Searching for “Phases” in Complex Simulation Output using Evolutionary Knowledge Discovery Techniques
Searching for “Phases” in Evidence-Led Common-Sense Secondary Analysis of Complex Simulation Specification Entities Simulation Output Output using Evolutionary ￼ Specification simulation rules and agent Comparison with evidence is facilitated if the The Model is run many times (in this case Knowledge Discovery behaviour is informed by available entities in the simulation correspond to entities 7000 times) each time using randomly Techniques evidence (as much as possible). Some that are observed in a naturalistic manner. selected parameter values (p1, p2,…). Many examples are listed below. A Small District different measures are recorded concerning Bruce Edmonds and other SCID the outcomes from different layers of the Project Members Initial party preference inherited model (m1, m2, … including the predicted Party preference can be linked to learning from variable, t). This provides a rich data set for parents (e.g. Verba, Scholzman et al. 2005) . the secondary analysis.Given that models that are adequate tomuch social phenomena will necessarily be People vote out of habithighly complex, we are left with the Going to the polls in one election will lead to a greater likelihood of returning to the polls in a subsequentnecessity of understanding them. The election (e.g. Gerber, Green et al. 2003) .approach here is to construct relevant butcomplex simulation models to start with Voting is a social norm(Data Integration Models) and then try and Civic duty is an important rationale for individual-level turnout (e.g. Riker and Ordeshook, 1968).model this with simpler models. People share the political views of their greater This data is then distributed over a space networks according to the values of a couple of the Probability of agreement within a network depends on the distribution of political opinion within one’s parameters (the grey background patchwork network (autoregressive networks) (e.g. Huckfeldt, indicates the density of this data in the space) Johnson, and Sprague, 2004). Emmigration Rate Electors can be mobilised to vote by family, friends and political parties Household members, friends and political parties will ask people to vote on election day (e.g. Cutts and Fieldhouse, 2009). A Household There are high amounts of homophily in social networksHere, to aid in this search, we use Individuals have more contact with similar people (e.g.evolutionary techniques to look for McPherson, Smith-Lovin et al. 2001).hypotheses about the model behaviour, butnot over the whole parameter “space” but Education increases the level of political interest The level of exposure to (political) information one israther to identify clusters where local exposed to increases when pursuing higher educationpatterns hold – maybe akin to “phases” (e.g. Lewis-Beck, 2008). Class Activitiesfound in some physical systems. These Agemight suggest context-dependent rules for Political experts are more influential within Ethnicity Etc. political discussion networksa simpler model or summaries of the Level-of-Political-Interest People will tend to listen to people they believe arecomplex model behaviour to use in the political experts (those who have higher levels of Memory(relative) validation of simpler models. political interest/involvement) (e.g. Huckfeldt, 2001). Propensity for Moving Nearbysecondary is achieved using a locallyevaluated Genetic Programming algorithm Satisfaction with the outcome of an election increases future turnoutwhich simultaneously develops arithmetic Positive reinforcement from voting will lead to further Discuss-politics-with person-23 blue expert=falsepredictors of a target output (voter turnout) voting (e.g. Bendor, Diermeier and Ting, 2003) . neighbour-network year=10 month=3and their scope. Lots-family-discussions year=10 month=2 Etc. Voting can be hindered by personal shocks The birth of a child disturbs habit (Plutzer, 2002). An Agent’s Memory of Events Voting varies with age What Next? The output (clusters and expressions) suggests hypotheses that can then: (a) be checked using specific simulation experiments and using Declining health, mobility, and energy levels impede standard statistical tests (b) be explored in simpler and more abstract models (in particular to capture any significant “phase” changes that these voting (e.g. Strate et al. 1989) indicate.