Combining Discrete and Continous

 Junior Member
 Posts: 14
 Joined: Fri Mar 29, 2002 3:39 am
Combining Discrete and Continous
Posted by ""Louis Macovsky, Dynamic BioSystems"" <dynbiosys@verizon.net> Greetings,
Somewhat related to the discussion of comparing modeling approaches or platforms...
Have any of you used a hybrid of discrete event modeling with SD?
The subject matter that I am modeling is for the most part in the recent literature modeled as petri nets using bayesian mathematics. I have since begun the study of applying these methods to simulation modeling.
What I envision is the marriage of DEV such as petri nets and SD in that SD would apply feedback and its endogenous delays on one hierarchal level and DEV modeling with petri nets in submodels. There are programs such as AnyLogic that bring agent based modeling together with SD; although I have not used such a combination I would think that the ""emergence"" from this combination (SD & ABM) might or should be different than combining DEV with SD. As I proceed I may find that bringing these ""paradigms"" (DEV and SD) together in a single model and simulation is not necessary, practical, or even wise, what do you think?
On the subject of bayesian math, have you or do you explicitly incorporate specific probability theory mathematics into your models (outside the builtin statistical functions and if/then/else statements provided in the SD software)?
Lou
Louis Macovsky
Dynamic BioSystems, LLC
Wilsonville, OR, USA
Posted by ""Louis Macovsky, Dynamic BioSystems"" <dynbiosys@verizon.net> posting date Tue, 27 Dec 2005 10:19:09 0800
Somewhat related to the discussion of comparing modeling approaches or platforms...
Have any of you used a hybrid of discrete event modeling with SD?
The subject matter that I am modeling is for the most part in the recent literature modeled as petri nets using bayesian mathematics. I have since begun the study of applying these methods to simulation modeling.
What I envision is the marriage of DEV such as petri nets and SD in that SD would apply feedback and its endogenous delays on one hierarchal level and DEV modeling with petri nets in submodels. There are programs such as AnyLogic that bring agent based modeling together with SD; although I have not used such a combination I would think that the ""emergence"" from this combination (SD & ABM) might or should be different than combining DEV with SD. As I proceed I may find that bringing these ""paradigms"" (DEV and SD) together in a single model and simulation is not necessary, practical, or even wise, what do you think?
On the subject of bayesian math, have you or do you explicitly incorporate specific probability theory mathematics into your models (outside the builtin statistical functions and if/then/else statements provided in the SD software)?
Lou
Louis Macovsky
Dynamic BioSystems, LLC
Wilsonville, OR, USA
Posted by ""Louis Macovsky, Dynamic BioSystems"" <dynbiosys@verizon.net> posting date Tue, 27 Dec 2005 10:19:09 0800

 Junior Member
 Posts: 14
 Joined: Fri Mar 29, 2002 3:39 am
Combining Discrete and Continous
Posted by Michael McDevitt <mmcdevitt@caci.com>
Louis,
I worked with others in my company to link a SIMPROCESS Discrete Event Model with a Vensim System Dynamics Model. For more information on CACI's hybrid systems architecture (HSA) and SIMPROCESS please contact me.
All the Best,
Mike McDevitt
Project Manager
CACI Dynamic Systems Inc.
10085 Scripps Ranch Court
San Diego, CA 92131
Posted by Michael McDevitt <mmcdevitt@caci.com>
posting date Wed, 28 Dec 2005 08:24:06 0800
Louis,
I worked with others in my company to link a SIMPROCESS Discrete Event Model with a Vensim System Dynamics Model. For more information on CACI's hybrid systems architecture (HSA) and SIMPROCESS please contact me.
All the Best,
Mike McDevitt
Project Manager
CACI Dynamic Systems Inc.
10085 Scripps Ranch Court
San Diego, CA 92131
Posted by Michael McDevitt <mmcdevitt@caci.com>
posting date Wed, 28 Dec 2005 08:24:06 0800

 Junior Member
 Posts: 14
 Joined: Fri Mar 29, 2002 3:39 am
Combining Discrete and Continous
Posted by ""Johnson, Scott T"" <Scott.Johnson2@bp.com>
You might find the following paper interesting (2002 Palermo
Conference):
Alternative Modeling Approaches: A Case Study in the Oil & Gas Industry, Scott Johnson, Bob Eberlein
The combination of SD and discrete event modeling is becoming common in BP.
Scott T Johnson, BP, Houston
Posted by ""Johnson, Scott T"" <Scott.Johnson2@bp.com>
posting date Wed, 28 Dec 2005 13:28:12 0600
You might find the following paper interesting (2002 Palermo
Conference):
Alternative Modeling Approaches: A Case Study in the Oil & Gas Industry, Scott Johnson, Bob Eberlein
The combination of SD and discrete event modeling is becoming common in BP.
Scott T Johnson, BP, Houston
Posted by ""Johnson, Scott T"" <Scott.Johnson2@bp.com>
posting date Wed, 28 Dec 2005 13:28:12 0600

 Junior Member
 Posts: 14
 Joined: Fri Mar 29, 2002 3:39 am
Combining Discrete and Continous
Posted by ""Joseph M. DeFee"" <jdefee@caci.com>
Mike, et al,
To reply to Louis' last question about incorporating probability theory mathematics into our simulation models, this can be done through any of our external interface techniques. For example events in SIMPROCESS models can have scripts associated with them that can call local run time libraries, Java classess, remote java calls (RMI), or web services via SOAP calls. By providing this capability, any method of introducing behavior can be integrated into a SIMPROCESS model. In fact, this is how we implemented the Vensim models with SIMPROCESS.
Mike C  please feel free to add comments if you have some other ideas.
thanks,
joe.
Posted by ""Joseph M. DeFee"" <jdefee@caci.com>
posting date Wed, 28 Dec 2005 10:51:58 0600
Mike, et al,
To reply to Louis' last question about incorporating probability theory mathematics into our simulation models, this can be done through any of our external interface techniques. For example events in SIMPROCESS models can have scripts associated with them that can call local run time libraries, Java classess, remote java calls (RMI), or web services via SOAP calls. By providing this capability, any method of introducing behavior can be integrated into a SIMPROCESS model. In fact, this is how we implemented the Vensim models with SIMPROCESS.
Mike C  please feel free to add comments if you have some other ideas.
thanks,
joe.
Posted by ""Joseph M. DeFee"" <jdefee@caci.com>
posting date Wed, 28 Dec 2005 10:51:58 0600

 Junior Member
 Posts: 14
 Joined: Fri Mar 29, 2002 3:39 am
Combining Discrete and Continous
Posted by ""Tobias Lorenz"" <space56@freenet.de>
Dear Louis,
We have been experimenting with combining all three methodologies (DES, ABM and SD) within single modells and i'd say it depends very much on the problem. Of course u always run the risk to loose the problem focus and modell a system. In order to not run into that trap it is necessary to have a very clear idea how the methodologies add their different core ideas to the explanation of the problem. If you don't want to make the model to complex you should check each part of the methodology upon its explanatory power. DES for example might be very useful to replicate organisational structures which can then be structured with feedback loops. Other possibilities might be to have some output variable of the interaction of some agents which can then be fed back to them, etc. I'd say there is a lot of integration possibilities but the idea behind each module of such an integrative model has to be checked carefully in order no to loose the problem focus. I would definitly like to exchange about this problems further if you should decide to proceed in that direction but would prefer not to do this over the list as these problems are research in progress at my employer...
I am not familiar with statistical tools but would definitly like to learn more about that...
Best regards
Tobias Lorenz DaimlerChrysler Research and Technology
Posted by ""Tobias Lorenz"" <space56@freenet.de>
posting date Wed, 28 Dec 2005 17:19:57 +0100
Dear Louis,
We have been experimenting with combining all three methodologies (DES, ABM and SD) within single modells and i'd say it depends very much on the problem. Of course u always run the risk to loose the problem focus and modell a system. In order to not run into that trap it is necessary to have a very clear idea how the methodologies add their different core ideas to the explanation of the problem. If you don't want to make the model to complex you should check each part of the methodology upon its explanatory power. DES for example might be very useful to replicate organisational structures which can then be structured with feedback loops. Other possibilities might be to have some output variable of the interaction of some agents which can then be fed back to them, etc. I'd say there is a lot of integration possibilities but the idea behind each module of such an integrative model has to be checked carefully in order no to loose the problem focus. I would definitly like to exchange about this problems further if you should decide to proceed in that direction but would prefer not to do this over the list as these problems are research in progress at my employer...
I am not familiar with statistical tools but would definitly like to learn more about that...
Best regards
Tobias Lorenz DaimlerChrysler Research and Technology
Posted by ""Tobias Lorenz"" <space56@freenet.de>
posting date Wed, 28 Dec 2005 17:19:57 +0100

 Junior Member
 Posts: 14
 Joined: Fri Mar 29, 2002 3:39 am
Combining Discrete and Continous
Posted by ""Geoff McDonnell"" <gmcdonne@bigpond.net.au>
Lou
In health systems simulation it is not uncommon to use DES/DEV with SD in modeling patient flows esp thru ED, ICU and ORs. DEV matches the way data is collected, but SD better represents the feedback effects associated with work pressure and other dynamic supplydemand interactions. The problem is (as John Morecroft described in a great paper at ISDC2005 referred to earlier in this thread) some people are trained (compelled) to see the world as ""constrained randomness"" while others (fewer!) focus on the SD paradigm of ""dynamic complexity"" Anylogic and the like allow combinations of DES, ABM, SD and bayesian nets, or whatever else. See http://www.xjtek.com/files/press/bnet.pdf for combination with Bayesian belief networks V5.4 Anylogic also has an example Agent based MMs model comparing AB and DES queues which is quite instructive. Also this week's Lancet Infectious Disease has a review of MAthematical
modelling: a tool for hospital infection control
(Lancet Infect Dis 2006; 6:3945 ) which discusses combinations of modeling approaches (which are called deterministic, stochastic, core groups and game theoretical approaches here !)
It is often wiser to start with a model using one method, then progressively add features of other methods. We find
that switching feedback effects on and off (such as the effect of ED wait
time on Did Not Shows) is often more convincing to domain experts than including them from the start.
regards
geoff
Dr Geoff McDonnell
Director Adaptive Care Systems
Simulation Research Fellow Centre for Health Informatics University of New South Wales Posted by ""Geoff McDonnell"" <gmcdonne@bigpond.net.au> posting date Thu, 29 Dec 2005 06:36:49 +1100
Lou
In health systems simulation it is not uncommon to use DES/DEV with SD in modeling patient flows esp thru ED, ICU and ORs. DEV matches the way data is collected, but SD better represents the feedback effects associated with work pressure and other dynamic supplydemand interactions. The problem is (as John Morecroft described in a great paper at ISDC2005 referred to earlier in this thread) some people are trained (compelled) to see the world as ""constrained randomness"" while others (fewer!) focus on the SD paradigm of ""dynamic complexity"" Anylogic and the like allow combinations of DES, ABM, SD and bayesian nets, or whatever else. See http://www.xjtek.com/files/press/bnet.pdf for combination with Bayesian belief networks V5.4 Anylogic also has an example Agent based MMs model comparing AB and DES queues which is quite instructive. Also this week's Lancet Infectious Disease has a review of MAthematical
modelling: a tool for hospital infection control
(Lancet Infect Dis 2006; 6:3945 ) which discusses combinations of modeling approaches (which are called deterministic, stochastic, core groups and game theoretical approaches here !)
It is often wiser to start with a model using one method, then progressively add features of other methods. We find
that switching feedback effects on and off (such as the effect of ED wait
time on Did Not Shows) is often more convincing to domain experts than including them from the start.
regards
geoff
Dr Geoff McDonnell
Director Adaptive Care Systems
Simulation Research Fellow Centre for Health Informatics University of New South Wales Posted by ""Geoff McDonnell"" <gmcdonne@bigpond.net.au> posting date Thu, 29 Dec 2005 06:36:49 +1100

 Junior Member
 Posts: 14
 Joined: Fri Mar 29, 2002 3:39 am
Combining Discrete and Continous
Posted by Bill Harris <bill_harris@facilitatedsystems.com>
""Louis Macovsky Dynamic BioSystems dynbiosys verizon.net"" writes:
>> Have any of you used a hybrid of discrete event modeling with SD?
Lou,
While I know what you mean, I view discrete (vs. continuous) and SD (vs. nonSD) as close to orthogonal dimensions in describing modeling. If SD is all about accumulation and feedback (as I've claimed in earlier discussions on this topic and John Sterman has substantiated), then there's no reason a Petri Net model, for example, can't also be an SD model. Places certainly accumulate tokens; all you need is feedback to have it be SD.
While I think we typically get benefit from the 10,000 meter view that comes by viewing the oftenlumpy real world as continuous, I don't think that's fundamental. I still have on my todo list the gaining of enough ML proficiency to do a useful SD model using CPN Tools.
Thoughts?
...........
See, for example, Paul Fishwick's _Simulation Model Design and Execution_ and his concept of multimodeling (http://www.cise.ufl.edu/~fishwick/resea ... earch.html).
Bill
 
Bill Harris
Facilitated Systems
Posted by Bill Harris <bill_harris@facilitatedsystems.com>
posting date Fri, 30 Dec 2005 22:09:42 0800
""Louis Macovsky Dynamic BioSystems dynbiosys verizon.net"" writes:
>> Have any of you used a hybrid of discrete event modeling with SD?
Lou,
While I know what you mean, I view discrete (vs. continuous) and SD (vs. nonSD) as close to orthogonal dimensions in describing modeling. If SD is all about accumulation and feedback (as I've claimed in earlier discussions on this topic and John Sterman has substantiated), then there's no reason a Petri Net model, for example, can't also be an SD model. Places certainly accumulate tokens; all you need is feedback to have it be SD.
While I think we typically get benefit from the 10,000 meter view that comes by viewing the oftenlumpy real world as continuous, I don't think that's fundamental. I still have on my todo list the gaining of enough ML proficiency to do a useful SD model using CPN Tools.
Thoughts?
...........
See, for example, Paul Fishwick's _Simulation Model Design and Execution_ and his concept of multimodeling (http://www.cise.ufl.edu/~fishwick/resea ... earch.html).
Bill
 
Bill Harris
Facilitated Systems
Posted by Bill Harris <bill_harris@facilitatedsystems.com>
posting date Fri, 30 Dec 2005 22:09:42 0800

 Junior Member
 Posts: 14
 Joined: Fri Mar 29, 2002 3:39 am
Combining Discrete and Continous
Posted by =?iso88591?Q?JeanJacques_Laubl=E9?= <jeanjacques.lauble@wanadoo.fr> Hi Lou
I have never tried to mix discrete event modelling to SD or even not tried to use discreet event modelling at all. Discrete event modelling is interested at the behaviour of individual events. It looks at reality very closely as explained by Bill Harris, and models reality exactly as it works. If you want to look at the behaviour of the whole system you have to aggregate the behaviours of all the individuals events.
SD makes the aggregation sooner with the inputs.
Practically this difference involves differences in both methods. Due to the closeness, the boundaries and the time horizon are smaller in discreet event. This tools is then less appropriate for strategic issues where many components need to be modelled. It is easier to use because it replicates reality very closely and generally during a short time period and it is then very easy to validate the models. Feed backs are not generally taken into consideration in discreet event maybe because they have not the time to build up. And feed backs make the study of any model much more complicate to build and to analyze. It is the reason why discreet event is more used then SD, I think.
My problems could have been studied with discreet event modelling but being more interested by the overall understanding of the reality then by operational functioning I preferred SD.
As to the possibility to mix both methods, it sounds to me difficult, because you cannot at the same time have large and short boundaries and horizon.
It is maybe possible to use submodels inside a SD time step, to describe reality closer. I have looked at the anylogic website and was interested how one could combine SD and discrete event modelling whether by agent based modelling or discrete event modelling, but to my perception most of the examples where dealing with discrete and I did not found one that was mixing SD and discrete. Maybe is has changed since, I must have a look. Regards. J.J. Laublé Allocar Strasbourg France. Posted by =?iso88591?Q?JeanJacques_Laubl=E9?= <jeanjacques.lauble@wanadoo.fr> posting date Sat, 31 Dec 2005 17:13:49 +0100
I have never tried to mix discrete event modelling to SD or even not tried to use discreet event modelling at all. Discrete event modelling is interested at the behaviour of individual events. It looks at reality very closely as explained by Bill Harris, and models reality exactly as it works. If you want to look at the behaviour of the whole system you have to aggregate the behaviours of all the individuals events.
SD makes the aggregation sooner with the inputs.
Practically this difference involves differences in both methods. Due to the closeness, the boundaries and the time horizon are smaller in discreet event. This tools is then less appropriate for strategic issues where many components need to be modelled. It is easier to use because it replicates reality very closely and generally during a short time period and it is then very easy to validate the models. Feed backs are not generally taken into consideration in discreet event maybe because they have not the time to build up. And feed backs make the study of any model much more complicate to build and to analyze. It is the reason why discreet event is more used then SD, I think.
My problems could have been studied with discreet event modelling but being more interested by the overall understanding of the reality then by operational functioning I preferred SD.
As to the possibility to mix both methods, it sounds to me difficult, because you cannot at the same time have large and short boundaries and horizon.
It is maybe possible to use submodels inside a SD time step, to describe reality closer. I have looked at the anylogic website and was interested how one could combine SD and discrete event modelling whether by agent based modelling or discrete event modelling, but to my perception most of the examples where dealing with discrete and I did not found one that was mixing SD and discrete. Maybe is has changed since, I must have a look. Regards. J.J. Laublé Allocar Strasbourg France. Posted by =?iso88591?Q?JeanJacques_Laubl=E9?= <jeanjacques.lauble@wanadoo.fr> posting date Sat, 31 Dec 2005 17:13:49 +0100

 Junior Member
 Posts: 14
 Joined: Fri Mar 29, 2002 3:39 am
Combining Discrete and Continous
Posted by ""Tobias Lorenz"" <space56@freenet.de>
Dear JeanJacques,
I have been thinking about the relation between discrete and continuous for a while myself now and i think there are cases, where also from a strategic or aggregated point of view, the discreteness of some phenomenon is essential for the understanding of a problem. Think of the introduction of a new law, this will sometime cause fluctuations, think of the last drop that makes a barrell overflow in whatever kind of problem issue. Sometimes there are certain states which will cause a system to change drastically and then they should be included, in SD software this probably happens via IfThenElse functions or equivalent logical functions, but upon using those, on should consider that they are not essential SD, they are the some kind of DiscreteEventSimulation, so that closely regarded every model using this kind of functions would already be a combined modell...
What do you think?
Best regards
Tobias Lorenz
Posted by ""Tobias Lorenz"" <space56@freenet.de>
posting date Mon, 2 Jan 2006 17:32:15 +0100
Dear JeanJacques,
I have been thinking about the relation between discrete and continuous for a while myself now and i think there are cases, where also from a strategic or aggregated point of view, the discreteness of some phenomenon is essential for the understanding of a problem. Think of the introduction of a new law, this will sometime cause fluctuations, think of the last drop that makes a barrell overflow in whatever kind of problem issue. Sometimes there are certain states which will cause a system to change drastically and then they should be included, in SD software this probably happens via IfThenElse functions or equivalent logical functions, but upon using those, on should consider that they are not essential SD, they are the some kind of DiscreteEventSimulation, so that closely regarded every model using this kind of functions would already be a combined modell...
What do you think?
Best regards
Tobias Lorenz
Posted by ""Tobias Lorenz"" <space56@freenet.de>
posting date Mon, 2 Jan 2006 17:32:15 +0100

 Junior Member
 Posts: 14
 Joined: Fri Mar 29, 2002 3:39 am
Combining Discrete and Continous
Posted by =?iso88591?Q?JeanJacques_Laubl=E9?= <jeanjacques.lauble@wanadoo.fr> Hi Tobi
It is a question of definition.
In continuous systems, the main characteristic is the aggregation of time into periods. You generally aggregate the inputs too but it is not an absolute necessity. In Powersim you can by example integrate discreete events and you can do it in Vensim too while being not managed by the software. But both use aggregated time period. In DES the time is not aggregated and discreet. In the two cases you mention the law case is still continuous because the overall model is continuous with one discreet event. One could say that in this case the model is the addition of two strictly continuous models. The second case, the drop is something different. It is a singularity in a continuous system, like the one you find in continous equations. This is too a continuous system. One could anynhow imagine hybrid systems with some data continuous or some discreet, but the important thing will be the treatment of the time, aggregated or not. And it will always be difficult to take a holistic continuous approach and an event centered approach at the same time. Doing it makes you aim at two very different even contradictory objectives at the same time and will make you build overly complicated models, reaching none of the objectives at all. It is like wanting to have a map with a definition of say 1/10000 and representing a territory like Germany. You will have a map that will be 100 meter large. Regards. JeanJacques Laublé Allocar Strasbourg France Posted by =?iso88591?Q?JeanJacques_Laubl=E9?= <jeanjacques.lauble@wanadoo.fr> posting date Tue, 3 Jan 2006 14:48:19 +0100
It is a question of definition.
In continuous systems, the main characteristic is the aggregation of time into periods. You generally aggregate the inputs too but it is not an absolute necessity. In Powersim you can by example integrate discreete events and you can do it in Vensim too while being not managed by the software. But both use aggregated time period. In DES the time is not aggregated and discreet. In the two cases you mention the law case is still continuous because the overall model is continuous with one discreet event. One could say that in this case the model is the addition of two strictly continuous models. The second case, the drop is something different. It is a singularity in a continuous system, like the one you find in continous equations. This is too a continuous system. One could anynhow imagine hybrid systems with some data continuous or some discreet, but the important thing will be the treatment of the time, aggregated or not. And it will always be difficult to take a holistic continuous approach and an event centered approach at the same time. Doing it makes you aim at two very different even contradictory objectives at the same time and will make you build overly complicated models, reaching none of the objectives at all. It is like wanting to have a map with a definition of say 1/10000 and representing a territory like Germany. You will have a map that will be 100 meter large. Regards. JeanJacques Laublé Allocar Strasbourg France Posted by =?iso88591?Q?JeanJacques_Laubl=E9?= <jeanjacques.lauble@wanadoo.fr> posting date Tue, 3 Jan 2006 14:48:19 +0100

 Junior Member
 Posts: 14
 Joined: Fri Mar 29, 2002 3:39 am
Combining Discrete and Continous
Posted by Bill Harris <bill_harris@facilitatedsystems.com>
""JeanJacques Laublé jeanjacques.lauble wanadoo.fr"" <systemdynamics@VENSIM.COM> writes:
>> It is a question of definition.
>> ...
>> In Powersim you can by example integrate discreete events and you can
>> do it in Vensim too while being not managed by the software.
Just a caution: SD is not defined as that which Vensim or Powersim or ithink does.
Bill
 Bill Harris
Facilitated Systems
Posted by Bill Harris <bill_harris@facilitatedsystems.com>
posting date Wed, 04 Jan 2006 22:32:33 0800
""JeanJacques Laublé jeanjacques.lauble wanadoo.fr"" <systemdynamics@VENSIM.COM> writes:
>> It is a question of definition.
>> ...
>> In Powersim you can by example integrate discreete events and you can
>> do it in Vensim too while being not managed by the software.
Just a caution: SD is not defined as that which Vensim or Powersim or ithink does.
Bill
 Bill Harris
Facilitated Systems
Posted by Bill Harris <bill_harris@facilitatedsystems.com>
posting date Wed, 04 Jan 2006 22:32:33 0800