Comments on: The Obscure Art of Datamodeling in Vensim http://blog.metasd.com/2009/11/the-obscure-art-of-datamodeling-in-vensim/?utm_source=rss&utm_medium=rss&utm_campaign=the-obscure-art-of-datamodeling-in-vensim Don't just do something, stand there! (Sometimes good policy in complex systems is counterintuitive) Sat, 18 May 2013 10:24:05 +0000 hourly 1 http://wordpress.org/?v=3.1.2 By: Bill Harris http://blog.metasd.com/2009/11/the-obscure-art-of-datamodeling-in-vensim/comment-page-1/#comment-1729 Bill Harris Fri, 04 Dec 2009 04:14:35 +0000 http://blog.metasd.com/?p=524#comment-1729 Perhaps put another way, I sense Vensim's approach uses Bayesian analysis as developed for control theory, while MCSim uses Bayesian analysis as developed for physics. I'm curious if they give essentially the same results or if there are times to choose one or the other (and what those criteria might be). Perhaps put another way, I sense Vensim’s approach uses Bayesian analysis as developed for control theory, while MCSim uses Bayesian analysis as developed for physics. I’m curious if they give essentially the same results or if there are times to choose one or the other (and what those criteria might be).

]]> By: Bill Harris http://blog.metasd.com/2009/11/the-obscure-art-of-datamodeling-in-vensim/comment-page-1/#comment-1728 Bill Harris Fri, 04 Dec 2009 04:09:03 +0000 http://blog.metasd.com/?p=524#comment-1728 Hi, Tom. In essence, MCSim does what OpenBUGS (http://mathstat.helsinki.fi/openbugs/) does, except that it integrates that with a differential equation solver using the Gear algorithm. http://en.wikipedia.org/wiki/MCSim and http://fredomatic.free.fr/ describe it, and the manual is at http://www.gnu.org/software/mcsim/mcsim.html. http://www.gnu.org/software/mcsim/mcsim.html#SEC49 begins to describe the setup of the Bayesian analysis, although it really helps to know a bit about MCMC sampling first and perhaps to find a few of Frederic Bois' journal articles. (It also helps to find the quick reference card on my site, if you want to use it as a DYNAMO equivalent.) And, yes, R is often one's good friend. While I've tended to use J both to set up more complex MCSim experiments and to analyze the data, I could have done that in R or Gnuplot, too, and R has a nice interface to OpenBUGS, as you probably know. Bill Hi, Tom.

In essence, MCSim does what OpenBUGS (http://mathstat.helsinki.fi/openbugs/) does, except that it integrates that with a differential equation solver using the Gear algorithm. http://en.wikipedia.org/wiki/MCSim and http://fredomatic.free.fr/ describe it, and the manual is at http://www.gnu.org/software/mcsim/mcsim.html. http://www.gnu.org/software/mcsim/mcsim.html#SEC49 begins to describe the setup of the Bayesian analysis, although it really helps to know a bit about MCMC sampling first and perhaps to find a few of Frederic Bois’ journal articles. (It also helps to find the quick reference card on my site, if you want to use it as a DYNAMO equivalent.)

And, yes, R is often one’s good friend. While I’ve tended to use J both to set up more complex MCSim experiments and to analyze the data, I could have done that in R or Gnuplot, too, and R has a nice interface to OpenBUGS, as you probably know.

Bill

]]> By: fiddaritt http://blog.metasd.com/2009/11/the-obscure-art-of-datamodeling-in-vensim/comment-page-1/#comment-1720 fiddaritt Thu, 03 Dec 2009 17:00:51 +0000 http://blog.metasd.com/?p=524#comment-1720 Hi Bill - I'm not familiar enough with MCSim to comment. A Kalman filter is essentially a Bayesian updating process for the model state, but I'm guessing that you mean something different. Now I'm curious ... I used to have a bookmark to a modeling software package that aimed at total replicability, but I can't find it at the moment. However, there are lots of ways to skin the cat. R scripts are fairly complete, for example. Tom Hi Bill -

I’m not familiar enough with MCSim to comment. A Kalman filter is essentially a Bayesian updating process for the model state, but I’m guessing that you mean something different. Now I’m curious …

I used to have a bookmark to a modeling software package that aimed at total replicability, but I can’t find it at the moment. However, there are lots of ways to skin the cat. R scripts are fairly complete, for example.

Tom

]]> By: Bill Harris http://blog.metasd.com/2009/11/the-obscure-art-of-datamodeling-in-vensim/comment-page-1/#comment-1681 Bill Harris Tue, 01 Dec 2009 03:22:27 +0000 http://blog.metasd.com/?p=524#comment-1681 Tom, very good comments. Most of what I want to say is "Me, too," so I'll refrain from that level of detail, but there are two points I'd like to make. First, I think the notion of data transparency and spreadsheets extends past data modeling for SD models. Even though I'm not as skilled as I'd like to be with it, I've used the (free) array language J (http://www.jsoftware.com/) in lieu of spreadsheets whenever I can for the past few years, and it's been quite a good ride. To stay on topic, I have used it to prepare data for SD models, and I have used it to analyze data from such models. My second point is more of a query than a point to make. I, too, see value in applying good data analysis (call it statistics, if you will) to SD model results in at least certain cases. For a variety of reasons, I've been exploring a Bayesian analysis approach using MCMC and the MCSim simulator for the past few years. I'm not an expert in the approach Vensim uses, but I gather it's more closely allied to Kalman filtering and that thread of inquiry. I'm curious if you've got any insights into how the two approaches might differ in terms of the insights they provide. Thanks for the ideas you share here. BTW, I don't want to take away from Vensim by mentioning MCSim; I suspect the use of multiple tools can help us become better at this stuff. There was research by Bill Curtis at the old MCC that suggested a correlation between the number of languages a programmer used and the programmer's skill at software. Perhaps the same thing applies to us. Bill Tom, very good comments. Most of what I want to say is “Me, too,” so I’ll refrain from that level of detail, but there are two points I’d like to make.

First, I think the notion of data transparency and spreadsheets extends past data modeling for SD models. Even though I’m not as skilled as I’d like to be with it, I’ve used the (free) array language J (http://www.jsoftware.com/) in lieu of spreadsheets whenever I can for the past few years, and it’s been quite a good ride. To stay on topic, I have used it to prepare data for SD models, and I have used it to analyze data from such models.

My second point is more of a query than a point to make. I, too, see value in applying good data analysis (call it statistics, if you will) to SD model results in at least certain cases. For a variety of reasons, I’ve been exploring a Bayesian analysis approach using MCMC and the MCSim simulator for the past few years. I’m not an expert in the approach Vensim uses, but I gather it’s more closely allied to Kalman filtering and that thread of inquiry. I’m curious if you’ve got any insights into how the two approaches might differ in terms of the insights they provide.

Thanks for the ideas you share here. BTW, I don’t want to take away from Vensim by mentioning MCSim; I suspect the use of multiple tools can help us become better at this stuff. There was research by Bill Curtis at the old MCC that suggested a correlation between the number of languages a programmer used and the programmer’s skill at software. Perhaps the same thing applies to us.

Bill

]]>