The Re Past Framework and Social Simulations Presented

The Re. Past Framework and Social Simulations Presented by Tim Furlong

Overview Re. Past Social Simulations implemented with Re. Past Santa Fe Artificial Stock Market Endogenizing Geopolitical Boundaries

Re. Past REcursive Porous Agent Simulation Toolkit Java class library University of Chicago Social Science Research Computing

Re. Past: Framework Base classes to be extended Engine class Agent class Environment class GUI displays, charts, graphs Utility classes Spatial representations Statistical RNGs

Generic approach Discrete event simulator Easy implementation Sugar. Scape(partial) : ~ 650 LOC Game of Life : ~ 750 LOC

Re. Past: Advantages Facilitates implementation Convenient representation of heterogeneous agents Support for geometric world models Garbage collection ‘Powerful’ visualization techniques Lars-Erik Cederman, “Endogenizing Geopolitical Boundaries with Agent-based Modeling”, prepared for Sackler Colloquium on “Adaptive Agents, Intelligence, and Emergent Human Organization: Capturing Complexity through Agent-based Modelling”, Oct. 2001.

Re. Past: Applications School voucher programs Consumer choice Decision making in closed regimes Modeling the size of wars Voting dynamics Self-organizing computer networks Multi-cellular tumors Repast Homepage – Projects and Publications : http: //repast. sourceforge. net/projects. html

Social Simulations Goal is to simulate observed behaviors with hypothesized model Several ‘flavors’ of simulation Statistical : global variables Agent-based : allows heterogeneous agents with varied and dynamic behavior

The Santa Fe Artificial Stock Market Re-Examined: Suggested Corrections Norman Ehrentreich

SFI-ASM: Introduction Simplistic stock market simulation Isolates learning speed of traders as critical parameter Based on original SFI-ASM Fixes faulty mutation operator Results not quite as compelling Interesting Re. Past model

SFI-ASM: Original Model N traders 1 unit risky stock, 20 000 units cash Each trader seeks to buy or sell stock based on expectations of profit Profit Fixed return of rf on cash assets Stock pays stochastic dividend

SFI-ASM: Stock Only one ‘stock’ in market Stock has price pt and dividend dt Dividend of stock at time t +1 Mean-reverting factor of (1 – ρ), but generally stochastic

SFI-ASM: Traders Risk aversion factor of λi Wealth at time t of Wi, t: stock + cash Optimal amount of stock based on expectations of profit

SFI-ASM: Expectation rules Market has descriptor Dt Bitstring of market conditions Each trader has own set of 100 rules Rule comprised of: Condition Forecast accuracy Fitness value

Condition is pattern matching rule String of {0, 1, #} Bits are technical or fundamental Forecast for rule j: (aj, bj)

Forecast Accuracy Fitness Value

SFI-ASM: Rule Evolution Genetic algorithm invoked after every K rounds of trading to evolve rules Mutation (p=0. 7) Crossover

SFI-ASM: Correction Original had faulty mutation operator Biased results to higher number of non-# bits Correct solution for rules is to converge to all-# bits Dividend and price too random to classify With new operator, rules always converge

SFI-ASM: Results Rules converge to all-# bits Reach homogeneous rational expectation equilibrium eventually With values for K < 100, complex trading emerges Harder to persuade the model to do this with the new mutation operator

Faster learners exploit slower learners Short-term trends In new model, only valid in beginning

Endogenizing Geopolitical Boundaries with Agent-based Modeling Lars-Erik Cederman

EGB: Introduction Agent-based modeling has potential to avoid reification of actors Reification: treating an abstract concept as concrete Long-term simulations require “sociational endogenization” of actors Actors must be internally dynamic

EGB: Background Essentialist perspective Ignore change of actors Fixed entities with attributes Sociational perspective Dynamic actors and relationships Context-sensitive

EGB: Endogenization Presents series of models to illustrate progression from reified actors to endogenous ones Modeling emergence of state borders Emergent Polarization (EP) Democratic Peace (DP) Nationalist Systems Change (NSC)

EGB: Emergent Polarization Models conquest and expansion of states Villages or counties on a finite 2 d grid States emerge as villages conquer neighbors State has capital based on original village Resources gathered from the territories depends on distance to capital

EGB: EP turn structure Five phases per turn Resource allocation Decisions Interaction Resource updating Structural change

Resource allocation Allocate troops to borders based on strength of neighbors Decisions Reciprocate aggressive action Attempt unprovoked attacks

Interaction Resolve conflicts based on balance of power Resource updating States gain resources from provinces Structural change Structure of defeated state altered by outcome of conflicts

Notes States can spread too thin, inviting attack from other neighbors and opening multiple fronts to conflict Can extend the model to allow alliances between states

EGB: Democratic Peace Adds categorical relationships to previous model Observed that democracies do not fight each other Add ‘democracy’ label to some states Democracies do not fight each other, and form a defensive coalition

Notes Difference in balance of power produces significant results Example of adding ‘categorical social’ processes • Threat evaluation is still relational

EGB: Nationalist Systems Change Introduce concept of actors separate from states : nations Nations and states sometimes coincide, but not always Each village has ‘cultural’ identity : string of trait values Nation is a pattern string of traits with wildcards

Nations founded and joined by agents Capitals more likely to found nations due to resources National identities have major impact on inter-state relations ‘irredentist’ invasions to conquer conationals not under ‘home rule’

EGB: Conclusions Agent-based simulations are better at modeling complex phenomenae than conventional approaches Treating actors as themselves emergent and internally dynamic is necessary to good simulation over long time scales

Questions?