Upload
edmund-chattoe-brown
View
120
Download
8
Embed Size (px)
Citation preview
What is simulation and what use is it?
Edmund ChattoeDepartment of Sociology
University of OxfordManor Road, Oxford, OX1 3UQ
[email protected]://www.sociology.ox.ac.uk/people/chattoe.html
Plan of the talk• Clearing the ground• Types of theorising• A simple example with methodological
implications• A more realistic example and contribution
to “live” sociological debates• Styles of simulation: A brief comparison• Conclusions
Simulation: A confusing term• Gaming or role playing: “Simulated” United Nations
for schools• Instrumental and descriptive simulation: Dealing with
messy calculus (Buffon)• A realist/empiricist approach to social theory:
Nothing to do with Baudrillard, PoMo and simulacra• A third type of representation for social processes:
neither a “mathematical” model, nor a narrative but a computer programme
• Simulation types: agent based, system dynamics
Mathematical theory: Lotka-Volterra• Let us assume that the prey in our model are rabbits and
that the predators are foxes. If we let R(t) and F(t) represent the number of rabbits and foxes, respectively, that are alive at time t, then the Lotka-Volterra model is:
• dR/dt = a*R - b*R*F• dF/dt = e*b*R*F - c*F• a is the natural growth rate of rabbits absent predation• c is the natural death rate of foxes absent food (rabbits)• b is the death rate per encounter of rabbits due to
predation• e is the efficiency of turning predated rabbits into foxes
Narrative theory: Marx• But with the development of industry, the proletariat not only increases in
number; it becomes concentrated in greater masses, its strength grows, and it feels that strength more. The various interests and conditions of life within the ranks of the proletariat are more and more equalised, in proportion as machinery obliterates all distinctions of labour, and nearly everywhere reduces wages to the same low level. The growing competition among the bourgeois, and the resulting commercial crises, make the wages of the workers ever more fluctuating. The increasing improvement of machinery, ever more rapidly developing, makes their livelihood more and more precarious; the collisions between individual workmen and individual bourgeois take more and more the character of collisions between two classes. Thereupon, the workers begin to form combinations (trade unions) against the bourgeois; they club together in order to keep up the rate of wages; they found permanent associations in order to make provision beforehand for these occasional revolts. Here and there, the contest breaks out into riots. (Communist Manifesto)
Simulated theory: Schelling example • Three state regular grid (red agent, green agent or
vacant site)• Red and green agents have two psychological
states (“satisfied” and “dis-satisfied”) based on an innate and fixed “preference” for sharing the type of their immediate neighbours
• If agent is satisfied, it stays still. If dis-satisfied, it moves to a randomly selected vacant site
• Randomly ordered updating for whole agent population determines each simulated “period”
Sample Initialisation
50% similarity
1500 agents and
1000 vacant sites
Two questions• How xenophobic do agents have to be to
produce segregation? (A percentage for the same neighbour requirement at or above which recognisable clustering results.)
• How does the type of clustering change for total xenophobia? (100% same neighbour requirement.)
• DON’T SPOIL IT IF YOU ALREADY KNOW THE ANSWERS!
Type A "Error": Non Xenophobic Clusters
80% similarity
Type B "Error": Xenophobic Non Clusters
50.4% similar: stopped after 50
periods
Simulated theory: Schelling again• to find-new-spot• rt random 360• fd random 10• if any other-turtles-here• [ find-new-spot ] ;; keep going until we find an unoccupied patch• end
• to update-patches• ask patches [• ;; in next two lines, we use "neighbors" to test the eight patches surrounding the current patch• set reds-nearby count neighbors with [any turtles-here with [color = red]]• set greens-nearby count neighbors with [any turtles-here with [color = green]]• set total-nearby reds-nearby + greens-nearby ]• end
• to update-turtles• ask turtles [• if color = red• [ set happy? reds-nearby >= ( %-similar-wanted * total-nearby / 100 ) ]• if color = green• [ set happy? greens-nearby >= ( %-similar-wanted * total-nearby / 100 ) ] ]• end
First two uses of simulation• Simulation as “complexoscope”: Just as a microscope
allows us to see things too small for the naked eye, a simulation allow us to understand things too complex for the “bare” brain. As the Schelling model shows, even quite simple systems are complex.
• Simulation as theory building tool: Even the simple Schelling model captures and solidifies the potentially abstruse notion of structuration. (A simulation is worth a thousand words?) In choosing, agents determine the “environment” which then influences their choice: “white flight”, tipping points.
Simulation and data: A distinctive relationship • What would we need to do to make the
Schelling model more realistic?• First: How do people classify neighbours
and make consequent relocation decisions?• Second: How “similar” are the clusters
produced by the simulation model and those observed in real urban settings?
• A combination of “traditional” qualitative and quantitative data (plus novel methods?)
Two types: No love lost
60% same type
preference
Three types: No love lost
60% same type
preference
Three types: The colonel’s lady ... reds and
greens both consider
each other as acceptable “company”
ISSUE OF “EQUIVALENCE
CLASSES”
The GT Box
QUALITATIVE
QUANTITATIVE
FALSIFICATION
RESEARCHDESIGN
Revisiting types of theory• Statistical models (found in quantitative research) make the
comparison between model and real system at the “aggregate” level but seldom specify an explicit micro mechanism generating the observed pattern. To my knowledge, no such mechanism has been independently tested even where proposed.
• Narrative theories (found in ethnography and “pure” social theory) describe individual states and interactions but ethnography seldom even attempts to generalise nowadays and simulations of social theories often don’t generate the outcomes hypothesised (Friedman example) because of complexity. Formalising theories is another interesting (if minority) use for simulation.
Case study: The strength of strict churches• Begins with Kelley and a potentially
counter-intuitive claim: The way to maintain a church is to ask more from adherents not less
• Statistical debate about whether this is true.• Problems with causality, contributions that
are hard to measure (differential association) and explanation
• Iannacconne RCT model of strict churches
The Iannaccone explanation• Worshippers face a time/money allocation
problem between secular and religious activities• Religion is a club good• This creates a free-rider problem• One solution is prohibiting secular activities• This often creates an enforcement problem• A solution is to effectively raise costs of
prohibited activities using apparently “irrational” practices (dietary restrictions, dress codes)
Unpacking this argument• Although intended as a RCT account of
worshippers, this is also an interesting functionalist account of church dynamics
• Churches that demand, prohibit and enforce simultaneously will thrive, others will not (based on income/membership constraints)
• Iannaccone proves an equilibrium result assuming unbounded rationality and perfect information
Building a simulation• Objection 1: Agents cannot choose over whole
space of allocations so have them compare only pairs of allocations at any instant.
• Objection 2: There is social structure not global knowledge. Comparators come (differentially) from self (choice), deliberate recruitment to new churches, own church members (social imitation) or other church members (social learning)
• Objection 3: The population of churches is dynamic with new creeds being born and churches with no members or income “dying”.
Interesting implications• Under “more realistic” assumptions the Iannaccone
result breaks down.• What kind of data do we need to build better
simulations? Meta-analysis of existing ethnography, theory driven comparative studies of successful and unsuccessful churches, different styles of “quantitative” data (contact diaries?)
• Can we use falsification based on more than one “dimension” of data: longitudinal church membership as well as cross sectional?
• Is functionalism coherent after all?
Types of simulation• Instrumental: Numerical integration,
probability distributions for hard functions• Microsimulation, system dynamics: Based
on assumptions of underlying stability in “transition probabilities”
• Agent based simulation: Grounds out all behaviour at the individual level. The only “parameters” are those used by agents themselves in their mental models.
Example: Trends in drug use (Caulkins)
LIGHTUSERS
HEAVYUSERS
NONUSERS
a g
b
L(t+1)=(1-a-b)L(t)+I(t), H(t+1)=(1-g)H(t)+bL(t)
Initiation
Issues• Presumed constancy of a, b, g, category
boundaries of use.• Non explanation of I(t).• If model design criterion is curve fitting, is
this explanation or data mining? (Should there be a distinctive box for “never used” and if so, where do we stop with building boxes based on something other than best fit?)
DrugChat Model• DTI Foresight Replication of DrugTalk
(Agar)• Explicit (but simple) representation of agent
social networks• Different types distinguished by behaviour
(“partying”) and evaluation (credibility of reported drug attitudes) rather than use level
• No parameter constancy assumptions just attributes and states at agent level
Simulation and data• Having a new method calls attention to the
need for new data/theory (Maslow): dynamic decision mechanisms, unstructured choices, large scale network structure, generic network properties
• It also shapes the collection and use of data (comparative statistics and qualitative similarity measures rather than model fit, “theory building” ethnography, systematic analysis of published research)
Institutional issues• Protective expectations: “Substantiveness”, “rigour”
and other exclusionary codes.• Size of simulator population: Quality as a function
of investment.• Length of existence of field: How to tell
exuberance from sloppiness?• Lack of infrastructure: professional training,
journals, conferences and so on• Raising standards internally (replication, data
protocols, systematic literature review)• The high frontier?
Conclusions• Simulation as complexoscope: No social reality
needed• Simulation as theory exploration tool: Need only a
stated theory and wise intuitions (Iannaccone)• Simulation as both “generative” and “falsifiable”
social science: Need real micro and macro data (Schelling)
• Simulation as a tool for “detheorising” theory (removing parameters and implicit assumptions) and developing interdisciplinary programmes of progressive research: Watch this space, I hope!
Now read on• Journal of Artificial Societies and Social Simulation
(JASSS): http://jasss.soc.surrey.ac.uk/JASSS.html• NetLogo: http://ccl.northwestern.edu/netlogo/• Gilbert, Nigel and Troitzsch, Klaus G. (2005) Simulation
for the Social Scientist (Open University Press).• BJS: ‘Using Simulation to Develop and Test Functionalist
Explanations: A Case Study of Dynamic Church Membership’, http://users.ox.ac.uk/~econec/bjs-1.doc
• DTI: <http://www.foresight.gov.uk/Brain_Science_Addiction_and_Drugs/Reports_and_Publications/DrugsFutures2025/Index.htm>