Der%20Agent%20

1
Advanced Computational Modelingof Social Systems
Lars-Erik Cederman and Luc Girardin Center for
Comparative and International Studies (CIS)
Swiss Federal Institute of Technology Zurich
(ETH) http//www.icr.ethz.ch/teaching/compmodels

2
Geosim

• Emergent Actors in World Politics (Princeton
  University Press, 1997)
• Inspired by Bremer and Mihalka (1977) and Cusack
  and Stoll (1990)
• Originally programmed in Pascal then ported to
  Swarm, and finally implemented in Repast

3
Model architecture
Actor
Actor
Relation
x,y res capital neighs
owner other twin act,res.. pol,prov
x,y res capital neighs
Relation
owner other twin act,res.. pol,prov
4
Main simulation loop
initiation phase
resource updating
resource allocation
decisions
inter- actions
structural change
5
Resource updating

• res resUnit
• for all provinces j of state i do
• res res resUnit

6
Resource allocation

• fixedRes(i,j) (1-propMobile) res / n
• mobileRes probMobile res
• for all relations j do
• in case i and j were fighting in the last
  period then
• mobileRes(i,j) res(j,i)/enemyRes(i)mobileRe
  s
• in case i and j were not fighting the last
  period then
• mobileRes(i,j)
• res(j,i)/(enemyRes(i)res(j,i))mobileRes
• res(i,j) fixedRes(i,j) mobileRes(i,j)

7
Decision rule of actor i

• for all external fronts j do
• if i or j fought in the previous period then
• attack j else cooperate with j Grim Trigger
• if there is no action on any front then
• select a neighboring state j
• with res(i,j)/res(j,i) gt superiorityThreshold
  do
• launch unprovoked attack against j
•  

8
Structural change conquest

• Conquest follows victorious battles
• Each attacker randomly selects a battle path
  consisting of an attacking province and a target
• The outcome depends on the targets nature
• if it is an atom, the whole target is absorbed
• if it is a capital, the target state collapses
• if it is a province, the target is absorbed

9
Guaranteeing territorial contiguity
Conquest... resulting in... partial state
collapse
"near abroad" cut off from capital
Target Province
Agent Province
j
i
10
Applying Geosim to world politics

• War-size distributions
• Democratic peace
• Nationalist insurgencies
• State-size distributions

11
Cumulative war-size plot, 1820-1997
Data Source Correlates of War Project (COW)
12
Self-organized criticality
Power-law distributed avalanches in a rice pile
Per Baks sand pile
13
Simulated cumulative war-size plot
log P(S gt s) (cumulative frequency)
log P(S gt s) 1.68 0.64 log s
N 218 R2 0.991
log s (severity)
See Modeling the Size of Wars American
Political Science Review Feb. 2003
14
Applying Geosim to world politics

• War-size distributions
• Democratic peace
• Nationalist insurgencies
• State-size distributions

15
Simulating global democratization
Source Cederman Gleditsch 2004
16
A simulated democratic outcome
t 0
t 10,000
17
Applying Geosim to world politics

• War-size distributions
• Democratic peace
• Nationalist insurgencies
• State-size distributions

18
4. Modeling civil wars

• Political economists argue that effectiveness of
  insurgency depends on projection of state power
  in rugged terrain rather than on ethnic cohesion
• But there is a big gap between macro-level
  results and postulated micro-level mechanisms
• Use computational modeling to articulate
  identity-based mechanisms of insurgency that also
  depend on state strength and rugged terrain

19
Main building blocks

• National identities
• Cultural map
• State system
• Territorial obstacles

20
The models telescoped phases
t 0
1000
1200
2200
Phase I Initialization
Phase II State formation Assimilation
Phase III Nation-building
Phase IV Civil war
identity- formation
nationalist collective action
assimilation
21
Sample run 3

• Geosim Insurgency Model

22
Applying Geosim to world politics

• War-size distributions
• Democratic peace
• Nationalist insurgencies
• State-size distributions

23
Puzzle

• Despite continuing progress, state sizes started
  declining in the late 19th century
• Lake and OMahony (2004) offer an explanation
  based on changes among democracies in the 19th
  and 20th centuries
• My argument nationalism caused the shift in
  state sizes

Technological progress
?
State size
24
Territorial state sizes
log Pr (S gt s)
log Pr (S gt s)
log S N(4.98, 1.02) MAE 0.048
log S N(5.31, 0.79) MAE 0.028
log s
1815
1998
log s
Data Lake et al.
25
Estimated means, 1815-1998
m
log s
Year
1800 1850 1900 1950 2000
26
Nested processes
27
A sample system at t 0
28
The sample system at t 2000
29
t 2054
30
t 2060
31
t 2813
32
Estimated m-values in 30 simulations
33
Simulated state sizes fitted by log-normal curve
log Pr(Sgts)
log Pr(Sgts)
log S N(1.28, 0.09) MAE 0.040
log S N(1.41, 0.10) MAE 0.046
log s
log s
t 2000
t 5000
34
Strategy Planned vs Reactive

• Goal-oriented planning
• Scan the possible options and find the sequence
  of actions that matches the goal
• Humans do little planning!
• Requires global knowledge!
• Reactive behaviors
• Use properties of the current situation, and use
  the output directly as a decision
• Need to express the problem differently, so that
  the goal can be reached incrementally

35
Goal-oriented planning
36
Reactive behavior
37
Rule-based systems

• Collection of if then statements that are
  used to manipulate variables
• If there is a limited number of situations, then
  this can be modeled as a finite-state machine
• Forward- and backward-chaining inference
• Match facts with rules to derive new facts
• Identify all the rules that could have led to the
  given fact

38
Subsumption architecture

• Set of horizontal layers
• The higher the layer, the greater the priority

Priority Behavior Condition
6 Retreat Low chance of winning
5 Evade Incoming threat
4 Attack Enemy present
3 Gather Low health
2 Investigate Possible enemy
1 Explore Always
39
Learning

• Optimization
• Attempt to find the solution to a known puzzle,
  which does not change other time
• Adaptation
• When the problem or the goal change, then
  adaptation is necessary
• Exploration vs Exploitation
• Attempt to cover all possible states by trying
  every action or confine to the set of actions
  known to be valuable

40
Exploration vs exploitation in chess
41
Various approaches

• Expert solution
• Rule-based or expert system, finite-state
  automata
• Expert guidance
• Neural network, optimization techniques,
  evolutionary algorithms, decision trees,
  reinforcement learning
• Imitation
• Exhaustive
• Brute force approach, dynamic programming
• Random
• Stochastic search no bias!
• Hybrid
• Learning classifier systems

42
GeoContest runs

• 60 x 60 grid
• Initial polarity of 1000
• 14 strategies
• 100000 iterationsper run
• More than 500 runs needed to accuratelydistingui
  sh the winner

t40000
43
GeoConstest results
44
GeoConstest results (cont.)
45
GeoConstest results (cont.)

• Gold PinkPanther (25 of games)
• Silver GeliStrategy (16 of games)
• Bronze Indiana Jones (15 of games)
• Heavily path-dependent no clear trend before
  performing many runs
• The set of present strategies affect the winning
  strategy (exploitation of weak strategy)