Behavioral microsimulation from SIMSOC List
From SystemsWiki
From Edmund Chattoe: http://www.sociology.ox.ac.uk/people/chattoe.html
At 9:17 am -0800 23/2/06, Bill Clancey wrote:
Our Brahms workpractice simulation system models "micro" behaviors (typically at 5 second increments), including communications, movements, use of facilities, and tools (e.g., camera). Brahms integrates with other simulations, such as life support and vehicle systems, to allow modeling of human-systems interaction. We have been working on Brahms for 13 years. We have simulated in particular a variety of mission operations including ground personnel and robots on the moon and Mars.
See http://www.agentisolutions.com, plus publications at http://bill.clancey.name
My recent publication, "The cognitive modeling of social behaviors," specifically addresses how to relate human factors (e.g., fatigue, hunger) to group behaviors (e.g., during meetings) and the coordination of work at a fine-grained level (including body posture), relating this to models of reasoning and attention.
At 7:17 pm -0500 23/2/06, Professor Ron Sun wrote:
for behavioral (or cognitive) social simulation:
R. Sun and I. Naveh, Simulating organizational decision-making using a cognitively realistic agent model. Journal of Artificial Societies and Social Simulation, Vol.7, No.3. 2004. http://jasss.soc.surrey.ac.uk/7/3/5.html
Ron Sun (Ed.), Cognition and Multi-Agent Interaction: Extending Cognitive Modeling to Social Simulation. Cambridge University Press, 2005.
see also the Web page: http://www.cogsci.rpi.edu/~rsun/mas.html
At 8:25 pm -0800 23/2/06, Carl Henning Reschke wrote:
Since I probably do not know the literature broad and deep enough, just a comment on the limits: on the mental side : perception / creativity / associative abilities / intuition. Relevant: Probably most the stuff from the BCL (von Foerster + coworkers), G. Günther, Rudolf Kaehr (now at Aberdeen), Bateson, Mead, would probably involve your sim agents to possibly become schizophrenic, paranoid and what else - which brings me to:
Limits inherent in computer programming: Only in German and more pop science than science: Peter Krieg Die paranoide Maschine - english summary on his / PILE website. In the end starts to define the (cognitive / mental side of the issue), but breaks off - somewhat unsatisfying. Relates to Borges and alle the others on library organization. Issue is information ordering, relating and flexibility (or lack thereof) in interpreting it.
how to do it nevertheless:
Behavior (directly): maybe Bruce Edmonds has sth.
Mentally: Gigerenzer / Peter Todd - theme of "Fast and Frugal heuristics"
Carl Henning http://www.stratevol.blogspot.com
At 10:37 am +0000 24/2/06, Paul Williamson wrote:
The leading UK expert in this field is: http://www.nottingham.ac.uk/economics/staff/details/alan_duncan.htm
At 12:14 pm +0100 24/2/06, Eva Ebenhoeh wrote:
Ebenhöh, E. 2006. Modeling non-linear common-pool resource experiments with boundedly rational agents, in: Sichman, J.S., Antunes, L. (eds.) 2006. MABS 2005, LNAI 3891, Springer-Verlag, 133-146
The approach is also described on http://www.usf.uos.de/~eebenhoe/forschung/adaptivetoolbox/
At 9:45 am -0800 24/2/06, Doug Samuelson wrote:
My impression, having done a bit of each, is that agent-based modelers and microsimulators tend not to talk to each other much.
At 2:33 pm -0500 26/2/06, Kim Bloomquist wrote:
Recently the IRS has developed several new simulation tools that incorporate significant behavioral elements. Some recent papers can be found on the IRS website. This link (http://www.irs.gov/pub/irs-soi/toder.pdf) is to a paper that describes a microsimulation model of taxpayer burden used to estimate the time and money cost to individuals of filing taxes (i.e., "lodging" taxes in British parlance). The IRS uses the model to evaluate the impact on burden of proposed changes to tax laws.
These next two papers (http://www.irs.gov/pub/irs-soi/04jacnta.pdf and http://www.irs.gov/pub/irs-soi/04desnta.pdf) are not simulation models but describe recent behavioral research that attempts to shed light on taxpayers' responses to a change in enforcement environment. The approach used is that of experimental economics, but the results could help to construct better microsimulation models of taxpayer behavior.
This paper (http://www.irs.gov/pub/irs-soi/04blonta.pdf) is one I wrote last year and recently published in the 2004 Proceedings of the National Tax Association. It describes a proto-type agent-based model to explore taxpayers' possible responses to changes in enforcement parameters, such as audit rate, penalty for underreporting and income "visibility." The model incorporates several behavioral elements including overweighting of audit probabilities, rudimentary social networks and agent heterogeneity.
Finally, I would point that President Bush's recent budget calls for specific funding of "dynamic" (read "behavioral") analysis to be applied to revenue forcasting in the Treasury Department. Although such behavioral analysis has been conducted for some years by Treasury, Congressional Budget Office and others to forecast the growth impacts of tax reduction, the present Administration wants to push this form of analysis to help bolster its tax proposals (or so my former boss, Dr. Eric Toder of the Urban Institute, seems to think and I concur).
At 11:48 am -0500 3/3/06, Snipper, Reuben (HHS/OS) wrote:
My office, the Assistant Secretary for Planning and Evaluation in the U.S. Department of Health and Human Services, uses large microsimulation models of the type mentioned in your email, especially TRIM (See http://trim.urban.org) for the purposes you do. Below I share some thoughts.
We have considered implementing behavioral effects into these models. The first difficulty we always face is that social science/economic research into the size and character of these effects is not very precise or specific.
For example, the Negative Income Tax Experiments (SIME/DIME, http://aspe.hhs.gov/hsp/SIME-DIME83/) are usually cited as showing a 10% reduction in work effort. When the results are examined closely however, this turns out to be quite misleading. First, the effect differs substantially by sex. Second, the results are not that everyone reduced their work effort by 10%, but rather that most people did not reduce their work effort at all and a small percentage reduced their effort a lot. Even further, those who reduced their work effort did so by lengthening the time between jobs, rather than by reducing their hours of work.
Such complicated results for large, social experiments are quite common and make it difficult to incorporate the results into microsimulation models. Less rigorous research methods have yielded widely differing effects among studies depending on the methods, questions, populations, etc., used in the study.
In other words, until research can provide more precise results, we haven't found it useful to include behavioral effects; their inclusion would just make the microsimulation results even less certain than they already are. Other issues that have deterred us include that the behavioral effects add a lot of complexity to each module and that we have to update the parameters of our models regularly (often annually) and the research is not updated that often.
If you want to pursue implementing behavioral effects into your microsimulation models, you might consider using dynamic models, since they can incorporate behavioral effects as well as other processes (especially demographics) together, in a consistent manner.
At 12:12 pm -0700 24/2/06, Black, Janice wrote:
I am currently working on and have had published an agent based model of a work group and its development of a culture that supports organizational learning. In this model individuals influence the group level which in turn influences the individuals across time. Our model has as its output several different things but in particular the group level support of a context that supports learning. You can read an article on our model in Leadership Quarterly the Jan 06 issue (here is the citation.)
Black, J. A., Oliver, R. L., Howell, J. P. & King, J. P. 2006. A Dynamic System Simulation of Leader and Group Effects on Context for Learning, The Leadership Quarterly, 17(1), 39-56.
The link is to a digital version of the article. I don't know if this is what you were interested in finding out but If I can help you in the modeling from the individual up I would be glad to share what we've been doing. We are currently working on the expansion of our model from two levels (ind & group) to three and more levels (ind-group-organization; ind-group-organization-multi-organization).
At 12:26 pm -0700 24/2/06, Ugo Merlone wrote:
I try to incorporate behavior into my agents, grounding the modeling process on classroom experiments.
The approach is described in http://jasss.soc.surrey.ac.uk/7/3/2.html
I have been using that approach in several models and now I am trying to give a stronger methodological basis relating to qualitive approaches.
At 1:50 pm -0600 24/2/06, William Lawless wrote:
you may be thinking about behavioral game theory; my colleague nick feltovich an economist at u houston works in that area and he might be able to help you; from my perspective, a new theory on uncertainty is needed because the general notion of a "rational" agent has broken down for several years; for a good review, see shafir, a colleague of tversky:
(Shafir, E., & LeBoeuf, R.A. (2002). "Rationality." Annual Review of Psychology 53: 491-517);
plus, here's a quote from mandel, business week online last october:
DEFYING EXPECTATIONS. In fact, Daniel Kahneman and Vernon Smith won the 2002 Nobel Prize in Economics for their work on behavioral and experimental economics. The writeup that accompanied their award observed:
Real-world decision-makers frequently appear not to evaluate uncertain events according to the laws of probability; nor do they seem to make decisions according to the theory of expected-utility maximization. In a series of studies, Kahneman -- in collaboration with the late Amos Tversky -- has shown that people are incapable of fully analyzing complex decision situations when the future consequences are uncertain.
SAYING TOO MUCH. In other words, Kahneman and Smith won their 2002 prize precisely for showing that people mostly don't behave the way that game theory assumes they do. Game theory is based on a finely honed sort of reasoning: "If I do this, then he'll do that, then I'll do this" ad infinitum, assessing the probability of different final outcomes. In reality, though, that's not how most people think or make decisions.
At 3:47 pm -0600 25/2/06, Guastello, Stephen wrote:
By making simulations "behavioural" are you asking how to establish their external validity with real-world behavior? If so the problem was defined over a decade ago, and to my chagrin, not addressed by the mainstay of "complexity" simulation research. The best reference I can offer is:
Guastello, S. J., & Rieke, M. L. (1994). Computer-based test interpretations as expert systems: Validity and viewpoints from artificial intelligence theory. Computers in Human Behavior, 10, 435-455.
At 9:10 pm -0500 26/2/06, Michael Wolfson wrote:
We (Statistics Canada) have been selling a cross-sectional model (the Social Policy Simulation Database and Model = SPSD/M) for about 20 years. It has a small but devoted user base, including federal and provincial finance and policy ministries, think tanks and a few academics.
You are correct in saying that folks have been "restive" for some time about the lack of behaviour in the models. However, this is not for want of being aware of the issue nor of trying. Rather it is a combination of (a) the research literature not having converged to consensus results; (b) the complexity of confronting the relatively simplified views of behaviour in the academic literature (e.g. micro econometrics of labour supply) with the complexity of actual program structures and tax provisions; and (c) difficulty communicating results inclusive of behavioural change to senior policy makers.
Our strategy for some years has been to leave the behavioural response analysis to our users in a post hoc or add on mode. To help them in this kind of analysis, the SPSD/M not only computes the "income effect" (i.e. the change in disposable income), but also a "price effect" (i.e. the change in effective marginal tax rate, though note that this is not always easy to define).
These topics were well covered in a 1991 or 1992 panel study on microsimulation carried out by the National Academy of Science in Washington DC (of which I was a member).
Statistics Canada has also been developing a family of dynamic longitudinal microsimulation models, one called LifePaths for social policy, the other called POHEM for POpulation HEalth Model. These do include behavioural equations, but no feedback from policy changes to behavioural patterns. These are both "industrial strength" models intended to support government policy analysis. The models are certainly open and extensible enough to all the inclusion of "endogenous behaviours". But it is proving to be more than enough work just to get the behaviours specified and estimated, even from "inside the cookie jar" of a national statistical agency that is among the best endowed in the world with richly multivariate microdata sources. (e.g. see Will et al, British Journal of Cancer, 2001 85(9) "First do no harm...").
More recently, we have been developing (in partnership with researchers in various countries) models of infectious disease spread. These have populations of co-evolving individuals who form "contact networks" dynamically, based on evolving characteristics. The most advanced application has been for alternative childhood vaccination policies in third world countries. Another is for the interaction between HIV and TB in sub-Saharan Africa.
Some years ago, we did one that was more theoretical, focusing on the impact of quality change on the measurement of economic growth -- with disturbing conclusions. It grew out of the Boskin investigation of problems with the CPI in the US, and was published in the Canadian Journal of Economics in April 1999, and a shorter version in R. Conte, R. Hegselmann and P. Terna (Eds.), Simulating Social Phenomena, Lecture Notes in Economics and Mathematical Systems 456, Springer. This has the most endogenous behaviour of any of our microsimulation models, with each individual endowed with a utility function, and using satisficing rules of thumb to try to maximize their (heterogeneous) utilities, but not always succeeding.
Michael Wolfson Assistant Chief Statistician, Analysis and Development Statistics Canada
At 12:34 pm -0500 24/2/06, Corinne Coen wrote:
I have a paper forthcoming in OBHDP (attached) that uses a laboratory experiment to measure people's real behavior in the context of a social dilemma. Subsequently, I use a simulation to extract a decision rule. I am beginning to work on a follow up paper that explores how a population of simulated people using the extracted rule behave. (There are intriguing emergent effects.) Perhaps this is something you are interested in? Regards, Corinne Coen SUNY at Buffalo
[Please contact her for the paper: "Seeking the Comparative Advantage: The Dynamics of Individual Cooperation in Single vs. Multiple Team Environments"]
At 2:40 pm +0100 24/2/06, Martine Antona wrote:
In this case the behaviour is represented through individudal educational and occupational choices (P.15). I guess this example is not "more behavioral" than the others you mentionned ; but just in case ... http://cormas.cirad.fr http://www.cirad.fr/ur/green
[Please contact her for the paper" Policy Options for Meeting the Millennium Development Goals in Brazil: Can micro-simulations help?"]
| Systems Thinking World Discussions Questions & Comments to Geoff McDonnell |
