Job Talk - James H. Fowler

Download Report

Transcript Job Talk - James H. Fowler

Agent Based Models
in Social Science
James Fowler
University of California, San Diego
The Big Picture: Collective Action

Cooperation

Alternative Models of Participation

Social Networks
Cooperation

Evolutionary models





Altruistic Punishment and the Origin of Cooperation
PNAS 2005
Second Order Defection Problem Solved?
Nature 2005
On the Origin of Prospect Theory
JOP, forthcoming
The Evolution of Overconfidence
Experiments




Egalitarian Motive and Altruistic Punishment
Nature 2005
Egalitarian Punishment in Humans
Nature 2007
The Role of Egalitarian Motives in Altruistic Punishment
The Neural Basis of Egalitarian Behavior
Alternative Models of Political Participation

Computational Models of Adaptive Voters and
Legislators





Parties, Mandates, and Voters: How Elections Shape the Future 2007
Policy-Motivated Parties in Dynamic Political Competition
JTP 2007
Habitual Voting and Behavioral Turnout
JOP 2006
A Tournament of Party Decision Rules
Empirical Models of Legislator Behavior



Dynamic Responsiveness in the U.S. Senate
AJPS 2005
Elections and Markets: The Effect of Partisan Orientation, Policy Risk, and
Mandates on the Economy
JOP 2006
Parties and Agenda-Setting in the Senate, 1973-1998
Alternative Models of Political Participation

Experiments







Altruism and Turnout
JOP 2006
Patience as a Political Virtue: Delayed Gratification and Turnout
Political Behavior 2006
Beyond the Self: Social Identity, Altruism, and Political Participation
JOP 2007
Social Preferences and Political Participation
When It's Not All About Me: Altruism, Participation, and Political Context
Partisans and Punishment in Public Goods Games
Genetics


The Genetic Basis of Political Participation
Southern California Twin Register at the University of Southern California: II
Twin Research and Human Genetics 2006
Political Social Networks

Voters



Dynamic Parties and Social Turnout: an Agent-Based Model
AJS 2005
Turnout in a Small World
Social Logic of Politics 2005
Legislators






Legislative Cosponsorship Networks in the U.S. House and Senate
Social Networks 2006
Connecting the Congress: A Study of Cosponsorship Networks
Political Analysis 2006
Community Structure in Congressional Networks
Legislative Success in a Small World: Social Network Analysis and the
Dynamics of Congressional Legislation
Co-Sponsorship Networks of Minority-Supported Legislation in the House
The Social Basis of Legislative Organization
Political Social Networks

Court Precedents


The Authority of Supreme Court Precedent
Social Networks, forthcoming
Network Analysis and the Law: Measuring the Legal Importance of
Supreme Court Precedents
Political Analysis, forthcoming
Other Social Networks

Political Science PhDs


Academic Citations


Social Networks in Political Science: Hiring and Placement of PhDs, 19602002
PS 2007
Does Self Citation Pay?
Scientometrics 2007
Health Study Participants



The Spread of Obesity in a Large Social Network Over 32 Years
New England Journal of Medicine 2007
Friends and Participation
Genetic Basis of Social Networks
What is an Agent Based Model?


Computer simulation of the global
consequences of local interactions of members
of a population
Types of agents




plants and animals in ecosystems (Boids)
vehicles in traffic
people in crowds
Political actors
What is an Agent Based Model?


“Boids” are simulations of bird flocking behavior
(Reynolds 1987)
Three rules of individual behavior

Separation


Alignment


point towards the average heading of other birds
Cohesion


avoid crowding other birds
move toward the center of the flock
Result is a very realistic portrayal of group motion in
flocks of birds, schools of fish, etc.
What is an Agent Based Model?

Comparison with formal models



Same mathematical abstraction of a given problem,
but uses simulation rather than mathematics to
“solve” model and derive comparative statics
Comparison with statistical models


Same attempt to analyze data,
but uses simulation data rather than real data
Advantages of Agent Based Modeling

Formal


Flexible


Easier to cope with complexity
(nonlinearities, discontinuities, heterogeneity)
Generative


Cognitively: agents can be “rational” or “adaptive”
Tractable


Assumptions laid bare
Helps create new hypotheses
Social Science from the Bottom Up

“If you didn’t grow it you didn’t show it.”
Disadvantages of Agent Based Modeling

Models too simple



Models too complicated



Not possible to assess causality (Cederman 1997)
What use is an existence proof?
Coding mistakes


Could be solved in closed-form (Axelrod 1984)
Closed-form solution always preferable
Many more lines of code than lines in typical formal proof
Data analysis

What part of the parameter space to search?
My Approach to Agent Based Modeling







Write down model
Solve as much as possible in closed-form
Justify simulation with mathematical description of the
complexity problem
Use real world to “tune” model
Make predictions
Check predictions against reality
Do comparative statics near real world parameters to
assess causality
Tournament Overview

A dynamic spatial account of multi-party multi-dimensional
political competition




Existing ABMs use a fixed set of predefined strategies,
typically in which all agents deploy the same rule.


is substantively plausible
generates a complex system that is analytically intractable
amenable to systematic and rigorous computational investigation using
agent based models (ABMs)
There as been little investigation of potential rules, or the performance of
different rules in competition with each other
The Axelrodian computer tournament is a good
methodology for doing this …

… while also offering great theoretical potential to be expanded into a more
comprehensive evolutionary system
Tournament ABM test-bed

We advertised a computer simulation tournament
with a $1000 prize for the action selection rule
winning most votes, in competition with all other
submitted rules over the very long run.

Tournament test-bed (in R) adapted from Laver
(APSR 2005)


The four rules investigated by Laver were declared pre-entered
but ineligible to win: Sticker, Aggregator, Hunter and Predator
Submitted rules constrained to use only published
information about party positions and support levels
during each past period and knowledge of own
supporters’ mean/median location
Departures from Laver (2005)





Distinction between inter-election (19/20) and election
(1/20) periods
Forced births (1/election) at random locations, as
opposed to endogenous births at fertile locations, à la
Laver and Schilperoord
De facto survival threshold (<10%, 2 consecutive
elections)
Rule designers’ knowledge of pre-entered rules
Diverse and indeterminate rule set to be competed
against
Tournament structure

There were 25 valid submissions – after several R&Rs for
rule violations, elimination of a pair of identical submissions
and of one in R code that would not run and we could not fix
– making 29 distinctive rules in all.






Five runs/rule (in which the rule in question was the first-born)
200,000 periods (10,000 elections)/run (after 20,000 period burn in)
Thus 145 runs, 29,000,000 periods and 1,450,000 elections in all
Brooks-Gelman tests used to infer convergence, in the sense that results
from all chains are statistically indistinguishable.
There was a completely unambiguous winner – not one of
the pre-entered rules
However only 9/25 submissions beat pre-announced
Sticker (i.e. select random location and never move)
Tournament algorithm portfolio

Center-seeking rules: use the vote-weighted centroid or median


Previous work suggests these are unlikely to succeed, a problem exacerbated in
a rule set with other species of the same rule
Tweaks of pre-entered rules: eg with “stay-alive” or “secret handshake”
mechanisms (see below)



Sticker is the baseline “static” rule for any dynamic rule to beat
Hunter was the previously most successful pre-entered rule
“Parasites” (move near successful agent): have a complex effect




Split successful “host” payoff so unlikely to win – especially in competition
with other species of parasite
But do systematically punish successful rules
No submitted rule had any defense against parasites
No submitted parasite anticipated other species of parasite
Tournament algorithm portfolio

Satisficing (stay-alive) rules: stay above the survival
threshold rather than maximize short-term support


“Secret handshake” rules: agent signals its presence to
other agents using the same rule (e.g. using a very
distinctive step size), who recognize it and avoid attacking
it


Substantively plausible but raise an important issue about agent time
preference – which only becomes evident in a dynamic setting
Substantively implausible (?) but, given 29 rules and random rule
selection, there was smallish a priori probability that an agent would
be in competition with another using the same rule
Inter-electoral explorers: use the 19 inter-election periods
to search (costlessly) for a good location on election day

Substantively plausible but raise an important issue about relative
costs of inter-electoral moves
Results: votes/rule
Results: votes/agent-using-rule
Results: agent longevity
Results: Pairwise performance
Vote Share By Type
KQSTRAT
SHUFFLE
GENETY
FI SHER
PRAGMATI ST
STICKY-HUNTER
PICK-AND-STICK
RAPTOR
HUNTER
HALF-AGGREGATOR
STICKER
NICHE-HUNTER
PATCHWORK
NICHE-PREDATOR
AGGREGATOR
CENTER-MASS
AVERAGE
FOOL-PROOF
PREDATOR
FOLLOW-THE-LEADER
INSATI ABLE-PREDATOR
MEDIAN-VOTER-SEEKER
PARASI TE
AVOIDER
ZENO
OVER-UNDER
MOVE-NEAR-SUCCESSFUL
JUMPER
BIGTENT
Number of Parties By Type
KQSTRAT
SHUFFLE
GENETY
FISHER
PRAGMATIST
STICKY-HUNT ER
PICK-AND-STICK
RAPTOR
HUNT ER
HALF-AGGREGATOR
STICKER
NICHE-HUNT ER
PATCHWORK
NICHE-PREDATOR
AGGREGATOR
CENTER-MASS
AVERAGE
FOOL-PROOF
PREDATOR
FOLLOW-T HE-LEADER
INSAT IABLE-PREDATOR
MEDIAN-VOTER-SEEKER
PARASITE
AVOIDER
ZENO
OVER-UNDER
MOVE-NEAR-SUCCESSFUL
JUMPER
BIGTENT
Results: run-off
Rule
Ru n off
Mean Median
vote
rank
Tourn ame n t
Mean Median
vote
rank
KQ-Strat
Pick-and-Stick
Sticky hunter
Genety
Pragm atist
Shuffle
Fisher
19.6
15.4
15.0
14.0
13.6
11.6
10.9
11.2
6.8
7.3
8.4
7.4
9.7
7.9
1
2
3
4
5
6
7
1
6
5
4
5
2
4
Results: No Secret Handshake
KQSTRA T
SHUFFLE
GENETY
FISHER
PRA GMA TIST
STICKY .HUNTER.MEDIA N.FINDER
PICK.A ND.STICK
RA PTOR
HA LF.A GGREGA TOR
HUNTER
STICKER
NICHE.HUNTER
PA TCHW ORK
NICHE.PREDA TOR
A GGREGA TOR
CENTER.MA SS
FOOL.PROOF
A V ERA GE
PREDA TOR
INSA TIA BLE.PREDA TOR
FOLLOW.THE.LEA DER
PA RA SITE
MEDIA N.V OTER.SEEKER
A V OIDER
ZENO
OV ER.UNDER
MOV E.NEA R.SUCCESSFUL
JUMPER
BIGTENT
0
200
400
600
800
Total Votes Received Per Simulation (Thousands)
1000
1200
Results: Evolutionary Reproduction
Characteristics of successful rules




KQ-strat focused on staying alive, protected itself against cannibalism with a
very distinctive step size, and became a parasite when below the survival
threshold
Shuffle was a pure staying-alive algorithm, non-parasitic and without
explicit cannibalism protection, though unlikely to attack itself since it tends
to avoid other agents
Genety had used prior simulations deploying the genetic algorithm to
optimize its parameters against a set of pre-submitted and anticipated rules.
It was not a parasite, had no protection against cannibalism and did not focus
on staying alive.
Fisher distinctively used the 19 inter-electoral periods to find the best
position at election time. However, it also satisficed by taking much smaller
steps when over the threshold
Characteristics of successful rules

Of the three other rules doing significantly better
than Hunter:





Sticky-Hunter/Median-Finder conditioned heavily on the
survival threshold
Pragmatist simply tweaked Aggregator by dragging it
somewhat towards the vote-weighted centroid
Pick-and-Stick simply tweaked Sticker by picking the best of
19 random locations explored in the first 19 post-birth interelection periods.
Pure center-seeking and parasite rules did badly
Set of successful rules was thus diverse – most
systematic pattern being to condition on the
survival threshold
14
12
KQSTRAT
FISHER
10
HUNTER GENETY
HALF-AGGREGATOR
STICKER
NICHE-HUNTER
AGGREGATOR
PREDATOR
AVERAGE
MEDIAN-VOTER-SEEKER
PARASITE
AVOIDER
8
ZENO
MOVE-NEAR-SUCCESSFUL
6
Average Vote Share (%)
Medium Eccentricity is Best
JUMPER
0.0
0.5
1.0
1.5
Eccentricity
(Distance From Center)
2.0
14
12
10
KQSTRAT
SHUFFLE
PICK-AND-STICK
GENETY
HUNTER
RAPTOR
STICKER
PATCHWORK
AGGREGATOR
PREDATOR
MEDIAN-VOTER-SEEKER
PARASITE
AVOIDER
8
ZENO
MOVE-NEAR-SUCCESSFUL
6
Average Vote Share (%)
Less Motion is Better
JUMPER
0.0
0.5
1.0
1.5
Motion
(Average Distance Moved From Previous Election)
Conclusion



Agent Based Models can help us assess causality in
social science
Tournaments can help bring human element into an
ABM
However, agent-based modelers must




Keep models simple
Check for closed-form solutions
Ground models in the real world
Work closely with statisticians (EI) and formal modelers
(TM)