Agent-Based Modeling of Conflict - ETH Z · § Agent-based models • Cognition based on behavioral...
Transcript of Agent-Based Modeling of Conflict - ETH Z · § Agent-based models • Cognition based on behavioral...
||
Prof. Dr. Karsten DonnayUniversity of Konstanz
Agent-Based Modeling of Conflict
125.11.2019Karsten Donnay
||
Outline
§ Introduc-on
§ ABMs in Conflict Modeling
§ Examples
§ Taking a closer look: Jerusalem
225.11.2019Karsten Donnay
||
Why do we fight?
§ Conflict as the result of irrationality• Human nature• Evolutionary perspective• Psychology
§ Conflict as the outcome of calculated, rational decisions• Bargaining failures• Arms races• Power shifts• Violence as a last resort of political struggle
25.11.2019 3Karsten Donnay
||
§ Classical: two opposing views• Hobbes: Endemic `warre’
life is `poore, nasty, brutish, and short’(Leviathan, 1651, 13)
• Rousseau: Harmonious livingAgriculture, demographic growth and privateproperty are what brought war
§ Human nature• Competition for food and water• Sex and reproduction, competition over partners• Rank, status and honor• Revenge and (blood) feuds
425.11.2019Karsten Donnay
Conflict as the result of irrationality
||
§ Evolutionary Perspective
• (Darwinian) struggle for survival against environment and each other• “Selfish gene”: evolutionary reward to protecting own kind• No perpetual conflict because others are strong, i.e., risks and costs
of (permanent) conflict are evolutionarily counterproductive
§ Psychology
• Frustration-Aggression: aggressiveness is produced by frustration andfrustrations high when discrepancy between expectations and reality
• Redirected aggression: towards minorities, foreign nationals, anyonethat is different than the in-group
• Authoritarian personality: poor self-image, identification with the aggressor
525.11.2019Karsten Donnay
Conflict as the result of irraKonality
||
§ Bargaining failures• Conflict is costly, should therefore always be able to strike a deal,
especially because it is inefficient, i.e., losing party always regrets
• Possible causes for bargaining failure
§ Indivisibility (e.g. holy sites in Jerusalem)§ Private information and misrepresentation§ Shifts in power through economic/social change§ Systemic uncertainties
§ Violence as a last resort of political struggle• Organized political struggles may not always achieve their goals,
especially against much stronger opponents (asymmetric conflict)
• Terrorism as a means of continuing the fight with other meansin order to achieve political goals
625.11.2019Karsten Donnay
Conflict as the outcome of calculated, rational decisions
||
§ Game Theory• Actors are strategic: forward and backward-looking
• Actors draw the best inferences given available knowledgeand constraints
• Maximize their payoffs given a set of preferences
§ Advantages: • Clarity, tractability
• Formal proofs
§ Disadvantages• Cumbersome to solve
• Strong and oFen simplisGc assumpGons aboutcogniGon and interacGons
725.11.2019Karsten Donnay
Modeling rational actors
||
§ Agent-based models• Cognition based on behavioral heuristics not necessarily
strategic considerations
• Large number of agents and complex systemic setupsthat can be very realistic depictions of conflicts
§ Advantages• Ease of use
• Ability to model large number of agents and includemany parameters
§ Disadvantages• Lack of transparency and potential for parameter overload
• Difficult to infer the core mechanism
825.11.2019Karsten Donnay
Modeling rational actors
||
Outline
§
§ ABMs in Conflict Modeling
§
§
925.11.2019Karsten Donnay
||
ABMs in Conflict Modeling
§ Why we need quantitative models of conflict:• Need better quantitative understanding of past and contemporary conflicts
(e.g. Afghanistan, Iraq) for informed policy decisions
• Patterns in empirical data ≠ understanding!
• Challenge: bridge gap between observed empirical conflict patterns andestablished theories of conflict
• In particular: connect theories of (individual or group) interaction withobserved macro patterns
1025.11.2019Karsten Donnay
||
ABMs in Conflict Modeling
§ Empirical patterns - Afghanistan
Bohorquez et al. Nature 462 (2009) Johnson et al. Science 333 (2011)
Casualty sizes Inter-event times
1125.11.2019Karsten Donnay
||
ABMs in Conflict Modeling
§ What we require of quan8ta8ve conflict models:
• Model parameters: observable or related to observable variablesand quan8fiable
• Models: applicable to a par8cular case but general enough to betransferable to a broader context
• Models should account for specifics of the country(popula8on, size, ethnicity etc.)
• BUT over-fiJng must be avoided!
1225.11.2019Karsten Donnay
||
ABMs in Conflict Modeling
§ Challenges of interdisciplinary approaches:• Natural science experience in modeling complex systems provides
unique insights for conflict research
• No impact on sociological or political science research if a modelthat “fits” empirical observations is NOT consistent with establishedtheories of conflict dynamics
• Quality of a model as a function of fit to empirical observations andplausibility of its mechanisms
1325.11.2019Karsten Donnay
||
ABMs in Conflict Modeling
§ Why use ABMs in conflict research• Political science: mainly regression analysis in quantitative studies
• Recently: agent-based models (ABMs) for detailed studies of specificconflict scenarios
• ABMs: simplified representation of empirical scenario
• Very useful where formal models fail
• Specifying and validating ABMs: use detailed empirical data to increaserealism and validity
1425.11.2019Karsten Donnay
||
ABMs in Conflict Modeling
§ Specifications of conflict ABMs
• Computational agents = relevant actor groups
• Microscopic interaction dynamics based on established theories
• Dynamics: who interacts with whom and when
• Simulation outputs: results of agent interactions(attack, migration etc.)
1525.11.2019Karsten Donnay
||
ABMs in Conflict Modeling
§ Empirical Data and Validation of ABMs• Empirical population data: realistic agent population and topography of
model landscape
• Simulation outputs: quantitative comparison to empirically observeddistributions of conflict indicators
• (Microscopic) Interaction mechanisms: qualitatively, ideally quantitatively,consistent with the empirical observations
• Test predictive power:§ In-sample or out-of-sample predictions§ Comparison to “null models”
161625.11.2019Karsten Donnay
||
Outline
§
§
§ Applications and Examples
§
171725.11.2019Karsten Donnay
||
Applications and Examples
§ Epstein’s model of Civil War
• Epstein, PNAS 99 (2002)
• Agent-based model of civil violence:
§ Central authority vs. decentralized rebellion
§ Central authority suppressing violence between two ethnic groups
• Main mechanisms:
§ Grievance leads to violence
§ Private grievances ≠ public actions
§ Police intercepts attacks with certain probability
1825.11.2019Karsten Donnay
||
Applica'ons and Examples
§ Epstein’s model of Civil War
Spatial outbursts
Epstein PNAS 99 (2002):7243-7250
Temporal outbursts
1925.11.2019Karsten Donnay
||
Applications and Examples
§ Epstein’s model of Civil War
http://ccl.northwestern.edu/netlogo/models/Rebellion2025.11.2019Karsten Donnay
||
Applications and Examples
§ Epstein’s model of Civil War§ Reproduces some stylized facts of revolutions/communal violence
§ BUT all model parameters unobservable
§ Interaction mechanisms somewhat arbitrary
§ Some results due to particular implementation (for details see http://ccl.northwestern.edu/netlogo/models/Rebellion )
§ No formal validation of results or mechanisms
2125.11.2019Karsten Donnay
||
Applications and Examples
§ Severity Size Distributions – Bohorquez et al. (2009)• Bohorquez et al. Nature 462 (2009)
• Agent-based model of insurgent violence:§ Security forces vs. insurgent groups§ Casualties in attacks scale with a group’s (military) strength
• Main mechanisms:§ Interaction probability scales with group strength§ Coalescence and fragmentation of groups as a consequence of
interaction§ Relative strengths of groups decide casualties in attacks
2225.11.2019Karsten Donnay
||
Applications and Examples
§ Severity Size Distributions – Bohorquez et al. (2009)
Bohorquez et al. Nature 462 (2009)
Size of events
2325.11.2019Karsten Donnay
||
Applications and Examples
§ Severity Size Distributions – Bohorquez et al. (2009)
• Excellent fit for a wide range of cases (4 fit parameters)
• Model parameters have empirical interpretation
• Problem of the modeling approach:assumed insurgent/military group dynamics NOT empirically plausible
§ Not realistic that military units fragment after insurgent attack§ Unsubstantiated that group fragmentation and coalescence is main driver of
insurgent/military group strength§ Empirically insurgent groups very often attack when own strength much smaller than
that of target: model predicts quick eradication of insurgency contrary to empirical situation
2425.11.2019Karsten Donnay
||
Outline
§
§
§
§ Taking a closer look: Jerusalem
2525.11.2019Karsten Donnay
||
Taking a closer look: Jerusalem
§ Violence in Jerusalem• Highly contested urban space
• High symbolic importance for theArab world
• Limited understanding of conflict mechanisms
• Only few quantitative studies
• Inform policy decisions?
2625.11.2019Karsten Donnay
||
Taking a closer look: Jerusalem
§ Agent-based model• Interactions of main population groups
§ Secular Jews§ Ultra-Orthodox Jews§ Palestinians§ Security Forces
• Realistic topology of Jerusalem seeded withempirical population data
• Model mechanisms based on Jerusalem specific literature
• Detailed violence and census data for 2001-09period available
2725.11.2019Karsten Donnay
Uninhabited/Industrial AreaPredominantly Secular/Moderate-Orthodox JewsPredominantly Ultra-Orthodox JewsPredominantly PalestiniansMunicipal Boundary; post-June 1967
Bhavnani, Donnay, Miodownik, Mor, Helbing. (2014). “Group Segregation and Urban Violence.” American Journal of Political Science 58(1): 226-245
||
Taking a closer look: Jerusalem
§ Key Ideas• empirical population data and geography: increase external validity
• violence data: optimize model to best match empirical situation
• validation:§ test predictive power in-sample§ compare predictions to simple statistical model§ consistency of model mechanisms with empirical reality
2825.11.2019Karsten Donnay
||
Taking a closer look: Jerusalem
§ Why this approach?• empirical grounding & formal validation:
§ external validity§ plausible model mechanisms
• building on literature: relevance for political science research
• goals:§ move beyond agent-based “toy models”
§ use framework for counterfactual studies (with policy implications)
292925.11.2019Karsten Donnay
||
Taking a closer look: Jerusalem
§ Evidence-driven computational modeling (EDM)
303025.11.2019Karsten Donnay
Bhavnani, Donnay, Miodownik, Mor, Helbing. (2014). “Group SegregaLon and Urban Violence.” American Journal of Poli0cal Science 58(1): 226-245
||
Taking a closer look: Jerusalem
§ Counterfactual analysis
313125.11.2019Karsten Donnay
Bhavnani, Donnay, Miodownik, Mor, Helbing. (2014). “Group Segregation and Urban Violence.” American Journal of Political Science 58(1): 226-245
Levels of Violenceno violencevery lowlowintermediatehighMunicipal Boundary; post-June 1967
Levels of Violenceno violencevery lowlowintermediatehigh
International Border; pre-June 1967Municipal Boundary; post-June 1967
||
Taking a closer look: Jerusalem
§ Implementation details:
• robust implementation as Java simulation platform
§ unit-testing for algorithmic consistency
§ counter measures against computingartifacts
§ validation against test scenarios
• full enumeration using supercomputing cluster
§ Euler cluster of ETH Zurich with >3000compute nodes
§ fine-grained enumeration of full model parameter space
§ sensitivity analysis for other key model parameters
323225.11.2019Karsten Donnay
||
Software tools
333325.11.2019Karsten Donnay
§ Open source and free:
• MESA for Python:https://github.com/projectmesa/mesa/
• Repast for Java:https://repast.github.io/
• NetLogo (not recommended for full-on modeling):https://ccl.northwestern.edu/netlogo/
||
Literature suggestions
• Bhavnani, Ravi, Karsten Donnay, Dan Miodownik, Maayan Mor, and Dirk Helbing. (2014). Group Segregation and Urban Violence. American Journal of Political Science 58(1): 226–245.
• de Marchi, Scott and Scott E. Page. (2014). Agent-Based Models. Annual Review of Political Science 17(1): 1–20.
• Epstein, Joshua M. (1999). Agent-Based Computational Models and Generative Social Science.Complexity 4(5): 41–60.
• Epstein, Joshua M. (2008). Why Model? Journal of Artificial Societies and Social Simulation 11(4): 12.
• Miller, John H. and Scott E. Page. (2004). The Standing Ovation Problem. Complexity 9(5): 8–16.
• Schelling, Thomas C. (1971). Dynamic Models of Segregation. Journal of Mathematical Sociology 1: 143–186.
• Weidmann, Nils and Idean Salehyan. (2013). Violence and Ethnic Segregation: A Computational Model Applied to Baghdad. International Studies Quarterly 57(1): 52–64.
• Axelrod, Robert. (1986). An Evolutionary Approach to Norms. American Political Science Review 80(4): 1095–1111.
• Axelrod, Robert. (1997). The Dissemination of Culture: A Model with Local Convergence and Global Polarization. Journal of Conflict Resolution 41(2): 203–226.
• Axtell, Robert L., Joshua M. Epstein, Jeffrey S. Dean, et al. (2002). Population Growth and Collapse in a Multiagent Model of the Kayenta Anasazi in Long House Valley. Proceedings of the National Academy of Sciences 99: 7275–7279.
343425.11.2019Karsten Donnay
||
Literature suggestions (continued)
• Bhavnani, Ravi and Dan Miodownik. (2008). Ethnic Polarization, Ethnic Salience, and Civil War. Journal of Conflict Resolution 53(1): 30–49.
• Bhavnani, Ravi, Michael G. Findley and James H. Kuklinski. (2009). Rumor Dynamics in Ethnic Violence.The Journal of Politics 71(3): 1–20.
• Cederman, Lars-Erik. (2003). Modeling the Size of Wars: From Billiard Balls to Sandpiles. American Journal of Political Science 97(1): 135–150.
• Centola, Damon, Robb Willer and Michael Macy. (2005). The Emperor's Dilemma: A Computational Model of Self-Enforcing Norms. American Journal of Sociology 110(4): 1009–1040.
• Epstein, Joshua M. (2002). Modeling Civil Violence: An Agent-Based Computational Approach.Proceedings of the National Academy of Sciences 99: 7243–7250.
• Helbing, Dirk. (2010). Pluralistic Modeling of Complex Systems. Science and Culture 76(9-10): 315– 329.
• Macy, Michael and Robert Willer. (2002). From Factors to Actors: Computational Sociology and Agent-Based Modelling. Annual Review of Sociology 28: 143–166.
• Schutte, Sebastian. (2010). Optimization and Falsification in Empirical Agent-Based Models. Journal of Artificial Societies and Social Simulation 13(1): 2.
353525.11.2019Karsten Donnay
||
§ If you have any questions, please get in touch!
Prof. Dr. Karsten DonnayAssistant Professor for Computational Social ScienceDepartment of Politics andPublic AdministrationUniversity of Konstanz
363625.11.2019Karsten Donnay