 |
 |
Well, I can't help but offer my two cents. First of two postings. This commentary from my perspective as someone involved in land-use modeling. Others may have more to add. On Jun 15, 2009, at 5:12 AM, Scott Moss wrote: > For me, the "meta-point" of the "what is the point" discussion is to > find a framework within which to evaluate the different approaches to > social simulation: agent based computational economics, econo- and > sociophysics, companion modelling, empirically driven policy > modelling, > and so on. I believe it to be an important problem that there are > parallel approaches to social simulation -- some though not all > embedded > in different conventional disciplines and some using specific > modelling > techniques -- with no constructive and persistent communication > amongst > them. So the point of the following three points is to suggest a > framework as a basis for such communication. It would be useful (at > least to me) to hear from adherents of these and other approaches to > social simulation whether these points apply to them and if not, then > what is missing or misguided. In the land-use modeling field, there are many both structured and informal means for modelers using different techniques to communicate, and there have been some deliberate efforts to compare modeling methods. There is, however, perhaps much more focus on the development of hybrid models that use multiple techniques for model calibration, parameterization, verification, and validation. But the discussions about verification and validation span alternative modeling methods. My perspective on the points below: > > > _Point 1:_ > > No one has argued that we can predict extreme events or policy impacts > though some argue that we should aim at such prediction or > forecasting. There have been some examples of successful prediction offered up. More on that separately. > Perhaps the first point should be to identify if outcomes can be > treated as being drawn from some underlying or persistent distribution > of events and whether that distribution satisfies the conditions for > forecasting or prediction to be feasible. I agree with this. > In either event: > > **our models should produce outcomes with distributions of (for > example) > absolute percentage first differences (= relative changes) in time > series that are not systematically distinguishable from the > distributions of relative changes in corresponding social > statistics.** > This is a bit specific and narrow--I am more an adherent of the more general pattern-oriented modeling perspective suggested by Grimm et al., with validation targets identified to match the problem domain, policy question, and/or goal of the model. > _Point 2:_ > > Our inability to forecast significant changes in social processes as > measured by indicator statistics coheres with evidence from fine grain > data series (e.g., financial markets and markets for fast moving > consumer goods) that the distribution of relative changes in such time > series data is leptokurtic (= fat-tailed = heavy tailed) and therefore > the law of large numbers does not apply. But if the goal is to reproduce distributions, the fat-tailed distributions do _not_ signal an inability to forecast. The mean of any distribution only happens sometimes; so basing predictions even on the mean of a normal distribution does not make any sense if you are interested in forecasting a particular event. > A natural and important task > is therefore > > **to identify the characteristics of social behaviour that would > explain > the observation of frequency distributions Yes, I agree, we need to develop process-based models that reproduce and explain observed outcome distributions. An example of this (that points out some of the modeling challenges) is replication of fractal patterns of land use. > that do not support > forecasting or prediction of important changes in statistical > indicators.** I don't understand this statement and how it relates to the first point. If we can reproduce a distribution, aren't we a step further towards prediction? > > _Point 3:_ > > There are multiple (and probably many) ways in which models can > generate > leptokurtic (heavy-tailed) distributions of relative changes. > Therefore, we need > > **to determine a set of criteria for choosing among alternative > representations of social processes yielding leptokurtic frequency > distributions of relative changes.** Yes, absolutely, and the many varieties of model that produce fractal land-use patterns is a perfect example. Grimm et al. suggest using multiple pattern targets-- for example, two models may each produce the same pattern outcomes, but only one might also replicate housing price dynamics in some acceptable way. > Dawn Cassandra Parker Assistant Professor, Department of Computational Social Science, Kransnow Institute for Advanced Study; Affiliate, Departments of Environmental Science and Policy, Geography, and Geoinformation and Earth Systems Science George Mason University 374 Research 1 4400 University Drive, MS 6B2 Fairfax, VA, USA 22030 +1-703-993-4640 (phone) +1-703-993-9290 (fax) dparker3 at gmu dot edu http://mason.gmu.edu/~dparker3