noc19-mg15 Lec 26- ELECTRE

IIT Kanpur July 201
20 followers
Dec 3, 2023
ELECTRE process, Liking set and Disliking set, Outranking
Welcome back my dear friends and dear students, a very good morning, good afternoon, good
evening wherever you are and welcome once again welcome to this course of DADM which
is data analysis and decision making to under the NPTEL MOOC series.
And as you know this course is total duration is 12 weeks which is 60 lectures that is 30
hours each a class being for or lecture being for half an hour and in a week we have 5 lectures.
So now as you can see from the slide we are in the 26th lecture which is the first lecture
in the 6th week and my name is Raghunandan Sengupta from the IME department IIT Kanpur
in India.
So if you remember we are considering the concept of Electra process and I did mention
very briefly though so there are 6 different methods for Electra but we will consider the
simplest one.
I did also mention that this is a multi criteria decision making where you make comparisons
for decision based on alternatives taking 2 at a time and you rank them and is almost
similar to HP method where you give scores in the HP method but here you will basically
find out 2 different criterias one set is that how much you like a decision with respect
to the other if you take the decision which you like like means consider you have decision
I and decision J I am just naming them as I and J or alternatives which are or the decisions
which are there based on criterias and if you take I with respect to J you get a positive
benefit and you basically put them in the set known as the concordance set or positive
ranking and if you take the decision which is not beneficial you basically formulate
the discordance set and then also you have a negative out ranking and combining them
you basically find out which decision basically best suits you considering the overall set
of criteria which you have.
So with this I will continue the discussion and if you remember also I mentioned that
we will consider the out ranking whether in the positive sense or negative sense be
to be of equal magnitude in the sense that if I win 100 rupees or if I lose 100 rupees
my level of liking and disliking would be of the same magnitude this is something to
do with in a way the squared error loss function and there are some reasons for that because
if you remember I did mentioned beginning of the this DADM 2 lecture series that having
a squared error loss is akin to considering the returns to be quadratic the utility function
to be quadratic and returns to be normally distributed such that there is a one-to-one
correspondence there and we will try to follow that and slowly relax and take the assumption
that asymmetric loss function where unequal penalization of out ranking in the positive
sense and the negative sense can be considered and you can model that also accordingly.
So this slide we have this was the ending slide for the last class which was the 25th
class but still I will continue discussing then. So thus intuitively one start with a
set A where set A has a1, a2, 1, 2, 3, 4 are the suffixes so a1, a2, a3 till an where the
number small n is basically a number of alternatives or decisions which you have so thus each will
basically have a1, a2, small a1, small a2, small a3 corresponding to the fact that what
is the weightage is you are going to give to those alternatives based on the criteria.
So let me continue that such that each a i that is i is equal to 1, 2, 3, 4 till n has
l number of set of criteria. So obviously a1 would be divided as so if you have a1 from
this so a1 technically would be a11 till a1l for all the l number criteria. Similarly ai
would be ai1 till ai capital L and you can basically formulate accordingly such that
each ai, i is equal to 1 to n has small l, l basically nomenclature being from 1 to capital
L number of criteria and one accomplishes one means the decision maker accomplishes
his or her ranking based on the collective the cumulative effect of all these different
alternatives of criteria on the ith alternative.
So we will combine them try to basically find out what is the collective effect and basically
take the decision. So if I continue reading it it says that one accomplishes his or her
ranking based on the collective cumulative effect of this L criteria when comparing any
two different alternatives say for example ai and aj, i not equal to j and i and j being
equal to 1 to n because if i and j are same that means you are comparing two alternatives
which are the same with respect to the criteria so obviously they would not make much sense.
So it is almost equal to the concept that you are taking the principal diagonal for
the HP where the numbers you have assigned is 1 that means it is difficult to take decision
i with respect to i if i and j are the same.
And the end of the comparison process we end up with the best choice set which is capital
A1 where capital A1 would basically be a subset of A that bold A would be the set of
all the A's which you consider so we will basically rank them according to order them
like A3 can be better than A4, A4 can be better than A10, A10 can be better than A1 so based
on the collective decision of the criteria.
One advantage of the electron process on the method is the fact that the final result which
is A1 when presented to the decision maker it is easy for him or her to take the final
view rather than what he should he or she would have taken with A only.
So you basically give him or her the choice set based on the criteria such that is easier
for him or her to take the decision.
So as in that process it in case it is difficult for the person to judge the n number of alternatives
in a rational method considering n can be large.
If you remember in the AHP method I did mention that trying to basically find out the ranking
process considering n is very large and you are doing going to do n C2 combinations may
be confusing because if you are trying to compare on the same criteria for all the alternatives
time and again like say for example style and you have basically 20 cars to compare
20 C2.
So style when you are going to compare obviously this consist of consistency will be violated
because then it becomes very difficult to make a rational choice.
So you want to basically choose alternatives in a rational manner so as to choose the so
called best set of elements amongst the n number of originally set which is already
there.
Now how does the process work?
In case if the ith alternative now I am considering the alternatives not in the criteria.
So if the ith alternative AI out ranks the jth alternative AJ then we say that AI is
better than AJ.
Now the question may come that if there is only one criteria then it is very easy I only
consider one criteria find out AI is better than AJ and my job is finished.
What if say for example if there are two criteria and I am considering the alternatives I and
AJ with respect to the first criteria AI is better than AJ and with respect to the
second criteria AG is better than AI.
So how do we do that I will come to that later.
So it means that the risk or loss whatever one wants to say risk or loss in the monetary
sense will say that for AI is not as much as that AJ.
So whenever we are saying that AI is better than AJ we are thinking from the loss perspective
that the loss of AI is just equal to AJ or is less.
So obviously we like to take AI as the best alternative with respect to AJ.
So we will say that and if I continue reading it means that in and it means that the risk
or loss whatever we want to say for AI is not much as much as AJ or AI is as good as
AJ and not as worse as AJ.
So whenever we are comparing if say for example there are two alternatives.
So say for example A1 and A2 and the criteria are two in number the first and the second
consider on with respect to the first A1 is better or just equal to and in the case of
the second criteria A1 is just equal to A2.
So even if if one of the criteria on based on the one of the criteria A1 is better.
So we will definitely take A1 and proceed and that question that if they are conflicting
criteria how we do that.
Now how one does one decides the so-called relative ranking between the alternatives
AI and AJ is a matter of prime importance to what and we will basically consider that
accordingly.
It is also worthwhile to mention that here is where the collective accumulative effect
of L number of criteria come into play such that we are able to consider all of them at
the same time or individually and find out the collective total effect of all the L number
of criteria's when we are trying to consider the alternative 1 and 2 or A I and J.
So electroranking system is non-transitive that means if say for example AI is better
than AJ and AL is better than AK we may not imply that say for example we should say it
is there is one error this should be changed.
So if AI so what we actually mean is if A1 is greater or better not greater better than
A2 1 and 2 are just arbitrarily A2 is better than A3 it does not imply that A1 is better
than A3 you have to basically consider that when considering the method of electroprocess
I will change that accordingly.
So obviously it means that 1 is greater than 2, 2 is better than 3 it does not obviously
always mean that 1 is better than 3.
In electra one uses the concept of concordance and discordance set or concept which is a
sort of level of liking and disliking concordance means I like so hence I give positive weights
or positive points discordance set is basically I dislike I means the decision maker dislikes
so hence he gives a level of points which gives the level of discordance or disliking.
So which one has to use this concordance and discordance set has to one has to use based
on the criterias between any two alternatives.
So it does not mean that if I give a concordance score between two alternatives it does not
mean that I am not going to assign a discordance score.
So if you remember I did mention that if I A1 and A2 are there alternatives and I like
A1 or the decision maker likes A1 so in the case A1 is chosen then obviously there is
a positive point but in case say for example due to some situation you are forced to choose
A2 then in that case you will basically give a level of disliking such that you basically
make a very rational decision that liking and disliking would be given as positive negative
points so that collectively we are able to balance that and take a rational decision.
So if we say that we like AI with respect to AJ when one chooses AI in place of AJ then
one assigns some score or points to the quantify this level of liking and this is known as
the concordance set or level.
On the same basis as I said if due to some reason one is forced to choose AI with respect
to AJ it is just the reverse not to choose AI with respect to AJ then the level of disliking
is objectively stated using the concept of discordance.
Now consider this there may be some liking and disliking and this liking comes out more
prominent so hence it will become a concordance set and if the disliking comes out to be more
important then it will be a discordance set and we will try to take a collective decision
based on both of them.
To deal with the outranking relationship by using pairwise comparison that means I am
trying to analyze AI with AJ among alternatives so such that under each of the criteria we
are able to do it separately.
So say for example for the L is equal to 1 we rank them A1 and A2 for L is equal to 2
with their rank them between A1 and A2 and continue doing that such that we take a collective
score based on the concordance and the discordant concept.
The outranking between two alternatives AI and AJ denoted by AI is better than AJ which
generally means even if the ith alternative is not dominating the jth one yet the decision
maker can choose the ith one because in that case the effect which the person has on trying
to choose AI is better than AJ.
So say for example if I am considering that the AI and AJs are ranked in such a way for
the criteria I am not getting any benefit for choosing any of the alternatives the first
and second based on the fact say for example I am considering the criteria.
But still I may be interested to take AI with respect to AJ in spite of the fact that
the alternatives which are there or several criteria which are there are not helping but
yet I will basically assign some concordance set or concordant points such that in the
long run when I take a decision of AI with respect to AJ I will get a positive benefit.
Alternatives are dominated if there is another alternative which excels them in one or more
criteria and equals in the remaining criteria.
So say for example I have AI and AJ these are the alternatives and L is equal to say
for example 1, 2, 3.
Now consider with respect to L is equal to 1 AI is greater than AJ with respect to 2
AI is equally disposed with respect with AJ and respect to 3 AI is equally disposed
with respect to AJ then obviously there is one concordance point hence we will be interested
to take AI because for this L is equal to 2 and 3 they are equally disposed in the sense
the alternatives.
We consider pair wise comparison of alternatives under each criteria by using some matching
index like monetary value and we will give that function as GI AJ that means for each
of this criteria we take the comparison of the alternatives compare them and assign scores
accordingly positive or negative whatever it is.
So once you specify the threshold so what is the difference of the threshold?
So if I take a functional mapping or the difference of the scores between AI and AJ for the criteria
1, criteria 2, criteria 3 so obviously if the difference is almost negligible so then
in that case I am indifferent in case if they are not so obviously I will be tempted to
take the decision 1 with respect to 2 that means alternative I with respect to J.
Hence the threshold which is given by the functional form of GI AJ minus GI AK it will
depend on the decision makers choice the concept of utility function would be very
important here and we will consider that later on.
The decision maker assigns weights or important factors to the criteria and that is done in
order to express one preference to show the relative importance of one alternative with
respect to the other and take decision accordingly.
This also comes from the fact of outranking liking and disliking comes from the fact that
the utility function or net value concept of any decision maker or making process is
important driving force or factor for the choices and their ranking for any particular
decision maker would also always hold because the decision maker is trying to utilize the
concept of ranking of liking and disliking in such a sense that he or she will give some
positive and negative points to this outranking method of liking and disliking and then take
the decision.
This method is convenient where there are when there are few criteria and a large number
of alternatives.
So, trying to compare the alternatives with few criterias this will definitely be advised
because you remember same thing was said about of the HP because if the number of different
comparison increases then there is a chance that the rationality or critical index may
be violated.
So, we will basically consider this algorithm in a very simple step like step 1, 2, 3, 4
and then also solve a very simple theoretical problem.
So, what for the step 1 is normalize the decision matrix and what we do.
So, we have a set x.
So, x is basically the scores which have been assigned with respect to some boundary value.
So, we will convert the entries in the decision matrix called the scaled normalized values
which has no dimension.
So, thus we will consider that the matrix A which we have where the cells are A 1, if
you consider along the rows it is A 1 1, A 1 2, A 1 3, A 1 4 so on and so forth.
So, we will normalize them considering either through the row or through the column.
This is the important fact which I had mentioned time and again in HP that you normalize using
the utility function 0.1 and you also ensure the normalization can be done along the row
or along the column, but remember the fact after normalization the sum should be equal
to 1.
So, we have basically have the relative ranking system in such a way that it is better for
you to compare.
So, here we take x ij as the ratio of a ij divided by the square root of the sum of the
squares of the a i's, a ij's corresponding to the sum for the row or the column.
So, please remember the normalization depends on the utility function.
So, thus the square concept which you have used has something to do with the utility
function which is quadratic which means that the returns which you are going to take is
normal.
So, this is just a simplistic one it can be logarithmic, it can be power utility function,
it can be Hara function whatever you want to take, but remember that as I had mentioned
that in AHP once you basically decide on the utility function for the decision maker
do not change the utility function for the decision maker when you are trying to compare
different type of alternatives or different type of criteria based for the same alternatives.
Thus when the normalized vector is obtained from A. So, you have x as the matrix it can
be a vector depending on how many alternatives and this criteria if you have.
So, x is basically the matrix which has the cell values we consider the topmost row it
will be x11, x12 so on and so forth till x1n and the last row would be xm1 till xmn and
remember what is m and n are basically the numbers of decisions on alternatives and the
number of criteria.
Here m as I just mentioned here m which is the third bullet point it mentions here m
is the number of alternatives shown along the rows and n is the number of criteria which
is basically shown along the columns.
The normalization which you do or which the decision maker does implicitly depends on
the utility or what he or she or you think each of the individual criteria is given giving
or contributing to each alternative on an individual basis.
So once you normalize the scores which you get say for example the score for one of the
criteria for one of the alternatives is 0.12 which means the total weightages or the overall
effect of that criteria on that alternative when you are trying to compare the alternatives
would be of 12 percent worth on a scale of 100.
Now obviously the question may arise that what if the weightages of this criteria changes
for a part per alternative so like say for example for the j ith one I give a weight
of 12 percent for the alternatives which you are going to consider.
So it will be say for example a1, a2 are the alternatives and l is the criteria so say
for example the normalized base is 12 percent.
So we will continue considering 12 percent for a2 also but if the question arises what
if we take 25 percent here and 12 percent here that concept we are not going to consider
in our problem because we will consider there is a changing weightages you are going to
assign to the criterias but we are not going to consider that.
So the last point mentions the normalization which you do implicitly depends on the utility
or what you think each individual criteria is given or contributing to each alternative
on an individual basis.
So consider this so I have the matrix A which is the so called non-normalized matrix the
points which are given without any units as per the figure with the values which are giving
along the rows is 2, 1, 2, 3, 4, 1, 1, 3, 2 here m is 3 n is 3 it can be m and n can
be different also but I am taking a simplistic problem.
So once you normalize using the quadratic concept x becomes the first cell is the values
which I will take and if I am doing it along the column so it will be 2 so this one which
is coming and this will be the sum of the squares of these.
So square root of 2 square plus 3 square plus 1 square so this value which you see is here.
Now say for example when I do it for the second one which is 3 so 3 and in that case you have
2 square plus 3 square plus 1 square this value is here similarly so this one was considered
this one was considered let I consider this so it becomes 1 divided by 2 square plus 1
square plus 2 square the value is here this is the one this is the one I am just marking
it with the color scheme.
So it is easier for you to understand where how the calculations have been done so you
can find it accordingly.
So you normalize the decision matrix now we multiply each of the column on the previous
matrix x which you have observed by the associated weights you are going to assign and the level
of importance you are going to assign to the decision.
So consider the weights are given as w1 for the first and till wn for the second for the
last one and here if you remember I mentioned just few minutes back that if I consider the
alternatives and we are also going to consider the fact which is a simplistic assumption
that the weights which are there for the alternatives for any of the criteria would be of equal
weightages.
So if it is 12 percent it remains 12 percent if it is 25 percent it remains 25 percent
we are not going to change that and that is why the weights are there.
So w1 remains at w1 for all things remember one thing the first row the first cell value
is w1 others are 0 that is why the weights are same for the second row the second cell
which is 2, 2, 2 is w2 rest values are 0 for the third row the third cell which is 3, 3
is w3 rest are 0.
So that is why I kept repeating it time and again the weights remain as it is and it should
also we remember that you have normalized so the sum of the weight should be 1.
So once you multiply the normalized x with the weights you will get the total weighted
sum of the alternatives and the criteria based on which you will proceed for the next step.
So with this I will close this first lecture for the 6th week which is the 26th lecture
and I will continue trying to solve the problem with all the steps which we will consider
very simple algorithm and try to also analyze that what if the ranking positive and negative
concordance and discordance are not of equal importance then how you will do we will consider
that later.
First let us consider they are of equal importance and positive weights and negative weights
would be same weights I am using in the very simplistic sense not from the mathematical
point of view.
So with this I will end this class and thank you very much for your attention have a nice
day.
