Aggregation Techniques

[Jean-Jacques Laublé jean-jacques]

Posted by Hi every body.


I am currently working on a model where I need

to aggregate the demand of the customers in a minimum of

classes while the behaviour of these customers relative to the

price is similar. My activity : short term car rental.

The criterion of classification is simple: they should have the same

probability to buy at a specific level of price.

The customers have different characteristics.

1.. Level of anticipation (the delay between the date of inquiry and the moment they will need the service in days.)
2.. The fact of having already bought our service in the past.
3.. The duration of the rent.
4.. The kilometres by day
5.. The position of the rent in the week, in the year or in holidays, to accommodate for the seasonality of the demand.
6.. The type of customer, private, business or association.
7.. The type of vehicle.
8.. Eventually the location of the agency.
9.. And maybe one to three others depending on the type of vehicle.

I have all the data that resume the past experiences about 100000 events every year that tells if a certain type of demands, has bought or not with a certain price.

I will have to model each type of vehicle one after the other for planning considerations, so the statistics will not go further then 10000 events by vehicle.

It is necessary for me to aggregate these demands so as to simulate different price politics and see the results in term of planning utilisation, availability, cash flow etc.


I have browsed on the SD mailing list, to see if there were any threads on classification, clustering, taxonomy, typology or aggregation discussing on the different ways to classify to better the aggregation process. I did not find anything.


I browsed on the net only on classification and found an universe of sites and software that use a lot of different techniques, based on rigorous methods but not equally adapted to all the subjects.

I decided to download some free software, easy to understand to make me familiar with the subject at first.

Does anybody have any experience on that subject?

Thank you in advance and best regards.

J.J. Laublé. Allocar, rent a car

Strasbourg France. Posted by posting date Sun, 16 Jan 2005 12:31:03 +0100

[Jean-Jacques Laublé jean-jacques]

Posted by Hi

I received a directly sent e-mail, that make me think that I have not well explained the aggregation problem.

This is the answer to my question and what I answerd back.

First answer:

Is this useful?

""To determine whether activities taking place serially can be aggregated, consider the average residence time of items in each stock (the average time between entering and exiting the stock). Stocks with short residence times relative to the time scale for the dynamics of interest generally do not need to be represented explicitly and can either be omitted or lumped into adjacent stocks."" A short residence time in two stocks relative to the time scale makes it unnecessary to distinguish between whether a unit is in one stock or the other.

""Parallel activities can legitimately be aggregated together if the individual flows are governed by similar decision rules and if the time the different items spend in the individual stocks is similar."" If the decision rules and times are the same, then for all practical purposes the flows are the same.



my back answer:



Hi



I thank you first for having made the effort to read my query.

I have browsed the Sterman's book and read the chapter 6 paragraphs on aggregation.

There is another place where he refers to the Forrester's market growth model it is in chapter 3, page 100 'defining the system boundary and the degree of aggregation are two of the most difficult steps in successful modelling.'



But all this does not solve my problem which I think needs further explanation.

I will try to explain it with a simplistic though unrealistic example:



Suppose that you have customers who want to rent cars and you want to determine the optimal pricing policy.

You know that there are mainly two characteristics in the demand:

the duration of each individual renting and the kilometres by day.

Suppose that you have only 6 experiences, and that you intend to consider

people that have short, middle and long durations, and short, middle and high

kilometres by day.

If you want to determine the behaviour of any demand based on the characteristics

you need to study the average behaviour of all the demands that share the same characteristics. There are 9 combinations of the different values of the characteristics,

And only 6 observations, which does not even make one observation for each classification

based on the characteristics. That does not make a valid statistic. I must then find another way to determine the behaviour of any demand. The only method is to construct another classification than the one based on each value of the characteristics, probably which a much smaller number of classes to make it possible to calculate valid statistics for them.



Example:

Experiences Duration of renting kilometres by day price proposition has bought or not

1. low middle 100 yes

2. high low 105 yes

3. high low 120 no

4. middle high 100 no

5. high high 90 no

6. high middle 70 yes



One can see that even with this simple case, it is not evident to find the classification that fits

better to the experiences

One can for example visually say that for a two classes partition, the first class fits better than the second one in the following two classes.

First classification

Class 1 : (low,middle) + (high,low)

Classe 2 : (middle,high) + (high,high) + (high middle)



Second classification :

Class 1: (low,middle) + (high,high) + (middle,high)

Class 2: (high,low) + (middle,high)



The first classification has a more homogenous behaviour relative to the price than the second one. It will then be easier to determine an average price behaviour for both classes of the

first classification.

Imagine now having 8 characteristics and 5000 observations.

There are rigorous methods that define precisely an order relation between different classifications that permits to find the classification that is better then the other.

There are many methods of classifications using different algorithms: trees, neural networks

etc.

I hope that I may have better explained my case.

Feel free to ask more details if something is not clear in my explanation.

Regards.

J.J. Laublé. Allocar rent a car.

Strasbourg France.
Posted by posting date Tue, 18 Jan 2005 09:26:31 +0100

[Geoff McDonnell gmcdonne bigpond]

Posted by ""Geoff McDonnell"" <gmcdonne@bigpond.net.au>
J.J. Laublé. Allocar rent a car wrote:.
you need to study the average behaviour of all the demands that share the
same characteristics. There are 9 combinations of the different values of
the characteristics,

And only 6 observations, which does not even make one observation for each
classification

based on the characteristics. That does not make a valid statistic. I must
then find another way to determine the behaviour of any demand. The only
method is to construct another classification than the one based on each
value of the characteristics, probably which a much smaller number of
classes to make it possible to calculate valid statistics for them.

Jean, I believe you need to use discrete choice methods to obtain valid
statistics. You may want to check out the work of Jordan Louviere, Ken
Train, and David Henscher in this area. Also you may find the problem easier
to conceptualise and model using a combination of SD and Agent based
methods, such as AnyLogic software from www.xjtek.com .
regards
geoff
Dr Geoff McDonnell
Director Adaptive Care Systems
Simulation Research Fellow Centre for Health Informatics UNSW
AUSTRALIA

gmcdonne@bigpond.net.au
Posted by ""Geoff McDonnell"" <gmcdonne@bigpond.net.au>
posting date Thu, 20 Jan 2005 16:23:47 +1100

[Jean-Jacques Laublé jean-jacques]

Posted by =?iso-8859-1?Q?Jean-Jacques_Laubl=E9?= <jean-jacques.lauble@wanadoo.fr>
Hi Geoff and thank you Geoff for your answer.



I have consulted the books from Louviere and Train on Amazon.

They are very close to my preoccupations. The only problem is with

the high density of long formulas mixing the habitual integrals with
probability

functions etc. Not that I do not understand them, but I fear that these
people

are more concerned with the theory than with VERY CONCRETE applications.

Do you know of any books about the same subjects but very applications
oriented.

I am definitely not interested by the academic side of the subject, but by
getting

concrete results.

About anylogic and agent based methods I have consulted their Web site and
read through the interesting presentation that anylogic made at the last SD
conference last year about the agent based methods.

I do not see what value is added by agent based modelling in my case.

I will have anyway to determine the behaviour of the customers, which needs
statistics whatever the method used. There are not so many agents and they
do not interact with each others, and if I can aggregate their behaviour, I
will be able to model using a traditional SD software. My model is
approximately finished, works and needs only the behaviour of the demand
that I generate stochastically to be calculated.

I have until now not been annoyed by the fact that the software I use,
Vensim, has only limited discreet features. I may later on, but I do not
have the time now to invest studying another software, especially if it is
based on different paradigms.

Another point to be considered is that studying the different classification
and clustering techniques is much easier to study than a new dynamic
paradigm. I know the time it took me to be able to produce usable dynamic
models and it is not finished.

Anyhow thanks for the help.

Regards.

J.J. Laublé. Allocar

Strasbourg. France.

Posted by =?iso-8859-1?Q?Jean-Jacques_Laubl=E9?= <jean-jacques.lauble@wanadoo.fr>
posting date Thu, 20 Jan 2005 17:05:47 +0100

[John Gunkler jgunkler sprintmail]

Posted by ""John Gunkler"" <jgunkler@sprintmail.com>
Jean-Jacques,

If I understand what you're doing (and I'm still not sure I do), I might
suggest that you first do statistical analyses of your data to discover how
many categories you really need to describe their probability of buying at a
particular price. Just because you CAN distinguish customers by demographic
(or other) characteristics doesn't mean that those differences make a
difference in their behavior -- and statistical analysis (maybe ANOVA? maybe
factor analysis?) is good at telling you which differences actually make a
difference.

Many, many times I have had clients tell me that ""the people on the east
coast of the U.S. act differently than those on the west coast"" and ""women
will obviously act differently than men"" and ""everyone knows that production
people have different opinions than salespeople"" (etc.) -- yet, when we do
the study we often find that these differences made no real difference.
This is good news for aggregation, because it says you CAN aggregate people
and still have their aggregated behavior be a good picture of how they act.

One of the things I don't understand is why you say you don't have enough
data. Don't you have 10,000 instances of renting decisions? You don't have
more than about 2,000 combinations of characteristics to study, do you?

Even if you do, there are experimental designs (for ANOVA - analysis of
variance) that allow you to test ""main effects"" separately (e.g., whether
duration of rental has an effect on willingness to buy at a particular
price) even though you may not be able to test ALL of the interactions
between main effects. You may discover that you only need two classes of
""duration"" (long, short) instead of three; or you may discover that duration
doesn't significantly effect willingness to purchase at a price at all.
Then, once you've determined which main effects matter, you can then
re-analyze the data for the interactions that seem most likely to matter.
Some ways of doing this are more legitimate than others (statistically) --
that is, sometimes you are supposed to analyze a new set of data rather than
re-analyzing the same set. But a very brilliant professor of mine once told
me, ""First, you must convince yourself that something may be happening that
warrants your further efforts. Don't worry too much about being rigorous
during the 'exploratory' phase of your research. Just dig into the data and
see what seems to be happening, if anything. Then, when you have a
hypothesis that seems worth testing, do it 'the right way' to demonstrate to
others if you are right.""

Posted by ""John Gunkler"" <jgunkler@sprintmail.com>
posting date Thu, 20 Jan 2005 08:56:53 -0600

[Andy Ford FordA mail.wsu.edu]

Posted by ""Andy Ford"" <FordA@mail.wsu.edu>
Folks,

I like Geoff McDonnell's suggestion to J. J. Lauble about aggregation
methods for simulation choice among different modes of transporation (such
as a particular rental car, the bus, etc.). In my opinion, this is the
domain of the discrete choice modelers, people like Ken Train and Dan
McFadden. I believe they would approach the discrete choice model with a
nested, multi-nominal logit model and J. J. Lauble's question on aggregation
would be rephrased as a question of the degree of nesting in the
multi-nomial logit model. They can do their work with either revealed
preferences (revealed by real decisions on which car to rent) or on stated
preferences. I think their approach is especially appealing with there are
more than two choices. When they come up with a good model of customer
choice among discrete options, we can implement their model of customer
choice within a larger, system dynamics model of the system that is
probably of interest to J. J. Lauble.

Andy Ford
Professor
Program in Environmental Science & Regional Planning
Washington State University
Pullman, WA 99164-4430

FordA@mail.wsu.edu
(509) 335-7846
http://www.wsu.edu/~forda
Posted by ""Andy Ford"" <FordA@mail.wsu.edu>
posting date Thu, 20 Jan 2005 12:17:56 -0800

[Andy Ford FordA mail.wsu.edu]

Posted by ""Andy Ford"" <FordA@mail.wsu.edu>
Folks,

I like Geoff McDonnell's suggestion to J. J. Lauble about aggregation
methods for simulation choice among different modes of transporation (such
as a particular rental car, the bus, etc.). In my opinion, this is the
domain of the discrete choice modelers, people like Ken Train and Dan
McFadden. I believe they would approach the discrete choice model with a
nested, multi-nominal logit model and J. J. Lauble's question on aggregation
would be rephrased as a question of the degree of nesting in the
multi-nomial logit model. They can do their work with either revealed
preferences (revealed by real decisions on which car to rent) or on stated
preferences. I think their approach is especially appealing with there are
more than two choices. When they come up with a good model of customer
choice among discrete options, we can implement their model of customer
choice within a larger, system dynamics model of the system that is
probably of interest to J. J. Lauble.

Andy Ford
Professor
Program in Environmental Science & Regional Planning
Washington State University
Pullman, WA 99164-4430

FordA@mail.wsu.edu
(509) 335-7846
http://www.wsu.edu/~forda
Posted by ""Andy Ford"" <FordA@mail.wsu.edu>
posting date Thu, 20 Jan 2005 12:17:56 -0800

[Jean-Jacques Laublé jean-jacques]

Posted by =?iso-8859-1?Q?Jean-Jacques_Laubl=E9?= <jean-jacques.lauble@wanadoo.fr>
Hi Joel



Thank your for your answer.

Generally the dynamic of a system can already be

studied at a high aggregation level, and going into details

does not necessarily add a great deal of value.


But in my case, the problem is that it is impossible to determine

the elasticity function of aggregated products whose behaviours are
heterogeneous.

If you try anyhow, you will find absurd behaviours like the same customer
that will buy a product say at 100 dollars won't buy it any more at 70
dollars. The elasticity function you get is no more coherent and goes
through multiple minimum and maximum.

If you force your elasticity function to be more coherent, you get other

aberrations. Expensive products will be always bought whatever the price
offered

and cheap one's will never be.

To get a coherent result you need to aggregate the products so that their

behaviour towards price is sufficiently identical.



I am sure that the problem must exist in other circumstances, where it is
mandatory to disaggregate to a certain level.

Here is where clustering techniques would help a lot.

Thank you again for your answer and best regards.

J.J. Laublé. Allocar

Strasbourg France.
Posted by =?iso-8859-1?Q?Jean-Jacques_Laubl=E9?= <jean-jacques.lauble@wanadoo.fr>
posting date Thu, 20 Jan 2005 17:33:23 +0100

[Jean-Jacques Laublé jean-jacques]

Posted by =?iso-8859-1?Q?Jean-Jacques_Laubl=E9?= <jean-jacques.lauble@wanadoo.fr>
Hi John



Thank you for answering to my question.


<If I understand what you're doing (and I'm still not sure I do), I might
<suggest that you first do statistical analyses of your data to discover how
<many categories you really need to describe their probability of buying at
a
<particular price. Just because you CAN distinguish customers by
demographic
<(or other) characteristics doesn't mean that those differences make a
<difference in their behavior



Perfectly understood.



< and statistical analysis (maybe ANOVA? maybe
< factor analysis?) is good at telling you which differences actually make a
< difference.
Simple regression usually is enough.

Sometimes just looking at the averages is enough.

When there is a great difference between two means, it is not absolutely

necessary to use ANOVA to test if the difference is significant or not.

Not to mention that in that case I do not need ANOVA at all, reasoning on
all

the observations (no sample).

For instance, over a total of 100000 observations (last year), if an
eventual

customer, has already been a client in the past, the probability of buying
is about

the double of one that has never rented before in our company.


< Many, many times I have had clients tell me that ""the people on the east
<coast of the U.S. act differently than those on the west coast"" and ""women
<will obviously act differently than men"" and ""everyone knows that
production
<people have different opinions than salespeople"" (etc.) -- yet, when we do
<the study we often find that these differences made no real difference.



The attributes I have chosen are all significant, I have tested them one
after

the other with simple regression analysis if the attribute is continuous or
other means otherwise.



<This is good news for aggregation, because it says you CAN aggregate people
<and still have their aggregated behavior be a good picture of how they act.
<One of the things I don't understand is why you say you don't have enough
<data. Don't you have 10,000 instances of renting decisions?



I have plenty of observations and I do not see where I said I hade not
enough data.

I have currently in a year 100000 observations.



You don't have
<more than about 2,000 combinations of characteristics to study, do you?
I have much more that 2000 'natural combinations' to study.

Generally for a continuous characteristic I cannot go under 3 classes

Little, middle, big.

For instance in duration, having 3 levels is probably not enough as duration
is

currently from one day to 90.



If one supposes that the average class by characteristic is three, this
makes 3 power 8



And 3 power 8 is about 9 power 4 equal more than 6400 Classes, combined by
Cartesian product.

But 2000 or 6000 does not change the problem. I need to make a minimum of
clustering

to get valid data.


<Even if you do, there are experimental designs (for ANOVA - analysis of
<variance) that allow you to test ""main effects"" separately (e.g., whether
<duration of rental has an effect on willingness to buy at a particular
<price) even though you may not be able to test ALL of the interactions
<between main effects. You may discover that you only need two classes of
<""duration"" (long, short) instead of three; or you may discover that
duration
<doesn't significantly effect willingness to purchase at a price at all.
<Then, once you've determined which main effects matter, you can then
<re-analyze the data for the interactions that seem most likely to matter.
<Some ways of doing this are more legitimate than others (statistically) --
<that is, sometimes you are supposed to analyze a new set of data rather
than
<re-analyzing the same set.

<I already did that job and kept only the characteristics that have enough
effects.





<But a very brilliant professor of mine once told
<me, ""First, you must convince yourself that something may be happening that
<warrants your further efforts. Don't worry too much about being rigorous
<during the 'exploratory' phase of your research. Just dig into the data
and
<see what seems to be happening, if anything. Then, when you have a
<hypothesis that seems worth testing, do it 'the right way' to demonstrate
to
<others if you are right.""



I can very easily dig into the data, as I have an agency near me, to make
all possible real

tests, to compare my hypothesis with real experiences. It is no more a
question of rigour

reality does not care about rigour.



There are three particular things to understand in my problem.



First: I have not a current inventory problem to resolve, like you find

in SD text books. I manage a planning, and contrary to an inventory, the
same

product (by example a particular day) can be sold at a different price,
depending on

the type of demand it is included in. It may be included in a demand for a
60 days location

and will not be sold with a high price, or as a stand alone Saturday, where
it can be sold

at a price four times higher. This is true for the anticipation.

If you want to sell the same product, 30 days in advance, you will have to
choose a price

that is four time less than if you sell it the day preceding the Saturday.

This particularity creates strong and short term dynamical feedbacks that
depend on the

pricing policy that you do not find with ordinary inventory problems.

The same case occurs with companies who have heavy fixed structural costs,
like airlines

companies, hotels and relatively small variable costs. We are between, We
have 40% fixed costs, 20 % half fixed and variable costs with the vehicles
(fixed vehicle costs that vary within months) and 40% full variable costs.



There is another second strong feedback generated by the fact that, if you
sell too much too quickly there is nothing more to sell, and you had
probably a too low price policy.





Second: Many SD problems try to understand the apparently inexplicable
behaviour of a system, and try to find an average policy that lessens the
negative effects of it.

This is not my problem. I try to find a pricing policy that needs to be
accurate enough to generate the supplementary margin that will increase
directly our benefit, or decrease our loss

(which is actually the case, our sector being in crisis). This accuracy
needs to be at the level of days, because we must find the optimal pricing
for every demand (of course in theory, we are in effect looking for a not so
bad pricing strategy and not an optimal one, that does not exist).



Third. Our problem has to do with pricing strategy which is dealing with
price elasticity.

Or I did not find a way to aggregate price elasticity of products that are
the same but can be sold differently at very different level of prices.

As I explain in a former mail, aggregating elasticity is very difficult
(impossible to my opinion if you want a price policy accurate enough)



I am actually downloading clustering and classification software and testing
them and until now, Bayesian techniques seem the most appropriate, but I am
only at the beginning of my research.

It has the advantage of finding automatically the best number of classes,
accepts non continuous characteristics example (older client or not, or
private, professional or association) , can handle missing values, (not
replaced by the average) and do not need the characteristics to be
independent (most important for me, some characteristics being not
independent from one another that forbids too making multiple regression
analysis).

I could use special techniques, to aggregate dependant characteristics, but
I am

looking for simple solutions.



Thank you again for your answer.

.......
Hi Andy



Thanks for answering to my question.


< I like Geoff McDonnell's suggestion to J. J. Lauble about aggregation
<methods for simulation choice among different modes of transportation (such
<as a particular rental car, the bus, etc.).



In fact the problem is not the choice among different ways of
transportations, but

between two solutions (buy or not buy a product) depending on the price.

There are cases where somebody can decide to use another solution, going by
train

instead of renting a car, but in 95% of the case, that kind of decision is
already taken

when people ask for a price. We have made post inquiries among people having
asked for

a price and not bought, and generally they reserved with a competitor.



<In my opinion, this is the domain of the discrete choice modelers, people
like Ken Train and <Dan McFadden. I believe they would approach the
discrete choice model with a
<nested, multi-nominal logit model and J. J. Lauble's question on
aggregation
<would be rephrased as a question of the degree of nesting in the
<multi-nomial logit model.



I tried to find books from Dan McFadden but did not find anything on Amazon.

I have to look for the definition of the multi-nominal logit model.



<They can do their work with either revealed
<preferences (revealed by real decisions on which car to rent) or on stated
<preferences. I think their approach is especially appealing with there are
<more than two choices. When they come up with a good model of customer
<choice among discrete options, we can implement their model of customer
<choice within a larger, system dynamics model of the system that is
<probably of interest to J. J. Lauble.



The same reflexion, my problems has to do with discreet choices, but
apparently with

only two choices. And instead of studying right away that problem with
difficult

to read books, I'd better study the matter with simple ideas first, simple
applications

too, and progressively step to a higher level of understanding if I find it
useful.

My initial question was, what classification technique is more appropriate
to my

classification problem among the many offered.

Loading software relative to the subject is a good way to learn about it.

Do you know of software, preferably downloadable freely at first, to test
it, that

deal with the subject of stated choices from Ken Train or Dan McFadden ?



Thank you for taking interest to my problem.


Regards.

J.J. Laublé Allocar

Strasbourg France


Posted by =?iso-8859-1?Q?Jean-Jacques_Laubl=E9?= <jean-jacques.lauble@wanadoo.fr>
posting date Fri, 21 Jan 2005 17:44:30 +0100

[Geoff McDonnell gmcdonne bigpond]

Posted by ""Geoff McDonnell"" <gmcdonne@bigpond.net.au>
Andy Ford wrote:

>> When they come up with a good model of customer
>> choice among discrete options, we can implement their model of customer
>> choice within a larger, system dynamics model of the system that is
>> probably of interest to J. J. Lauble.
>>

Thanks for expanding on my little knowledge of the area, Andy
The advantage of using multi-method object -oriented software is that the
practical problem can be modeled at multiple levels of aggregation. We are
using SD to form an aggregate macro view of context and agent based to
model individual behaviour. In AnyLogic the multi-nominal logit (or probit
or the like models) of discrete choice methods are represented by object
statechart transitions (e.g. from ""aware"" state to ""buy"" state) -- the
transition is the custom probability function estimated by the discrete
choice method. Nesting is easily represented by multi-level objects
representing subgroups. Each object can have multiple statecharts.

THe AnyLogic SD tutorial and Bass diffusion (SD and Agent based AB versions)
examples in the trial download from www.xjtek.com illustrate this.

J. J. Lauble is of course right in assuming that the best software is the
one you know .

However if you wish to represent the deep knowledge of a specific practical
area of interest that involves real people, then you may be faced with
modeling a combination of system constraints and individual behaviors. We
are finding combining SD and AB views of the world in the same head can be
quite difficult, so I'd advise to to try to team up with an AB head
regards
geoff
Dr Geoff McDonnell
Director Adaptive Care Systems
Simulation Research Fellow Centre for Health Informatics
University of New South Wales
Sydney AUSTRALIA
gmcdonne@bigpond.net.au





Posted by ""Geoff McDonnell"" <gmcdonne@bigpond.net.au>
posting date Sat, 22 Jan 2005 07:03:46 +1100

[Jean-Jacques Laublé jean-jacques]

Posted by =?iso-8859-1?Q?Jean-Jacques_Laubl=E9?= <jean-jacques.lauble@wanadoo.fr>
Geoff McDonnel wrote

<THe AnyLogic SD tutorial and Bass diffusion (SD and Agent based AB
versions)
<examples in the trial download from www.xjtek.com illustrate this.



I do not see the interest of studying the Bass diffusion examples of xjtek

when there is a 15 days trial version of their software.

I have worked with many computers tools (Algol, Cobol, Fortran, PL1,
Multiple basics, VBA, Assemblers, Macro Assemblers, Sql based languages etc.
along 40 years ) and I think that dynamics software are at least two levels
of difficulty higher than commonly used languages.

The level of technicality and experience needed is much higher with dynamic
oriented tools, due to the treatment of dynamic phenomenon.

15 days is ridiculous to see the real interest of such a tool.

Especially for me, as I have unfortunately not much time for my SD work.

I have currently other functions, being the Ceo of my business.

The tool you use is only a part of all the means used to resolve a problem.

Of course a more adapted tool is better. But it is too easy to be fooled,
and one must be sure to have the tool adapted to your case.

I have in the beginning of the eighties, switched from a basic implemented
by Pick, an English system, and working on a mini computer, to PL1, after
having asked an advice from an expert in the field.

After one year, working on a mainframe, I decided to switch a 5000 lines
program to a micro computer and in basic again with absolutely no problems.

I recently downloaded a trial version from Powersim to see how this more
discreet oriented software could help me, as I have, I recognize, some
discreetness in my problems.

I studied their first example, a simple bank account capitalisation scheme.

It said that if you wanted to do the same thing, you would get a very messy
model in a software having not the discreet features exposed.

I wrote a model in Vensim, giving exactly the same results, which was simple
than their model!

It dissuaded me to continue the trial, which I must recognize was a 60 days
trial, which is already better than 15 days.

I may try it again, if I have the time, but the maximum dimensions of
subscripts is only 2 in anylogic.

In my model, I have already 2 dimensioned arrays, and will probably add one
more, and eventually two more. Not to mention that simulating stochastic
data input, I am obliged for optimization purpose, to add another subscript,
generally 100 to 1000 dimensions, to launch parallel simulations with
different random numbers, from which I calculate the mean which I generally
optimize, but eventually standard deviations too.

That makes at least 4 dimensioned arrays. Or course these models are slow to
run.

I heard more then a year ago, that Vensim was working on stochastic
optimization, and I hope that it may help.

Anyhow I propose to send you if you are interested, a simplified example of
my model when completely tested and used, which is the more important part,
to show you how I modelled all this in standard SD. I cannot do it right
now, it needs to be translated and well documented. But it is always a good
thing to show one's work to more experienced people.

I can of course not get in the discussion about discreet choice methods, not
knowing them.



< J. J. Lauble is of course right in assuming that the best software is the
< one you know .



I did not quite say that. I said that changing a software is very expensive

and one must be sure of the utility of such an action.



Thank you for your answer and best regards.

J.J. Laublé. Allocar

Strasbourg. France.





Posted by =?iso-8859-1?Q?Jean-Jacques_Laubl=E9?= <jean-jacques.lauble@wanadoo.fr>
posting date Sat, 22 Jan 2005 12:56:54 +0100

[Geoff McDonnell gmcdonne bigpond]

Posted by ""Geoff McDonnell"" <gmcdonne@bigpond.net.au>

J.J. Laublé wrote>

>>I do not see the interest of studying the Bass diffusion examples of xjtek


>>when there is a 15 days trial version of their software.


>>15 days is ridiculous to see the real interest of such a tool.


To clarify, the software becomes save disabled at 15 days (our colleagues
have had this trial period extended on request).
I believe you can spend as long as you want on working through the existing
examples


>>In my model, I have already 2 dimensioned arrays, and will probably add one
>>more, and eventually two more. Not to mention that simulating stochastic
>>data input, I am obliged for optimization purpose, to add another

subscript,

>>generally 100 to 1000 dimensions, to launch parallel simulations with
>>different random numbers, from which I calculate the mean which I generally
>>optimize, but eventually standard deviations too.


>>That makes at least 4 dimensioned arrays. Or course these models are slow

to

>>run.


One of the main reasons we found Anylogic attractive is that instead of
using multi-dimensioned arrays, in many instances the problem can be modeled
more simply with multiple UMLstatecharts for a single object.

Also it is packaged with the Optquest genetic algorithm optimiser


>>Anyhow I propose to send you if you are interested, a simplified example of
>>my model when completely tested and used, which is the more important part,
>>to show you how I modelled all this in standard SD.


Of course I'd be delighted to see it when available. Why not submit it to a
future SD conference too, (if you can find the time)?

regards
geoff mcdonnell
Posted by ""Geoff McDonnell"" <gmcdonne@bigpond.net.au>
posting date Mon, 24 Jan 2005 08:07:37 +1100

[Jean-Jacques Laublé jean-jacques]

Posted by =?iso-8859-1?Q?Jean-Jacques_Laubl=E9?= <jean-jacques.lauble@wanadoo.fr>
Hi Geoff


< To clarify, the software becomes save disabled at 15 days (our colleagues


I may test the software if I have once the time.

< Of course I'd be delighted to see it when available.



I will send it to you and to anybody interested in that problem.



< Why not submit it to a future SD conference too, (if you can find the
time)?



About submitting something to the SD conference, I am not a SD professional,
only a user who makes his models by himself.

I do not have the expertise, to bring something relevant to the SD
community, and to justify the effort needed to do it. On top of that I think
that the people exposing with the posters, have

something more or less to sell, whether directly or through a lobbying
process or by some sort of interest. I think that anyhow, having never been
in a SD conference, I cannot have an opinion about it. I may go to the next
year conference in Nijmegen which is at three hours drive from Strasbourg,
Boston is too far away.



Besides I have not measured the real benefit of my modelling effort and I
can at the end, just find out that the differences generated by the various
policies, do not justify choosing one among the others and making the effort
to change. I would prefer talking or showing something that has effectively
brought substantial benefits (monetary or not).



This is the real problem I think about SD or related techniques. It is
difficult to estimate the

expected benefit and the expected time needed for such studies. If you know
of sources about this fundamental problem, I would be glad to know them.

Working in a competitive environment, I must unfortunately severely and very
realistically

orient my efforts towards what has the best chance to be profitable.

This means, more profits than costs and especially with a minimum of risk
(high average expected profit with low standard deviation).

I must acknowledge that this is not the case with my SD project, which is
still more based on

personal trust and intellectual interest than on strict management
principles.

I cannot for instance expose the project to my board of administrators, who
would instantly

ask me how I justified the investment effort (mainly my personal time).

I may once expose it, but only when I will have demonstrated the
effectiveness of the

method.



To go back to the aggregation problem, I will do it with pencil and papers.

I prefer to stick to the following principle: not use something that I do
not master

and going little step by little step. I fear that using statistical
techniques may generate that

harmful distance Yaman Barlas was talking about coming from the use of too
sophisticated tools.

Posted by =?iso-8859-1?Q?Jean-Jacques_Laubl=E9?= <jean-jacques.lauble@wanadoo.fr>
posting date Mon, 24 Jan 2005 11:38:55 +0100

[Jean-Jacques Laublé jean-jacques]

Posted by =?iso-8859-1?Q?Jean-Jacques_Laubl=E9?= <jean-jacques.lauble@wanadoo.fr>
Hi Geoff
I think that the last part of my last message was not sent. Probably an end
of page character in the text when pasting.

< I asked a world expert on DCM whether he could recommend anything more
< practically relevant to your problem. Here is his answer
<Rental car companies in the US of A have been using DCM to predict demand

>> <for cars for years. What is less obvious is whether one can predict the
>> <rental DURATION and TIMING, which is what he's asking. So, as far as I

< know

>> is that DCM's accurately predict the average (mean) duration, but have
>> nothing to say about time of day or stock. If you knew the mean and the
>> variance, and you had historic data on duration and time of day, this

<would be a piece of cake.
< he implied he was too busy just doing this stuff for large companies to be
< of much help

Sorry but I do not know what DCM is.
I have studied mainly American rent a car software, like TSD or Blue Bird.
One is using a revenue management add-in who is provided by the Rubicon
Group, specialized in yield management. I had three years ago contacted the
Rubicon group to have more details about their solutions. They were using
very simple methods, mainly moving up and down prices depending on the level
of already recorded future utilisations (the backlog of reservations).
I have already used such a method for years, which is better than nothing
but whose justification is very empirical.
I have of course all the possible data about past experiences, including the
time, and for years.
But it is not so much the prediction of demand that is difficult to do, but
its classification and calculating the average behaviour of the class
towards the price.

< That's about all on this subject for now. Bonne chance!
Thank you for your support, and I would prefer not need any chance, and be
sure in advance of the results I will get as I explained before!
Best regards.
J.J. Laublé. Allocar, rent a car
Strasbourg, France.





Posted by =?iso-8859-1?Q?Jean-Jacques_Laubl=E9?= <jean-jacques.lauble@wanadoo.fr>
posting date Tue, 25 Jan 2005 14:58:18 +0100