Klaus G. Troitzsch Universität Koblenz-Landau ESSA Summer School 2010

1. 50 Years of Social Simulation: Why We Need Agent-Based Social Simulation (and Why Other Approaches Fail), Klaus G. Troitzsch Universität Koblenz-Landau ESSA Summer School 2010 50 Years of Social Simulation

2. Outline • Simulation from the 1960s to 2010 • historical background • main features of some of the approaches • system dynamics, microsimulation, discrete event analysis, sociophysics, cellular automata • early extensions • some first conclusions • Why complex social systems are even more complex than other complex systems • peculiarities of human social systems • requirements for computational social science • and how they can be met: recent extensions 50 Years of Social Simulation

3. From World Models to Multi-Agent Models 50 Years of Social Simulation

4. Early Simulations • 1963 Simulmatics • Simulation as Science Fiction: Simulacron 3 (1964) • movies after this novel (“Welt am Draht” [“World on Wire”], Reiner Werner Fassbinder; “13th Floor”; “MATRIX”) 50 Years of Social Simulation

5. More Early Simulations • Microanalytical simulation of effects of tax and transfer regulations (since 1957) • Club of Rome simulations by Forrester and the Meadows group (early 1970s) • Thomas Schelling’s segregation model (1971) • Abelson’s and Bernstein’s referendum campaign simulation (1963) • Kirk’s and Coleman’s simulation of human behaviour in small groups (1968) • The Global 2000 Report to the President [Jimmy Carter], ed. Council on Environmental Quality and U.S. Department of State (1980) 50 Years of Social Simulation

6. Outline • Simulation from the 1960s to 2010 • historical background • main features of some of the approaches • system dynamics, microsimulation, discrete event analysis, sociophysics, cellular automata • early extensions • some first conclusions • Why complex social systems are even more complex than other complex systems • peculiarities of human social systems • requirements for computational social science • and how they can be met: recent extensions 50 Years of Social Simulation

7. System Dynamics • obviously has its roots in systems of differential equations from which it seems to differ mostly in two technical aspects: • discrete time is used as a coarse approximation for continuous time to achieve numerical solutions, and • functions of all kinds, including table functions, can be used with the help of the available tools like DYNAMO or STELLA. • is restricted to the macro level in that it models a part of reality (the ‘target system’) as an undifferentiated whole, whose properties are then described with a multitude of attributes which typically come as ‘level’ and ‘rate’ variables representing the state of the whole target system and its changes, respectively. 50 Years of Social Simulation

8. A PowerSim example 50 Years of Social Simulation

9. World Models Systems Dynamics and DYNAMO have received public interest mainly because they were used to build large world models: • WORLD2 (World Dynamics, Forrester 1970) • WORLD3 (The Dynamics of Growth in a Finite World, Meadows et al. 1974) • WORLD3 revisited (Beyond the Limits, Meadows et al. 1992) • WORLD3 (The 30-Year Update, Meadows et al. 2004) 50 Years of Social Simulation

10. Main Features of Forrester’s World Model (1) 50 Years of Social Simulation

11. Main Features of Forrester’s World Model (2) 50 Years of Social Simulation

12. WORLD2 complete All these feedback loops are, of course, tied together by auxiliaries and controlled by constants not shown in the previous diagrams. 50 Years of Social Simulation

13. WORLD2: Results Prediction results of Forrester’s WORLD2 model for births, deaths and population size 50 Years of Social Simulation

14. Retrodiction Retrodiction results of Forrester’s WORLD2 model for births, deaths and population size are obviously wrong. 50 Years of Social Simulation

15. Types of Validity • With Zeigler we should distinguish between three types of validity and three different stages of model validation (and development): • replicative validity: the model matches data already acquired from the real system (retrodiction), • predictive validity: the model matches data before data are acquired from the real system, • structural validity: the model “not only reproduces the observed real system behaviour, but truly reflects the way in which the real system operates to produce this behaviour.” • [Zeigler 1976:5] 50 Years of Social Simulation

16. Outline • Simulation from the 1960s to 2010 • historical background • main features of some of the approaches • system dynamics, microsimulation, discrete event analysis, sociophysics, cellular automata • early extensions • some first conclusions • Why complex social systems are even more complex than other complex systems • peculiarities of human social systems • requirements for computational social science • and how they can be met: recent extensions 50 Years of Social Simulation

17. The microsimulation approach • Microanalytic simulation models were first developed to predict demographic processes and their consequences for tax and transfer systems (Orcutt 1986). They consist of two levels at least: • the level of individuals or households (or in the rare case of simulating enterprises, the level of enterprises) • the aggregate level (e.g. national economy level) • More sophisticated MSMs distinguish between the individual and the household levels, thus facilitating models in which persons move between households and can found and dissolve new households (e.g. by marriage and divorce). 50 Years of Social Simulation

18. … what the founding fathers said … • “. . . in microanalytical modelling, operating characteristics can be used at their appropriate level of aggregation with needed aggregate values of variables being obtained by aggregating microentity variables generated by microentity operating characteristics” [Orcutt 1986, p. 14]. The main advantage of this kind of procedure is that • “available understanding about the behaviour of entities met in everyday experience can be used ... to generate univariate and multivariate distributions” [ibid.]. 50 Years of Social Simulation

19. Types of micro simulation • The classical micro simulation comes in three different types, the first of which is most common, but does not actually describe a (stochastic) process: • “static micro simulation”: change of the demographic structure of the model population is performed by reweighting the age class according to external information; • “dynamic micro simulation”: change of the demographic structure of the model population is performed by ageing the model persons individually (and by having them give birth to new persons, and by having them die) according to life tables; • “longitudinal micro simulation”: simulation is done on an age cohort and over the whole life of this cohort, thus omitting a population’s age structure (but children of the cohort members may still be simulated). 50 Years of Social Simulation

20. How it proceeds • All types of micro simulation, in contrast to many other simulation approaches, are data driven instead of concept driven: • Starting from data of a population or rather a sample from some population, normally on the nation state level, • this approach models the individual behaviour in terms of reproduction, education and employment, • simulates this individual behaviour and • aggregates it to the population level in order to generate predictions about the future age or employment structure. 50 Years of Social Simulation

21. How it proceeds current population with all properties of all individuals future population with all properties of all individuals real process sampling projection representative sample with selected properties predicted sample with selected attributes updated for all individuals simulation 50 Years of Social Simulation

22. Subprocesses • To realise the simulation, several subprocesses have to be modelled: • demographic processes: ageing, birth, death, marriage, divorce, regional mobility, household formation and dissolution • participation in education and employment, employment income • social transfers • taxes and social security • consumption • wealth 50 Years of Social Simulation

23. Subprocesses 50 Years of Social Simulation

24. Structure of a typical micro simulation model • Initialise the individuals from an empirical data base • Link them together according to their current household structure and to other information on networks (kinship or friendship networks, where the latter information will usually not be available) • Then, for every simulated period • organise the marriage market, • and for every simulated individual • increase its age, • decide whether it dies, • decide whether, if it represents a woman, it gives birth to one or more children, • decide whether, if it represents a person currently married, it is divorced, • decide whether and whom it will marry, • decide whether it will move from one household to another or form a new household, • decide on transitions in education and employment, respectively • and execute all these transitions and changes. • Store all the data needed for the analysis and interpretation of the simulated history and perhaps output some intermediate results. • Analyse and interpret the collected data, aggregate them, calculate distributions etc. 50 Years of Social Simulation

25. An alternative: event orientation instead of period orientation (1) • Usually microsimulation proceeds in a period-oriented manner. • Every agent has to check in every period whether anything happens with respect to it. • Alternatively, the simulation could proceed from event to event, and every event generats one or more new events: • At the time of birth, the events “child enters school” and “mother gives birth to another child” are scheduled for some time in the future (the waiting time being distributed according to some frequency distribution): • enter school • P(tschool = tbirth+5 = 0.2), • P(tschool = tbirth+6 = 0.5), • P(tschool = tbirth+7 = 0.3) • next birth • P(tnextbirth < tthisbrth+1 = 0.0), • P(tthisbrth+1 < tnextbirth < tthisbrth+25 = f(tnextbirth < tthisbrth)), • P(tnextbirth > tthisbrth+25 = 0.0) 50 Years of Social Simulation

26. An alternative: event orientation instead of period orientation (2) • Event-oriented agent-based microsimulation makes it necessary to look for other types of parameters than in period-oriented microsimulation: • instead of an age-dependent probability of giving birth to a child during the next period (year) we need an estimate of (e.g.) the frequency distribution of the time between the birth of the first and the second child, • instead of the age-dependent probability of marrying next year we need an estimate of the frequency distribution of the time between (e.g.) the time a person finishes school and the time when (s)he tries to find a partner: at the time of this event (s)he will look around for partners whose respective events are scheduled for the next very short period of time and select the best match from them, • instead of an age-dependent probability to die within the next period, we need the distribution of lifetimes; • some of these distributions are easily estimated, others are not. 50 Years of Social Simulation

27. Are microsimulation microentities agents? • Agents are • autonomous: they apply rules to beliefs and make decisions, perhaps also plans • reactive: they perceive stimuli from their environment and respond to them • proactive: they have goals which they try to achieve • socially capable: they have models of their environment and of other agents, and they can communicate with other agents [at least in Aparicio Diaz/Fent 2005] 50 Years of Social Simulation

28. UMDBS as one tool for micro simulation • micro data base • model • parameters / coefficients (life tables …) Universal Micro DataBase System UMDBS (Windows) [Sauerbier 2000, http://www.fh-friedberg.de/sauerbier/umdbs] 50 Years of Social Simulation

29. Output • tables • graphs • distributions (one- and two-dimensional) • queries 50 Years of Social Simulation

30. A pessimistic view • What such a micro analytical simulation model yields is in a way prediction, but not in the strict sense. • It is the outcome of one realisation of a stochastic process whose parameters are not exactly known but estimated on the base of more or less reliable empirical data. • The distribution of the outcome of this stochastic process can only be estimated (as it were, on a higher level of estimation) if a large number of parallel runs of the same model was run; then confidence intervals can be estimated on a Monte Carlo base. • After this time-consuming procedure we arrive at an estimate of the distribution of, e.g., the age distribution among women ten years from now, or of the distribution of the proportion of people over 65 with living daughters (to nurse them in case of sickness) — but only for the one set of parameters with which we initialised our simulation model earlier on, and not much is then known about the sensitivity, namely the dependence of the distribution of the outcomes of the stochastic process on slight changes on one or several of the large number of input parameters. 50 Years of Social Simulation

31. … and the optimistic view • Results of micro analytical simulation models have their value as they show possible paths into the future, • and Monte Carlo simulations of this type even show the reliability of the predictions, while multiple runs of similarly parameterised models give a first glance at the validity of the model: • if there is no sensitive dependence on initial conditions then the problem of estimating parameters is not a hard one. • And if we happen to have a long panel or a series of cross-sections then we can validate our model in comparing results of simulations of past periods with the empirical data of the same period. 50 Years of Social Simulation

32. Outline • Simulation from the 1960s to 2010 • historical background • main features of some of the approaches • system dynamics, microsimulation, discrete event analysis, sociophysics, cellular automata • early extensions • some first conclusions • Why complex social systems are even more complex than other complex systems • peculiarities of human social systems • requirements for computational social science • and how they can be met: recent extensions 50 Years of Social Simulation

33. Models from Econophysics and Sociophysics • Opinion formation or product choice • Simple case: two alternative opinions (“yes”/ “no”) or two alternative products (“MS-DOS” / “MacOS” or “VHS” / “Betamax”) • Probability of choice depends on global majorities • Typical approach: 50 Years of Social Simulation

34. Opinion formation in one population • NetLogo model 50 Years of Social Simulation

35. Opinion formation in several disjoint populations • NetLogo model 50 Years of Social Simulation

36. Outline • Simulation from the 1960s to 2010 • historical background • main features of some of the approaches • system dynamics, microsimulation, discrete event analysis, sociophysics, cellular automata • early extensions • some first conclusions • Why complex social systems are even more complex than other complex systems • peculiarities of human social systems • requirements for computational social science • and how they can be met: recent extensions 50 Years of Social Simulation

37. Cellular Automata • Defining features • Standard examples • Social science examples 50 Years of Social Simulation

38. A grid of cells 50 Years of Social Simulation

39. Defining features • A grid or lattice of a large number of identical cells in a regular array • e.g. a square • Each cell can be in one of a (small) set of states • e.g. ‘dead’ or ‘alive’ • Changes in a cell’s state are controlled by rules 50 Years of Social Simulation

40. Defining features (ii) • The cell’s rules depend only on the state of the cell and its local neighbours • e.g. the immediately surrounding cells • Consequently cells can only influence their immediate neighbours 50 Years of Social Simulation

41. Defining features (iii) • Simulated time proceeds in discrete steps • often called steps, cycles or generations • At each step, the state of every cell (at time t+1) is calculated using the states of neighbouring cells at time t. 50 Years of Social Simulation

42. Famous examples • The Game of Life • rules: • a ‘living’ cell remains alive if it has 2 or 3 living neighbours, otherwise it dies • a ‘dead’ cell stays dead unless it has exactly 3 living neighbours, when it bursts into life. 50 Years of Social Simulation

43. A Life sequence 50 Years of Social Simulation

44. The Game-Of-Life Glider 50 Years of Social Simulation

45. Neighbourhoods • von Neumann neighbourhood • Moore neighbourhood North East South West North North-east East South-east South South-west West North-west 50 Years of Social Simulation

46. The universe Right neighbour is left edge cell Bottom neighbour is top edge cell 50 Years of Social Simulation

47. Spreading gossip 50 Years of Social Simulation

48. Majority rule Starting configuration: 50% random ‘on’ Rule: ‘on’ if 5 or more Moore neighbours and self are ‘on’, ‘off’ if 5 or more Moore neighbours and self are ‘off’ Result: stable blocks of ‘on’ and ‘off’ form 50 Years of Social Simulation

49. The effect of individual differences At start Later Rule: majority rule with uniform random threshold variation (if 4 neighbours on and 4 off, new state is either on or off at random) 50 Years of Social Simulation

50. Outline • Simulation from the 1960s to 2010 • historical background • main features of some of the approaches • system dynamics, microsimulation, discrete event analysis, sociophysics, cellular automata • early extensions • some first conclusions • Why complex social systems are even more complex than other complex systems • peculiarities of human social systems • requirements for computational social science • and how they can be met: recent extensions 50 Years of Social Simulation