognitive Agent Types • Schema Learner: Social Group Agents simulate inter-group dynamics • Agents represent social groups • Popular Support between groups is computed • Each agent has an individual Neural Network • Used in two previous DoD studies • Network Learner: Social Role Agents simulate intra-group dynamics • Agents represent individuals • Conditional Probability Tables read in that have population characteristics and behavior frequencies • Corrupt behaviors categorized by whether stealing or bribing, and who is doing it • Bayesian networks generate data and are used to simulate development of social role networks and frequencies of behaviors: coevolving genetic algorithms allow them to learn new behaviors and different new choices.

14. Nexus Schema Learner

15. Nexus Schema Learner uses the Narrative Paradigm to model Popular Support • A Social Group Agent for each ethnic, insurgent, or government group • Social Group Agents interpret and reinterpret actions in context • Of the actions that have happened before • Of ideological similarity to the actor • Of the network of support that agents have for each other • In the light of new actions, Social Group Agents can • Reassign blame for past actions • Change belief in the trustworthiness of other agents • Declare new support levels for other agents • Social Group Agents compute support levels with narrative-rational minds • They assign blame, belief in trustworthiness, and support so that the other agents have consistent behavior patterns (Narrative Coherence). • They assign blame, belief in trustworthiness, and support as it supports their own identity and values (Narrative Fidelity). • Social Group Agents have a historical consciousness and reinterpret blame for actions

16. Social Group Agents Each agent has a “group mind” (or you can say… it’s the leader’s mind). This “group mind” fits in with SME statements of group opinion. This group mind is a constraint satisfaction neural network, used to compute group support of other agents Constraint satisfaction neural networks simulate coherence Duong, Deborah Vakas. A System of IAC Neural Networks as the Basis for Self-Organization in a Sociological Dynamical System Simulation. Master’s Thesis, The University of Alabama at Birmingham, June 1991. http://www.scs.gmu.edu/~dduong/behavioralScience.pdf. (published 1995) Thagard, Paul. Coherence in Thought and Action. Cambridge Ma: MIT Press 1999. Read, Stephen and Lynn Miller. “On the Dynamic Construction of Meaning: An Interactive Activation and Competition Model of Social Perception,” Connectionist Models of Social Reasoning and Social Behavior. London: LEA1998.

17. Social Group Agent Simulation Loop All groups compute their support for all other groups Group A computes support for group B, with neural net, based on: the past actions of group B what other groups group B supports how close group A is in ideology to group B The neural net takes into account not only the historical actions of a group but the action of their friends, and their friend’s friends, etc., into a coherent picture that looks at higher order relations The enemy of my enemy is my friend: from Heider’s balance theory Cognitive Dissonance when triad rules are broken All groups make public declarations of support. There is a hook for adding “believed support” if it differs from public declarations. Groups modify their neural nets according to their new beliefs of public support. Groups act based on their objectives Groups re-compute support, etc.

18. Each Social Group Agent Mind is a Boltzmann Machine Neural Network

19. The Boltzmann Machine A Constraint Satisfaction Neural Network Nodes may represent states of the world Links may represent how much evidence of one state supports evidence of the other state An Ising Spin model, similar to the Hopfield Network Stochastic: Uses Simulated Annealing Nodes are turned on randomly at first Each node computes its activation based on (activation of nodes linked to it X weight at synapse) Randomness dissipates slowly so that network can settle on more consonant states Represents a Paradigm of consonant states Nodes compute activation over and over until all converge on a steady state Settled upon state represent an internally consistent set of evidence, and a consensus of the evidence on the state. Can represent a paradigm shift, a change in belief in who is responsible for the group’s problems

20. Example: The Necker Cube Simon, Dennis. “The Boltzmann Machine Necker Cube Example” http://www.cs.cf.ac.uk/Dave/JAVA/boltzman/Necker.html

21. Architecture of Nexus’ Neural Networks Columns are social groups First Row: Support Nodes Second Row: Trust Nodes Remaining Rows: Blame Nodes for Salient Historical Events One event per row Ordered: bottom are future events Evidence Node Court-type “Proof” they’ve witnessed, before the minds spin Ideology Node (not pictured) Ideological similarity of groups (connects to the trust nodes)

22. Nexus’ Social Group Agents act according to the Narrative Paradigm Social Group Agents seek a coherent picture of events Agents may reinterpret the blame for events in the past depending on present beliefs Agents choose alliances according to a coherent picture Coherence includes consonance with historical events, in an ongoing story Higher order vulnerabilities in support network may be brought to light Useful in campaigns to strengthen or weaken support Works by way of cognitive dissonance theory, or the minimization of incoherent facts Nexus has lots of explanatory power for a variety of Irregular Warfare Scenarios For example, Insurgent-agents may incite Rodney King incidents, make videos of it and distribute it, planned by looking ahead to what actions would advantage them. They choose Rodney King because they see that to keep an alliance with population, government agents would be forced to prosecute police, separating police from government. Groups can represent countries and Gandhi type protests can represent actions that cause them to try to keep ideologically similar to their allies, so their allies will trust them. Ghandi’s tactics wouldn’t work for Nazis, because the Nazis allies were outwardly brutal It could model why an ideology that is dissonant may become popular, for example, a Nazi insurgency succeeding in a civilized country. Humiliation by enemies after WW1 was so strong, and Germany wanted to preserve superior identity so much, that moral dissonance could be minimized.

23. UNCLASS Example: MCCDC Study • Data from country Subject Matter Experts (SMEs) • Ideological similarity estimations for ideological node input into trust nodes • Initial support levels between groups and for the government • 22 relevant historical events in evidence node input to blame nodes (for past events) • “Indirect or Direct” Data, SME predictions of group reactions as evidence node input to blame nodes (for future events) • Data elucidated by MCCDC through a technique that assigns a numerical value to SME vocabulary • Popular support derived from vocabulary based on values from Affective Control Theory data (a combination of emotional valiance and power) • Study measured ten social group’s interpretations of US direct action vs. indirect action to help their country in a natural disaster, and how it affected their support for the government • Groups were displaced persons, the urban poor, the urban middle class, old money, illicit organizations, the police, the army, the church, the government, and the insurgency

24. MCCDC Study Results • All groups except displaced persons were for the government and against the insurgents • However, the structure of their support for each other combined with historical events made the government somewhat vulnerable. • All groups except the old money and the displaced persons changed their attitudes slightly more towards the insurgency when the US helped (while still supporting the government). • It did not matter very much which kind of help. • None of the groups, except for the urban middle class, had any different support levels for the government or the insurgents when direct and indirect action was compared. • Only the urban middle class liked the insurgents a little bit more when the US action was direct than when it was indirect.

25. Nexus Network Learner

26. Nexus Network Learner: Generality • Nexus Network Learner social theory is general • Input to the simulation is general • Bayesian Network of traits and behaviors • Social Role Network that defines flows of resources • Nexus Network Learner learns in a general manner • Arbitrary resource allocation or network choice behaviors may be learned • Learning based on pragmatic utility • Good for many IW Warfare Scenarios • We often want to effect social structure change

27. General Input Data: Traits and Behaviors • A Bayesian Network that holds the demographic characteristics of a population • Determined traits such as gender, ethnicity, age • Situational traits such as employment • Behaviors such as propensity to steal or bribe • Attributes looked for in the choice of network partners, such as “look for ethnicity in employment”

28. General Input Data: Social Role Networks • Networks, such as Kinship, Trade, and Bureaucratic • Roles, such as Wife, Customer, Employee that belong to the networks • Role Definitions • Corresponding role ,such as Parent and Child • Criteria for choosing a role, such as to choose a wife you must be a male of working age • Distribution for the number in a role, such as a Parent has a mean of 4 children and a standard deviation of 2 • Derived roles, such as a Mother is a female parent • Accounts that belong to roles • Roles are responsible for accounts, such as the Employer is responsible for distributing the Payroll, or the Head of Household is responsible for distributing the Household Budget • To distribute to another network, an agent may switch roles, such as an Employer in the Bureaucratic Network becomes a Purchaser in the Trade network • Allocation distributions to other accounts, depending on traits, such as Stealing • Exogenous accounts such as Foreign Aid • How each transactions effects utility of individuals

29. General Learning of Network Structure and Behaviors • Which attributes to keep constant and which are learned are also input to the simulation • For example, Gender and Ethnicity are constant for an agent, but Stealing is learned • Agents start out with the behaviors that the Bayesian Network gives them, from country data • Network choices based on attributes • For example, choosing an employee based on whether a bribe is offered based on ethnicity • Behaviors based on attributes • For example, tendency to steal based given other traits like location, employment, income, etc. • During the Simulation, networks are built and funds flow • Networks attrit and learned network choices are made • Funds are allocated through networks based on learned behaviors • The utility of choices are kept track of • Own or role relation utility is an input • Agents learn other network choices and behaviors based on utility • Many new sets of behaviors and network choices are tested until the best are found

30. Performing Tests with Nexus Network Learner • A wide variety of tests relevant to IW may be performed • For example, new network formations and behaviors may be tested based on… • The effect of different utility functions • For example, make agents care only for self rather than larger social network • The effect of different penalties • For example, a penalty attribute that encodes different fines • The effect of different exogenous resources • For example, test resource rents or foreign aid • Monte Carlo methods reveal if new structures are the result of different CONOPS • Nexus is stochastic and can give a confidence interval

31. Nexus Network Learner Flexibility • As each agent learns, all the agents coevolve, making them very adaptive • Every agent has its own private learning algorithm • Their behaviors effect the larger social structure and the larger social structure effects their behaviors • Micro-Macro Integration is modeled • They can adapt to data from other simulations and to initial country data as well • The learning algorithm in each agent makes the adaptation to data flexible • BOA (Bayesian Optimization Algorithm) can start learning from initial data • In the calibration phase, agents to adapt to initial data, so that they generate it though their perceptions and motivations • Thus they “explain” the data, going from correlation to cause • This greater ability to ingest data also allows them to meld with other simulations in a composition • Together, composed simulations create a coherent picture of the social environment • Conflicts are resolved through mutual adaptation

32. Open Source Software • A wide variety of open source software help Nexus Network Learner to adapt and be general • Weka Data Mining Toolkit: • Bayesian Networks • Easy entry of country data through Graphical User Interface • Any data can be easily converted to Bayesian Networks • Visualization of Statistical Data • Repast Agent Based Development Toolkit • Easy Entry of Agent Behaviors • Visualizations of Social Networks using Jung

33. Use Case: Modeling Corruption • Corruption in African society is said to be the result of conflicting social networks • Kinship Network vs. Bureaucratic Network • Kinship Network has patron-client roles and many obligations • Bureaucratic Network has merit-based impersonal roles • Rules on how to choose network relations (based on merit?) and how to distribute resources (based on kin obligations?) differ • Nexus Network Learner easily modeled Kinship, Bureaucratic, and Trade Networks • Nexus Network Learner easily represented the roles and role relations of both networks • Includes 65 roles, including roles and role relations important to Matrilocal and Patrilocal ethnicities • Nexus Network Learner could easily represent the corrupt behaviors which change distribution of resources in the networks • There are said to be eight basic types of corruption, that Nexus models with basic stealing and bribing behaviors, that occur in different sectors, and during both network choice and transactions • Nexus models the moving of funds from Bureaucratic to Kinship networks based on behavior traits • Ways to move from the rules of Kinship to the Rules of Bureaucracy may be explored • Different penalties, utilities, and exogenous funds may be entered to explore the effects of DIMEFIL actions

34. Summary • Nexus is a robust, general, flexible tool for modeling Social Cognition, Social Role Networks and resource flows through those networks • Usable in a wide variety of IW scenarios • Nexus can get to the crux of the issues in IW because it models agent motivations and interpretations • Nexus easily ingests real world data because it can adapt to it • Nexus works well in a composition of other simulations because it can adapt to them • Nexus offers Monte Carlo exploration of the effects of DIMEFIL actions on Social Structure