Hi all,
I have a theoretical question.
I'm currently working on a optimization model (the DICe model) . One of the goals of this model is to define the optimal policies for reducing emission, so I can use VENSIM to optimize and find the optimal values of these policies.
That's fine.
Let's imagine now I have also historical data for some variables. I'd like to calibrate some parameters of the model to those historical data.
Is it correct to run the optimization for calibration purposes firstly and then, once I've got the parameter values, run the optimization again to find the optimal policy?
What's the differences, pros and cons of this approach versus using time series techniquest to estimate the parameters and then put the values in the VENSIM equations?
I hope I was clear...
Thanks in advance
Giovanni
Combining optimization for calibration and policy evaluation
Dice model
Hi Giovanni
I am not a specialist of ecology. So I can only give you some ideas.
Coyle’s book studies well the optimization problem and advices to have a good knowledge
of the model by using it through repeated simulations prior to using optimization techniques
that can be used in very different ways: what objective functions to choose, how to define
the structural parameters that will drive the generation of adequate policies, etc…
When you study this you understand that optimization and calibration must be used with care
and necessitates a very good understanding of the subject and the model.
I looked at the DICE4 model. It has 161 symbols (above my level of ability of mastering a model within a reasonable time)
65 constants and 11 data which are probably exogenous input.
What is its purpose?
One should first determine which constants are unchangeable, which are known, how many are
unknown and need to be known, which have the main influence on the behavior relative to the
purpose of the model. If there are a lot of constants that need to be found out and are significant,
calibration will be difficult and may give equivalent solutions relative to the objective but very
different from one another.
I doubt that there is a definitive answer to your question. It depends on the subject and one solution
is to use both techniques and compare the solutions. Using regression techniques does not take
dynamic effects into consideration and then calibration may be more appropriate, but in theory.
One solution is to calibrate putting only one parameter at a time into the optimization plane and try
to get insights from that experience. I personally think that SD is a tool for thinking and learning
about a subject, and helping choose robust policies, but not for giving definitive results as the future
is always unknown.
One solution is to simplify the dice model to start with an easier model to understand and progressivily add more material only if one is sure of its usefulness.
There are not too many loops and they are relatively short.
But they should be well studied to get a feel of the overall dynamic of the model.
Regards.
JJ
I am not a specialist of ecology. So I can only give you some ideas.
Coyle’s book studies well the optimization problem and advices to have a good knowledge
of the model by using it through repeated simulations prior to using optimization techniques
that can be used in very different ways: what objective functions to choose, how to define
the structural parameters that will drive the generation of adequate policies, etc…
When you study this you understand that optimization and calibration must be used with care
and necessitates a very good understanding of the subject and the model.
I looked at the DICE4 model. It has 161 symbols (above my level of ability of mastering a model within a reasonable time)
65 constants and 11 data which are probably exogenous input.
What is its purpose?
One should first determine which constants are unchangeable, which are known, how many are
unknown and need to be known, which have the main influence on the behavior relative to the
purpose of the model. If there are a lot of constants that need to be found out and are significant,
calibration will be difficult and may give equivalent solutions relative to the objective but very
different from one another.
I doubt that there is a definitive answer to your question. It depends on the subject and one solution
is to use both techniques and compare the solutions. Using regression techniques does not take
dynamic effects into consideration and then calibration may be more appropriate, but in theory.
One solution is to calibrate putting only one parameter at a time into the optimization plane and try
to get insights from that experience. I personally think that SD is a tool for thinking and learning
about a subject, and helping choose robust policies, but not for giving definitive results as the future
is always unknown.
One solution is to simplify the dice model to start with an easier model to understand and progressivily add more material only if one is sure of its usefulness.
There are not too many loops and they are relatively short.
But they should be well studied to get a feel of the overall dynamic of the model.
Regards.
JJ
Optimization
Hi JJ,
Thanks for the insights.
I'll certanly follow your suggestion to work on smallest portion of the model to get a better understanding.
Can you give me the title of the book you are referring to?
Ciao!
G
Thanks for the insights.
I'll certanly follow your suggestion to work on smallest portion of the model to get a better understanding.
Can you give me the title of the book you are referring to?
Ciao!
G
Dice model
Hi
There are two ways to simplify the model.
The first is to take away some parts of the model and the second is too study different parts of the model independantly from one another.
These two approaches will give different results.
The second one must be done with care as there may be loops flowing through different parts of the model.
Separating the model into different parts may suppress the
dynamic effects of these loops.
It is easier to split a model with short loops than with longer ones. The dice model having mainly short loops, spliting it should be easier
I think that there is one long loop.
The book is 'system dynamic modelling: a practical approach' and can be bought through Amazon.
It is in the short bibliography from the SDS web site:
http://www.systemdynamics.org/short_bibliography.htm
It is the best book I have read about system dynamics.
It is better to translate the examples in the book into Vensim models which is an excellent exercise. I have done it.
Regards.
JJ
[Edited on 6-2-2009 by LAUJJL]
There are two ways to simplify the model.
The first is to take away some parts of the model and the second is too study different parts of the model independantly from one another.
These two approaches will give different results.
The second one must be done with care as there may be loops flowing through different parts of the model.
Separating the model into different parts may suppress the
dynamic effects of these loops.
It is easier to split a model with short loops than with longer ones. The dice model having mainly short loops, spliting it should be easier
I think that there is one long loop.
The book is 'system dynamic modelling: a practical approach' and can be bought through Amazon.
It is in the short bibliography from the SDS web site:
http://www.systemdynamics.org/short_bibliography.htm
It is the best book I have read about system dynamics.
It is better to translate the examples in the book into Vensim models which is an excellent exercise. I have done it.
Regards.
JJ
[Edited on 6-2-2009 by LAUJJL]