Thank you so much for
the introduction and thank you for

the invitation and the wonderful
opportunity to present my research Europe.

You'll have to excuse me for
just a minute.

Because for some reason I see that the
computer is not connected to electricity

which is a minor problem because my
slide will stop in about twenty minutes.

So if you do that.

I'm sort of.

Open here.

But if it's.

Behind.

It's my apology for

that than my apology for my red color
is being black on the overhead for

some reason there is there is no
synchronization was your color isn't mine.

So Red is a little black but

I hope it will be OK thank you again for
do it this wonderful invitation.

The title of my thought because
a little complicated database and

improve adventures will become
clear as I go along and

then I try to squeeze
science engineering and

management in a coherent fashion into
a single sentence and I hope I manage.

But really what I mean by that I hope
will become a little clearer later on

I'll talk about research that I've been
involved with for over a decade now.

And have been many people who have
been my partners doing just research.

First and foremost many students those

stars are those happy on some already
graduated the others are still there.

I've had their collaborators doing
a brutal analysis that this takes

many people with Wharton working group.

Theoretical partners and of course
industry partners because I need the data

and then there is the Technion C center
which was mentioned in the introduction

and I show you more provide more
details about it as we go along.

A little history.

My personal history was already described
in mathematics computer science that

is the separations right.

Research.

Then my first job was in the business
school and currently I'm in an I.E.

faculty which I could think of one
could think about one could think of as

being the convex hull of all
those interesting station is or

that I've visited all along and
that's actually useful to recall because

it really affects the way I look at
things through all these lenses together.

I've been teaching of course type of
service engineering at the Technion for

those who fifteen years
now all of the material

that I'm talking about appears
either in the course or

in one menu of the course and
your modem will come to look through it.

It's all in English.

It's old publicly accessible.

There is research on call centers it
started Technion about ten years ago and

we became more serious of Alltel
of research about five years ago.

Just research came from I.B.M.
in a joint research program that

goes under the heading O.C.R. which
stands for open collaborative research

that's a program that I.B.M.
has started in order to

encourage partnerships between university
and I could email was no strings attached.

That's kind of unusual.

We have been enjoying to support for
three years.

And again the tech.

Which I shall return to momentarily few
words about service engineering and

service science this is not a common
term definitely not a common discipline.

But it is becoming so

it's now being recognized or as I say here
there are emerging them of disciplines.

There are no universities
worldwide that teach courses and

sometimes programs in service
engineering and service science.

You just browse through and
you're just here in S.S.M.E.

in Google and
you'll come to a website of I.B.M.

as it seems to me stands for service
science management and engineering and.

Well I.B.M.
has their story about why it is important

to develop service sciences
service management nowadays.

Programme one service science
it's goes under the we're

going to see service
enterprise engineering.

Germany has programs in
service engineering etc etc

My take of it is that I try to
understand really fundamental phenomena

that are unique to services and
which are common across up occasions and

I shall explain this in
more detail as we go along.

My main fields about the cases
of call centers and hospitals.

These will be my main sources for today's
lecture and as was mentioned I try to

simplify complex realities in a way
that is the meaning both for and

of the research and
I would just want to make a comment that

the fact that the ultimate results or
Delta results are simple doesn't

necessarily mean that the models
are simple the models could go deep into

mathematics it's just the outcome of
the model is simple and actually a couple.

So it could be exported either
through the classroom or

through with management perhaps it's.

Obvious to you with the time frame that
I have it's impossible for me to go into

the details of the whole to do things so I
just will explain to you what can be done.

And hopefully this will just make you
curious about how things can be done.

A little about the title.

So the classical stuff scientific
proof which starts was measurement.

Goes on was developing a model
then experiment things the model

validating the model refining
it if necessary and vice versa.

This is being day to day reality for
physics biology and many other sciences.

In recent years because
of the human complexity.

The need for such a pro dynamo sort of
arose in transportation in the comic so

we have behavioral economics and
behavior transportation but for

investment in years and operations
researchers which is where I'm coming from

this paradigm is of course
not very natural and

what I'm trying to do and
what I've found is that it is useful.

So this is one way of depicting those
three wards science engineering and

management in a way that I think
the interaction among them will become.

So everything starts was data was
measurement and then as I say you just

build a model based on this data you
validate the model and you keep cycling

around here which is to say class like
I said it's classical sense if it.

Now the markets do each a certain
point a level of maturity

that makes them amenable for application
and this here comes the engineer and

the junior takes those mature battle
models and transfers the model and

the knowledge to the managers for
applications the managers implement them

improve their system get feedback
on that and they cycle around here

until something needs arise which cannot
be accommodated by existing models and

here comes the engineer again and
transfers those needs to decide this.

The scientists measures again and
he wrote the cycling around here.

So you have here a whole cycle
between science and union manage my.

We professors are focusing on
this art which we develop we do

science we teach engineers and the hope is
that they transferred the knowledge that

we developed to actual practice and
it helps me think about my role and

the role of students and the role of
your search in this whole big universe.

So I hope this gives some idea
of what I meant by engineering

versus management person is science.

OK the first for requisite for
everything that I do

is data as you saw before I start
everything with data and the data it's

reusing is quite unique in the sense
that it's not really aggregated data.

So what I'm saying here.

I need data at the level
of individual transaction.

For each service transaction be.

A phone service in a call center or
pictures visiting the hospital or

browsing in an internet.

I would like to have the whole operation
will history of this transaction

from the second it enters
the system until it leaves.

OK so I call that timestamps of events.

Now the sources for the data as I said
are mainly telephone call centers and

hospitals but of course we have data from
other sources are the service operations

face to face services the ministry of
data for example the port system and

we're expanding.

So we're now trying to

trying to convince hospitals to
install are not afraid the system so

that we get I don't know if all of you
know but are afraid the stands for.

Read your frequency identifiers and

think of it as that chip that
broadcasts location in real time.

OK so
we're trying to install our if sensors.

For example in hospitals so that in real
time we shall know where every patient

is and hopefully also where every
staff member is and perhaps everywhere

every bed is in order to be able
to trek motion within the hospital

in the resolution level that is similar
to what we do actually in poll centers.

I'll give you a small example over an hour
if I get the kitchen in a minute to

demonstrate the difficulties of
actually doing database research.

But before that I want to pause for
a commercial These allow me and

the commercial is about
the scene at the So

this is the logo of the tech and
there is a cog wheel around here.

And as you can see we change.

It's called wheel to a telephone.

To express the fact that this is what we
do in fact it's four years old because

this is how old the lab is and by now this
logo is already up Salit people kind of

asking What is this
because there is no OK so

I have to change the logo because of that
and I also have to add to it something

that represents health care if you have
regional ideas I'll be happy to hear that.

Anyway this is the logo of our lab.

What we do in this lab we
create They are reports that or

is that support research and
so for example.

Well the first database that we actually
had was about ten years ago and

it was a database of a small bank

in total about three hundred fifty
thousand calls over the whole year.

But this really paved the way
this was done jointly It was

called if it worked and and and.

But having this data set already
was enough to convince the world

that it's really quite exciting and
interesting.

And this really paved the way for
collecting or

getting larger much larger databases.

So for
example the first database that we had.

We thought is from a U.S. bank which
includes about covers two hundred

to two and a half years of data one
hundred twenty million colds forty

million calls handled by agents hundred
and eighty million calls handled by

the answering machine and there is about
there are about thousand agents more or

less operating discoursing So
it's three hundred fifty thousand calls

per week as opposed to play or
year which was our first post and I.

Then we have a cell phone company
database an Israeli bank.

The same one.

But now it's group and
it's much much larger and

data from an Israeli hospital
the story with this Israeli bank

is kind of interesting and
the state that we got into is default.

And this is really a bank is archiving
door data at the call center

every night and

when they're done with your archiving they
deposit all the data in our laboratory.

So every morning I have
the data of yesterday and

it is being cleaned up the magically
accumulates and builds up and

eventually those call centers
could actually connect to

the laboratory work with my server use our
interfaces in C. all of the flaws that

happened to them yesterday in order to be
able to correct them almost immediately.

So this is as close as one can get to real
time monitoring of the data because you

cannot really monitor or archive
data in real time in a call center.

There is simply too much data and
will slow down the operation so

doing it at night is the best we can and
we just get as close to

the the best possible by having this data
fed right into our lab every morning.

Is the software that we developed in
order to access the data in a lot

of the pictures that you will see or graph
that you will see a record be created was

and I demonstrated this this
morning to some of you and

I like to refer to it as extend Vironment
for graphical exploratory data analysis.

He is an acronym that was coined
by John to grace that decision and

thirty four years ago maybe
he didn't have in mind.

Exactly.

Maybe he did have in mind
exactly what we do today.

You could not do it over and over there.

As we do it now but now actually the idea
is to be able to sit next to a computer

will ship out almost immediately
get an answer to your question.

And just keep doing as I say
exploratory data analysis.

That is available for it's free for

academic use it's available
on the server of the.

And you just register and

then can access it and
get from it as I said Wonderful pictures.

Now just to demonstrate a little of
the difficulties that are involved in

getting data at to a state where
you can actually use it so

we were actually involved
in a small it's not small.

It's small data.

Data wise it's small but
it was a drill for

all the mass casualty event at
the Rambam hospital in Haifa and

what we did is we put our faith the tags
on the forty seriously injured.

People who actually
participated in the drill and

then we tracked their entrance to
the system and the exits to the system.

We also tried to track the motion of
doctors and nurses through the system but

for some reason they didn't work and
I was very disappointed what had happened.

But anyway from this yellow line.

This was a chemical mess casualty
event from beyond the yellow line old

the casualties are all you wash your
clean we put a tag on all of them and

then we tracked him when he entered
the hospital and I'm just mentioning here.

The directory twenty people are doing
these observations manually

in order to keep track in real
time of the state of the drill.

So if an hour of effort are for the system
is found useful then it will be a very

huge benefit to monitor or
to help monitor or disable thing.

Now why am I showing you that well as I
said there are only forty patients that

were tracked.

And here is the data of those
board patients entering and

exiting the system and you see here next.

It is missing here is the entry is missing
here is exit time is the entry time here

is wrong time etc so that percent of
the opposite patients are flawed.

Now think about what would happen if you
had to clean sixty thousand calls per day

and that's basically
the difficulties that are uncommon.

So there is a lot of expertise
accumulated in actually

cleaning the data in preparing it for
teaching and research and show.

Get them but it way do it now.

OK but I'm not going to for
through it again.

What does the data.

Tell us.

OK so here's an example from
the first database to ever.

Thirst back that we ever analyzed and

what you see here is a simple
histogram of service time an important

service time the time that the agent
picks up the phone and talks to you.

It's not including waiting
time a simple histogram OK.

This is time and

this is the fraction of coal is that
actually lasted that amount of.

We had one year worth of data.

This is the data covering generally to
October and this is the data covering

November December two months and
you see that they are very very similar

in fact when you take the log
of that service time and

plot its histogram you will discover
that you get almost perfect so

durations of service calls in a call
center are logged normally distributed for

some mysterious reason enough to
return to this in a minute but

there is another difference here.

I mean there is a difference here between
the November December and the general

York Tolbert histograms in this
difference is right here around Oregon.

You see in November in January to October.

There are about seven percent of the coal
the last of the less than ten seconds.

What can you do in a ten second school
basically pick up the receiver and hang up

OK until this time I had been
interested in and customers

I'm sorry customers of course centers
abandoning when they are fed up waiting.

Now this is the first time that I
encountered the opposite phenomena.

What you see here is agents
abandoning customers.

What would cause them to
abandon the customers.

Why would they hang up ten seconds or less
than ten seconds after a day that they.

Well one reason is
a distorted incentive system

in many call centers agents are judged
by how many calls they made per day or

what is the average duration of the cult
and what is an easier way to improve your

statistics then just hanging up
on customers and other ways and

this is actually what happened here is
that the engines were simply exhausted.

They were they're overworked and

this is this was their way to
tell the system that I'm fed up.

I can't have that anymore but at any rate
you really discovered those phenomena and

that this was a very important lesson for
me.

You realize that if I plotted these
histograms in the resolution of

half the minutes.

I would not have discovered this
phenomena the natural high scale

the trek performance in a call center
is seconds and every system has its

natural time scale which you want
to really track events with OK.

The other point is that we discovered an
interesting phenomenon the law of normal

distribution which I said is not the only
and it's also quite surprising for

those in no program limit theorem
the central limit theorem there would

recognize that the log normal distribution
has a natural multiplicative structure.

It doesn't have an additive structure and

want to expect that the duration
of a phone call will have for

some reason an additive structure and
as it turns out then this was quite

surprising to me the log normal is
what we call infinitely divisible

that's something that is very unclear and
very difficult to prove mathematically but

it somehow gives some all I say I
don't want to say credibility but

if someone explains why would the normal
normal the peer as the distributions is

appropriate to describe service durations
but there is still the question is

why specifically look normal and not any
other infinitely divisible distribution.

This is left for future research.

OK now this is a broad picture of
the service rep told long does

the service take.

But as I said these are simple models for
complex realities and here is.

The complex reality.

This is what a service conversation looks
like from at the resolution of seconds.

You see you start here.

Then there is a Hello there is a request.

It's a flow chart of a phone call
at the resolution of seconds and

we create three Stipe models for
those who know the technical term.

Sometimes the fixes are exponential
as they ought to be theoretically and

sometimes they're not so
you can see the difference the mean.

And there is a standard deviation.

And when the face is exponentially
distributed the meanest indication ought

to be close to each other and sometimes
they are thirty four thirty two and

sometimes they're not forty
nine twenty four and so

it becomes an interesting challenge
to fit the face distribution.

Even when you have all those phrases.

Granulated at that level.

OK So this is a map of a phone call
this is a very complicated process.

Another point of view on this conversation
is from the agent's point of view.

So what you see here is joint
work with no advance you and

Shannon from from from North Carolina.

This this simply tracking phone
conversations over six months for

a single agent service time.

How long it took half a year.

OK And it's kind of obvious that
there is a decreasing trend here.

So services become shorter as
time goes on agents tend to learn

even that individual level
service times are still

look normal distributed as you can see
here and what you're trying to understand

here is the learning effects of agents
while they talk over the phone.

Well sometimes they behave
normally indeed they learn so

this is not the service duration but
the service.

This is that means service that right.

So it's decreasing in time.

You see that sometimes when agents
going to break this one less.

For three months and came back they
forgot to actually what they did and

then they went up again and sometimes
they did change the nature of the job and

their conversations became longer.

So here you see actually
twelve the learning curves

of agents this is no longer
to serve a duration.

It's now the service rate and as you can
see some agents increase their service

rate in time they become better but
some of them become worse a slowdown and

when you have thousand agents and you pay
them lots of money is salaries it's very

important to be able to track this
individual behavior as one goes along.

OK so the learning issue is one important.

Why should I bother was performance at
the level of an individual agent and

at the resolution of seconds because as it
happens when you have a large call center.

It turns out that if you increase
effort service duration by one second.

Suddenly you have to find
several million dollars to fund.

It when you do the mathematics you'll
discover that you have thousand salary

Speer year one second per each
cold added or added to the effort.

It becomes a lot a lot of money and
similarly if you save one second.

Then suddenly you've got free money
that you can spend on other things so

many years of call centers are going
to show associates huge importance

the duration of the call because of
these very very large cost figures and

basically what we're doing is we're doing
classical industrial engineering time and

motion studies but
the result going down to the shop floor.

We do through our data and I know lies
ing it the way we actually done so

we do work design.

We do work or design based on
what I showed you before and

we actually are now going to do
the same thing but with the I.V.

our I.V. our stands for
interactive voice response.

It turns out for example that in
banks seventy percent of the services

are carried out through the I.V.
our people don't talk to an agent but

it may sing lead there is
hardly any research to support.

To design proper design No five.

You are in the way the I.P.R.
interacts was desisted itself.

There is very very little if
a hole that can help managers or

designers actually create good systems and
so

here is his the realm of the art and
I'm not going to go into the details but

you see it looks quite clear that there
are very several peaks there in order to

fit a distribution to such an eye
view artist to such an eye duration.

You need to be a mixture of
several distributions and

we have automatic tools and no software
that allows us to fit make sure.

So in this particular case I needed about
seven log normal distributions before and

mind you I push a button and in a couple
in in ten seconds twenty seconds.

I'm getting an answer to what is the best
mixture of seven lognormal distribution is

that will fit my data.

This is a flow chart of the I.V.

are exchanged it is similar
to what you saw before.

Let me not expand on that.

Some of the pictures arise
not in hospitals as well.

So what you see here is
the length of stay in Israeli

hospital measured in days
looking in DIE as well.

It's like this from Singapore and
they look quite similar.

So look normality prevails
all over the world.

Now if you plot the same histogram but

instead of days in years
resolution of hours.

You get this picture.

So it's the same thing.

It's just now it's hours here it's days.

So you have a log maybe envelope but
there are those peaks that arise.

Every day.

What is the source of those peaks
what it turns out that they have

to do with the release policy in
the hospital because in Israel patients

are released typically at three o'clock.

Plus Minus some in Singapore
they're released at one o'clock.

Plus Minus up.

So when you plot the when you plot
the this length of stay of patients.

It was written of our.

Then this India release rate of patients
shows up and then you get old those peaks.

Spaced by by date and again one would
ask why should I bother was it.

Of course it affects this
resolution is very important.

It affects stuffing decisions it affects
bad management within the hospital and

many other operational challenges

must address must be addressed
using this resolution.

So you see this is an interesting picture
and you'll see a similar picture later on

when I get well I gave you
a little of the history and

I said that I started with
school center of this was about

ten years ago the call center is a huge
industry and I don't want to go into

the details just to impress upon
you a little the fact that.

It is argued that there are between
two to four percent of the U.S.

workforce employed in call centers.

So there are more people actually
answering phone calls that

are working in agriculture.

That's a fact and similarly in health
care are more nurses than people that

work in our culture just impress upon
you decide is of the service sector.

So these are really sectors the deserve
the attention of Engineers operations.

Researchers managers etc If you've
never been to a call center.

Well that's what it looks
like it looks like your

computer lab not very exotic it's just
average very taken from the Internet.

You'll find many pictures like that but
sometimes it can also look like this.

That's a call center.

Obviously if you see the people
because I don't know right now so

obviously in a call center like
that they don't need my help.

They need somebody else who designed
the whole working environment but

it can get better and that and that's why
sometimes these environments are cold

the sweatshops of
the twenty first century.

So not everything is bright and high tech.

Call centers.

But this is more and
more familiar picture of a call center

from our point of view as an engineer or
a service engineer.

OK.

It takes me about forty minutes

to go in details over the slide which of
course I'm not going to do so let me just.

And I'm using it as an introduction to
the course disservice engineering course.

So let me just give you
a glimpse of what you see here.

First of all there is color
coding you see the yellow is

a function that is to be carried out
at the coal said OK So forecasting for

example I want to forecast Khomeini people
will probably do more is an obvious

function to be carried out that of course
underneath the function there is in

purple the scientific discipline that
is required to support that function so

naturally statistics supporting
forecasting now whenever there is

more than one discipline that is
required to support that function.

I indicated that in blue emphasized
a multidisciplinary needs in running

a post and so for example when you want
to analyze the phenomena of abandonment.

You need to combine psychology and
sophistic.

And when you need to do what we call a
skills based voting which I should explain

to you in a minute then you need marketing
human resources Operations Research and

Information System people
talking to each other.

OK So
oftentimes you need really is into the.

Multi-disciplinary approach and

I'm using that to actually do to convince
my students that they chose the right.

Profession because the only place that
a technician and I bet the Georgia Tech

as well where all of these disciplines are
under one roof is industry junior right.

So that's good for the profession.

What do I mean by let me just
focus on two points here.

Cueing I mean what the what does
weight look like in a call center and

a function of skills based wealthy.

OK What you see here is
an actual map from a call center

that matches customer
types to server skills so

customers are over many many types
that require many types of for.

Services agents have many many skills and

skills based routing is the algorithm that
matches in real time types the skills.

That's why it's called
skills based routing OK.

Marketing is the function doesn't
the that is in charge of second getting

the customers into their types or
their worth to the organization

Human Resource Management is in charge
of segregate in the order determining

the skills that are required in order
to manage all this customer base and

this really determines various
functions such as who you do.

Who do you actually hire how
long you train them etc So

the level of a specialist.

It's quite important here.

And so this is human resource management
your patients researchers they're in

charge of developing the algorithms
that match that the customer types

with the server pools and the information
systems are in charge of developing

the infrastructure that is required
in order to get this very big system.

Board and all these as I said before
must talk to each other in order for

ultimately a call center to work and
that's I think an example

that shows you that just you require
this market discipline or education.

You know or to be able to do to manage or
engineer successfully such an operation.

What do we do I mean when we see something
like this obviously it's very very

complex.

So how do we actually do research that
supports US based quality So one approach

is to try and identify a simple typologies
was isn't this very complex structure.

So you see this is a V. model.

This is an X. model.

This is a reversed the model.

So we identify special apologies and
then we develop models

that are tailored to the public another
way of simplifying reality is to

identify characteristics that somehow
help one success simplify thing.

So for example when services turned out to
be the last dependent last dependent means

that it doesn't really matter who you are
served by the service the service is that

the determined by the plus of the
customers OK So this the private prepaid

will have the same service regardless
work they will get the service from so

when you have such a phenomena discovered
and sometimes you do then it turns out

that this very complex the poet you can be
reduced from a theoretical point of view.

If the model.

If the architecture does
significantly simpler to allies and

here I'm mentioning three excellent
Ph D's that were actually

focusing on working on these
on these simplifications so

cold out that was supervised
by Jim Di going to shake out.

And if I go by ward with so
these are really ongoing.

This is really ongoing research
that people do Ph D.'s on

these are quite recent.

It's another simplification
is when the services are not.

As dependent by Paul
dependent which means that

it doesn't really matter what customer you
are if you are served by this particular

poll eight then your service
will be always of the industry.

When this is the case then the very
general policy actually reduces

mathematically to ever first vith apology
which makes Again things much simpler.

So this is another way and

for these we developed theoretical models
which are actually applicable control

whites who control these very complex
environments and so we call such phenomena

state space called apps because this
is a much simpler state this one and

this is what what is meant by
equivalent broken control problems

which of course I don't have time
to get in and that point I want

to mention is or just to show you is the
experience of waiting in a call center.

So what you see here is
three histograms of waiting.

I'm in a call center this
is not the service time but

this is the time until you get served.

OK so ever.

At a given resolution sometimes waiting
time could be almost perfectly exponential

but sometimes it could look like this and
sometimes it could look like this.

Which reminds you i hope of the hospital
picture that you saw before.

Let's go to resolution is every minute.

There is a peak.

But every day there every minute.

There is the did resolution
is very different.

Why is there a jump your
every In ten seconds.

Well this is that when
taken from the US band.

It consists of four call centers
interacting with each other and

the particle works as follows you
calling New York and if you are not and

should was in ten seconds then you
join into queue a poll that is

being served by all four
call centers together and

you stand stand away in this interview
until you are being answered and

this ten second here shows you the system
works well because hardly anyone

is waiting more than ten seconds and
why ten seconds and not immediately.

Why don't you need to be an interview
because the exchange of information

is taxing and
then the system will just crash.

If you have to update it every second.

So the ten seconds was found based on
experience to be at a level that gives you

what you actually need and why are there
jumps every sixty seconds here.

Well the protocol here is
this follows you call and

you join the queue was
a certain priority and

based on how important you are your
priority is upgraded every sixty seconds.

So there is an operating priority in the
increase in the priorities also functional

by for you on the bottom line is that
there is an increased likelihood of you

joining or getting service.

Whenever your priority is bumped up.

And then this manifest itself was
these jumps every sixty seconds.

Interestingly we didn't we didn't have
any theory that it was capable of.

Analyzing or or
covering this particular algorithms and

based in jointly was it had
been trying to understand.

I wouldn't actually conform to
the NYMEX that I just described.

OK Here are some pictures from a good
picture show the picture from whole center

here are some pictures from a hospital.

I'm not going to explain
that in detail and

an analogy to what I showed you before
that some things can not be as nice.

Well here is a picture in a hospital that
I got from a colleague when I visited

China.

That's This reminds you i
hope of the picture that

the flow chart that you saw
before of the call center.

But actually if you look at the details
it is not this is the flow of a patient

within an emergency department using
the same coding that was used before.

But as you recognize many of the functions
that existed before do exist here as well.

So we do want to forecast.

How many patients will arrive to more
in order to know how many doctors or

nurses are required.

We do have skills based routing
was in the emergency department.

But this goes on to different names or
other type of routing.

We have surprisingly enough patients
abandoning emergency departments

It even has a name.

It's called left without being seen and

there are about six seven percent of the
patients that actually leave the emergency

department before actually
being seen by a doctor and

then naturally most of them return
later on it to much of your state.

So there are many similarities between
their environment of the call center and

the environment of the emergency
department not to mention defect that both

of them serve as the gate
to the operation.

This is the first encounter
of the customer with

the operation of your mercy Department and
the call centers and

through this encounter the customer
forms their opinion about the.

Of the organization for
better or for worse.

And that's something for us to take into
account when we design such systems and

taking into account
the human human factor.

So the point that I want to show you that
I'm showing you that is to emphasize

defect that I'm not doing service
engineering of both centers and I'm not

doing service engineering of emergency
departments I'm doing service engineering

here again and the tools that I develop
are applicable to call centers and

immerse the apartments like in many other
operations as well we have been involved

in several research
projects in hospital and

I don't think I have much time
to go over the details here so

let me just focus just give
you details about two of them.

This is OK.

The analogy of skills based relative
in an emergency department.

This is flow controlled within the it

could be operationally based or it could
be clinically based What do I mean by that

clinically based means that there is
a process that is encountered by a patient

when they entered the emergency department
the first if you want to go hear

those over a particular type or year etc
etc So this is one particular architecture

architecture of an emergency appointment
but rather think can be also operation you

based as opposed to clinically based for
example a fast track model.

What is a fast track model you try to
identify those patients that we remain

in the emergency department for
short time and

you give them high priority
shortest processing time for us.

Now those who actually get
short service time could be

because they are light the heart or
they could be also severely hurt but

in both cases they will stay within
the emergency part and for short time and

you really try to identify them as fast as
possible in order to clear up the beds for

those who will come later.

So those are two different
architectures and there are you.

And you see there are two
other architectures.

So they're in short there are various ways
of being able to operate an emergency

department various flow
control policies and

what we tried interest in this project to
do is to understand which one is best and

of course that's too broad of a question
we try to understand which one is best.

Under what circumstances we try to tailor
the strategy to the circumstances and

the circumstances typically
is the patient mix.

So for example when you have an early
population to dominate your patients.

Then a fast track would be
preferable because most of them

will be hospitalized and you really want
to discover this is fastest possible and

then they get them flow into the system
and let the beds empty and be ready for

others so that's one project that
we've been working on the other one is

a student from Singapore and
a home for him kind of the Technion.

And it has to do again with flow
control within the emergency department

from a different angle and
the question is quite interesting and

a quick question is this full.

There are two doors in the emergency
department the door the entry door and

the exit door and went up been
a physician becomes free Idol.

Then he asked himself who should I see
next those close to the exit door or

those close to the to
the to the entry door.

The reason you want to give priority
to those close to the entry doors

because you don't know about them yet.

So you need to learn about them and
in order to do that reality properly and

reason you want to give priority to the
exit door is because you want to free up

the beds and just let others flow into the
emergency department and not block it and

it's an interesting question of who do
you give priority to and when and how and

in this project what we did is we
just formulated it eventually as

this is actually quoting the physician
that manages this emerging apartment

when I asked him.

What do you think just try to
articulate for me your priorities.

He said I would like first to
adhere to reality constraints.

So high point and
type two are to see if you're to even

worry about your immediate jump and
serve their lives but three four and five.

I think you need to see within half
an hour for the sake of the story.

For within an hour it
type five was in five.

It was in two hours.

So each arriving type has
a deadline associated Was it

that is based on the political priority.

And so he wanted to say I want first to
adhere to those three articles trains and

then once I do that.

Get the rest out of there as fast as
possible and this is exactly what

the the problem that we formulated we
minimized congestion costs either waiting

cost or
number of beds occupied subject to it.

Hearing to those constraints and
it turns out that in heavy traffic.

There isn't very interesting and
that very nice way.

And that's what we are for
many of these problems in solving it.

And that's something that is
not presently written up and

we followed the framework of the plumbing
person and considered mission control.

Instead of just minimizing cost but
omitted my.

But the situation is quite simple.

OK so it's a record of press perquisite
for my for everything I say was data.

No the second predicate and there
are two major prerequisites The second

purpose it is models and there
are broadly speaking two types of models.

They're fluid models which capture effort
is there based on laws of large numbers

they capture what I call
the predictable variability and

they give rise to that
I'm interested models.

For example this would be an arrival
rate which is actually derive a process

through this drill that I was

telling you about from where we
installed those are if ID And so

I'm not going to talk about your models
they're basically Dynamical Systems that.

That can be developed in order to
capture predictive abilities in

the casting system.

The other type of models which
are more more interesting for

the purpose of today's talk arced the cast
and so let me give you a short brief

introduction to the it's OK traditional
curing theory predicts that

service quality and service deficiency
must be traded off against each other.

What do I mean by take the simplest doing
model and one many of you studied it.

I believe in their courses so
it's a single serve a king who if you want

the server to be highly utilized You
have no choice but to live with the fact

that your customers will be waiting on
average ten times their effort service.

OK so if you want highly highly utilized
server what I call your the congestion

index which is the average waiting time
divided by ever service time will be

approximately five minutes Service
fifty minutes on average wait.

OK.

This must This is a reality.

You must live with and

only nine percent of the customers are
actually served immediately upon arrival.

Most are delayed and delayed for
a long time as you can see but

I say heavily loaded using systems with
congestion index zero point one So

one hundred times less than
here actually do exist and

they are prevalent and
here I give you three examples.

Well so when the congestion
index is here is point one.

This means that waiting time is one order
of magnitude less then service time.

Right.

OK so in call centers the natural way
to measure waiting time is in seconds.

The natural way to measure service
time is in minutes in transportation.

When you look for
parking in the congestion city.

So maybe you feel like
it's a long time but

you still measure naturally the time
you search for parking it minutes but

the parking time itself is
naturally measured in hours.

Hospitals.

When you wait to be hospitalized.

Then the waiting time seems to be very
long but still the natural way to measure

it is in hours while the hospitalization
time is naturally measured in days.

So these are three examples where
waiting time is one order of magnitude

less than service times which is very
different than what you see here.

Moreover in Paul centers for example about

it's not unusual that fifty percent of the
when they run well it's not unusual that

fifty percent of the customers are
actually served need to be upon arrival.

Once they go beyond the ideology
answering machine hurdle there

being search results too much weight and
that's again something that

really stands in contrast to what you see
here the nine percent fortunate costumes.

So this gives rise to a very
different dynamics that is of course

due to the fact that there
are many many servers running.

I mean serving catering to the systems and
this gives rise to what I've been calling

all what it is now being referred
to as the Q It's an operational

regime were both high quality
of service and hi Paul T.

of high efficiency of responses
is a pain simultaneously.

Now you would recognize the Q.B. also
arises and is used in mathematics right.

It's what ever demonstrandum it's the end
of every path America proof acuity is also

a quantum electrodynamics which is what
Richard Feynman got the Nobel Prize for

but for us it's quality and
efficiency driven.

OK so what I would like to explain to
you briefly is what is meant by dequeue

regime for cueing systems and for this.

I'll take the simplest system possible.

We call it a a for abandonment.

It was first used by Paul in the forty's
and it's really a birth and death process

as simple as the one to study in the most
basic of the chaotic process of course.

What do we have this model assumes an
arrival process that is posts on server.

Time to visit exponentially
distribute exhibited

patience time that is also exponentially
distributed and number of service.

So therefore parameters that characterize
this model and arrive a great service

great impatience rate abandon individual
abandon rate in number of service.

Maybe to new things for you is
the impatient so this impatience is like

an internal clock that every customer
comes Was that a call center.

If the clock rings before the customer
gets serviced the abandon if they get

service before the clock rings they
get OK that's what they do not mix of

the systems and it's easy to write
down the transition rate that

the ground is reversed impressed with
this both in the process and analyzed.

OK let's test these ere long
a primitives this something of

the Assumption is the director is a plus
on service your actions are exponential

impatience is an exponential there
are other implicit assumptions as well.

For example the primitives
are independent.

It's assumed that impatience and
service directions are independent.

It's assumes that customers and
service are what you use it is you that

the service discipline is first
come serve for first served.

And many others I would like to validate
support or refute these assumptions.

OK arrivals to a call center arrive as
the call center these are rival rates.

You can see that they are actually
time during over the day.

So if that old the arrival will not
be homogeneous boss on it will be

not homogeneous plus on
Incidentally this is an arrival

rate from a course that foreman over it.

Over seven hundred helpdesk in
the United States in one thousand nine

hundred five and

this isn't a driver rate from an English
call center nine hundred fifty nine.

So there were whole centers also
nine hundred fifty nine which again

just underscores the fact that there is
service science there to be discovered and

this is the call center to be analyzed
in one thousand nine hundred nine.

So it's far as those two
hump shape are concerned.

Not much has changed in
the world of School Center for

the last fifty or sixty years.

OK so I said if the arrival is
the process as it was in process

then it will be in one homogeneous possum
process and that's something that can be

overcome because we can assume that it's
maybe piecewise constant arrival rate and

then analyze every hour as though
it was almost units was done but

it's more interesting and more complicated
because it turns out that the rivals to

a call center are what
we call over dispersed

the over dispersed in the sense that there
is much more variability in the arrivals

then what the process on model
predicts how do I discovered that

I take one hour say Mondays between
ten to eleven and I look at all.

Regular Mondays in a call center over a
year I plough I calculated the effort and

I calculate the standard the various And
if this was applied a sample from a possum

distribution those two numbers mean and
variance should be close to each other but

it turns out the difference is way larger
than the mean which is what we call

over dispersion over dispersion
relative to the plus on hypothesis.

So we have a time you know what you
just process which is over dispersed.

So it's definitely not
your every day was a boss.

We know already.

December's times are not
exponentially distributed.

We know that patience is not exponentially

distributed what you see here is
the Huzzard rate of impatience or

predictions of lost of customers within
a call center that has a rate for

those who don't know what it is former you
think of this graph as the likelihood of

you abandoning the next second different
that you have survived waiting so far.

OK so the value about one hundred is
the likelihood that you abandon in the one

hundred first second given the fact that
you have survived hundred seconds of

waiting and what you see here is the heart
of a hazard rate of impatience in

a call center and you see the first of all
if it was exponential the being fat but

off it's not and it actually
exhibits some form of a behavior.

There are here ups and downs.

There also.
Here are two curves by the way one of them

corresponding to V.I.P. customers.

One of them corresponding to regular
customers who do you think has more

patience V.A.P. customers the regular
customers regular board patient.

OK The blue one is the regular customers
and you see that every second.

They're more likely to abandon
than the V.O.I.P. customer.

So who is more patient.

Why maybe maybe.

OK so it's a very interesting question and
it actually records that the IP customers

are more patient than
regular customers and

this raises the question of
what do we actually measure.

Maybe we don't measure patients.

Maybe we measure needs and V.A.P.

customers have typically V.A.P.
services and they need it more.

So you will be waiting there for your
transact to give your transaction order to

the bank fuming from anger but
you will still be

waiting there until you'll be answered
because you have no other choice right.

So we measure the swing the way with which
is not necessarily patience but never mind

from a statistical point of view I don't
distinguish between those two things.

Now what are those ups and downs.

So this up right at the beginning
corresponds to those who are not willing

to wait at all but
there is a bump here at sixty seconds.

And this was very consistent.

There was always a bump
in sixty seconds and

it took us about one year until I
mean I was working on that was there

are bound to Porton and
then he told me that's impossible.

The something going on in this call center
at sixty seconds that we don't understand

and then it occurred to me that maybe
after one year analyzing this data maybe I

ought to call the school center and

see what's going on because I
was not a customer of that bank.

So I went back to Israel.

I got the permission to call the call
center and listen to what's going on and

sure enough after sixty seconds.

There was an announcement and

the announcement actually
says the following year.

Number three in line in the first one
has been waiting for two minutes.

That's the nature of the announcement of
course this is not what you would like to

listen to your one you want to
know how long you'll be waiting.

But what the system is telling
you hey here's a puzzle.

You were number fair.

They're waiting two minutes guess
how long you will be waiting.

But the point of the matter is
that what this message is doing

is it reminds you that
you're actually waiting.

And then you ask what am i actually
doing here and you hang up and

this isn't complete opposite to what
the designers of the system had in mind

because they put this mess this year
in order to convince you to stay.

But this message is actually doing is
the reverse effect of just encouraging

customers psychologically subconsciously
we actually abandon the queue.

And there are many surprises like that
when you analyze data from this point of.

OK so just there is a reminder.

I put this graph just
to convince you that.

Patience is an exponential so you see
patience is an exponential it turns Also

it turns out also that service time and
impatience are not independent.

This is ongoing research we talk of in

the student at the Technion We
call from Jerusalem.

We analyzed the dependence between
service times and impatience and

it turns out that customers are willing
to wait more for long services.

OK And we discovered that and

that's very important by the way
let me just make this in a minute.

Summarizing what we found out
there over the not post on

service times are not exponential
patience is not exponential

building blocks are not independent
customers are not the one genius.

Has to be returned for service in short
none of the early long hypothesis applies.

Which of course raises the natural
question whether Iran has it all relevant

in my answer is in fact
the yes it is relevant and

it raises another challenge to
understand why is it relevant.

What makes it work so well so
I don't know how much time doing heads.

If at all.

I should finish them.

OK So it raises a very important question.

Theoretically and
practically is actually relevant and

what I'm trying to do is just to give
you a small little of the idea of why I

believe it is and proving it
using either data in practice or

just developing theory that
supports this positive answer

come from a radical point of view it all
started in the work was there Brown and

his students and colleagues at
Wharton this was published in jazz.

Some years ago.

We actually already there.

Fit in there like a model to this
complex reality of the call center.

And it already was very
very interesting and

very intriguing how come this
very simple model works so

well to describe reality but it didn't
always work sometimes it feels work and

then it raises the question of first of
all in order to understand why it works.

I want to be able to justify its
robustness and then I want also to be able

to chart the boundaries beyond
which it's not applicable and

when I chart these boundaries This gives
rise to new models new phenomena and

you theory did I would like to develop
the way I try to address this new new

new theory is doing theoretical theory
that is based on a sympathetic analysis

because when you do the same topics then
really the essential scum out of the model

doesn't thought it theory tells us what
is important and what is less important.

And this is exactly why I would like
to this is exactly what did what

is being used to identify those features

that matter as opposed to other features
that actually wash out in the limit.

So this is where double disc
you do regime extra rises and

it really helps explain why
the Earth gate is so robust and

they can make just a side comment
I was motivated by moderate.

Large service systems which has which have
hundreds or maybe thousands of servers but

it turns out the B.S. in power
theory is the extremely accurate and

it's accurate actually to systems
that consist of five ten and

fifteen servers and this is not only
interesting from a theoretical point of

view which has been addressed recently
by researchers in Holland but

it is actually very important
from a practical point of view

because by having this a synthetic
theory relevant for small systems.

It makes the conclusion of the theory
relevant to small services them's

such as health care.

So we can haul in call centers we
have hundreds of agents but we don't

have hundreds of nurses in an emergency
department one hundred of doctors.

So if the lessons learned from
the asymptotic theory are applicable

to small scale service system.

Suddenly the application spectrum
is broadening significantly and

this is extremely important.

OK what is let me just now the last
I guess the last point I want to

convey because it is explain a little
what the Q Would you regime is so

that you'll have an idea of what
I'm talking about and even so

the basic quantity which is
actually used to describe

those a synthetic regimes that I was in
to get is what we call the offered load.

Let me know the offer told by our R.
is equal.

Let's think about a stationary
system is equal to arrival rate.

Hi I'm average service time.

So for example if you have twenty
five calls per minute on average and

each call brings with it.

Four minutes of every purge of work then
you have one hundred minutes of work

arriving every minute and
then you need the proximity hundred

minutes of agents to cater to
deeds hundred minutes of work.

So we should live around this
stuff in level of hundred and

the question is do you want to be
exactly hundred more than hundred.

LESS THAN hundred if you'll
be much more than hundred

then you will be driven
to a regime that is.

Quality driven because you were you were
concerned was the quality of the customer

service quality.

If you will be below the hundred then
you're driven to a regime that is called

efficiency driven that is for the priority
here is to have the agents being efficient

and if you look around this hundred
then you'll be in the regime that is

the quality and efficiency driven regime
and now why is it the subtle regime.

So let me just are offered
load Islam the times new and

I'm the overview and
it's the number of agents.

That's to the full on mental exercise.

Suppose we keep the number
of agents fixed and

increased the offered load to infinity.

What will happen to
the probability of the late.

So we have a fixed number
of agents more and more and

more customers calling us obviously more
and more and more obviously in the limit.

Almost all of them will
be delayed in the green.

OK let's fix lambda are increased and

to infinity increase the number of service
to infinity then obviously less and less

customers will wait in the probability
of delay will converge to zero.

So you see that by having large and
relative to land or large land Oliver to

you live in the extremes of almost
all customers delayed or almost

none of the customers delayed and the
queue the regime aims right in the middle

as I told you before in well run call
centers fifty percent of the customers

actually get an answer immediately and
fifty percent are delayed in queue.

How can this happen.

This cannot happen by letting either
an increase to infinity unilaterally.

They must do it jointly and that must do
it very very delicate The relative to each

other and that's exactly to do with
the regime so there was the quality driven

regime as I said you stuff your system
by our last Delta times or there is

a deficiency driven regime where you stuff
your system by our minus gamma school.

About And there is a quality and

efficiency driven regime
that is not intuitive.

It's a mathematical story and
the story says in order to be round our

Which is where you want to live a wrong
means square with neighborhood of our.

So the square dequeue of the regime is
actually the one where the number of

agents is approximately equal
to our plus some constant

squirt of our word this constant can
be negative or positive from practice

in practice purposes the constant
beta is between minus one to plus one

this regime was already introduced by
earlier about one hundred years ago it was

mathematically substantiated by how often
and what it is about thirty years ago and

that's why it's often goes under
the name health in the health and

what regime and at the Technion We
just adopted this regime to models

that have abandonments in them and
this happened more recently and

this is actually the regime that is
more useful for call centers because

the abandonment is such an important
phenomenon in a call center and

that's a regime where actually expected
waiting time is one over squirt of N.

and is the number of servers expected
service time and if you think about

the system was hundred servers one of
us grew out of it is one of a ten and

that's why you get one order of magnitude
less waiting for out the service.

OK we understand the security
regime very very well there is

even mathematical theorems
that go where was the Q.

of the regime which I don't want
to because of time to put it.

I don't want to explain to you before me.

Does that come out of the asymptotic
analysis and let me just go back and

here was the question of
why does during a work.

I raise the question I said none of
the assumptions actually prevail and then

I said I'm using a synthetic theory to
try and identify why the model is robust.

So basically speaking we're
taking a very complex model and

the leave the notation here aside and
actually reducing this very

complex model in one way or another
to an earlier game because it's very.

A simple model.

If you saw before.

So for example when we have
over dispersed arrivals.

Then we have a natural way of
measuring how over dispersed it is and

then we associate it with be associated
with this over dispersion parameter

called it see if C. is equal to half.

This is the natural dispersion
level of a possum process but

if sees the knowledge in half
then it's over dispersed and

it turns out that over to respond to an
over dispersed arrival rate then of Robert

process then the staffing level will be
not our plus Be it a score of our But

our plus beta our to the power of C.
were C. is greater than half so

you are paying for this over dispersion
by having more agents on board so

that he can cop cope with this excessive
variability and there is a simple.

There are statistical theory supporting
that that this is really a natural model

and there is a synthetic theory supporting
that that says that this would be

the right stuff and gravel that goes
business over dispersed environment.

Gen patients.

OK patients is not exponential Well it
turns out from asymptotic theory that only

the behavior of the impatiens distribution
around the origin is the mattress and

this goes to the fact that in the Cuban
regime waiting times are short so

only the behavior of the patients
around Short's waiting

times are extremely important and this is
something that people here led by research

by Jim actually discovered and
we discover new things in Israel as well.

So we have reducing did General
the problem of general in patients with

the problem of behavior in the origin and
mind you when you look at the exponential

distribution devalue of the origin is the
parameter for the expression distribution.

It's the value of the density of the
origin is the parameter of the expansion.

So basically the rule of
thumb goes as follows.

You take all the formulas if
you develop for the four D.

a model and you replace
the parameter value of the extent.

Distribution by did the value
of the density of the origin

in all of the formulas are applicable and
its go and

there is another question of how do
you estimate the value of the origin.

There is an easy way of doing that and
I don't have time to explain it.

General Services.

This is the big world and

it's only now starting to be understand
we don't understand it get very well.

There are deep reasons why
General Services are very hard to analyze

when you have many servers around.

But from an empirical point of
view we sometimes discovered that

there is insensitivity to
the to the distribution and

you can capture a lot of the performance
that you want to get through

the exponential distribution even though
the service distribution is not normal.

But the coefficient of radiation is
not is not a huge deal in practice and

consequently that works quite well but
would you use customer service.

I told you about a mission to
phenomenal states basically lapse.

So we reduced it to mention all of the and
when we're lucky.

We have you can reduce it to
a single dimension one dimension and

then it becomes basically on allies in
an earlier time varying environment.

There was a very interesting story here.

There was a way actually.

OK Very Q.'s in general are very very
difficult creatures to analyze but the way

we cope with time varying Q It's actually
changing a little the question we don't

analyze time varying Q.'s what we ask is
the fully what would the stuffing level.

What is the required staffing
level such that this time varying

system will be behave like it is not so

we have kind of varying service levels
that respond to Prime for an gloats

such that at any given time customers we
actually call the call center will face

the saying service level and what happens
there is something that is quite amazing.

And again there is no theory
if able to explain it.

This Some think it is that the.

Any given time the system be if
though it is a steady state system.

It's a remarkable phenomena.

It goes to some Everything principle
that we don't understand but

there is a said The point is
you have a time for rank system

you stuff it by time varying
in level of servers and

then essentially what you get is a steady
state system at any given time and

disturbance a system happens to be
the order gate which you all love to know.

So again it's a RID reduction of this time
during environment are not a system and

then dependant building blocks
as I said long service times.

I mean long service Times Associated
was long impatience here.

Well that there's a long long story here.

I just want to say one point
that when you want to estimate

the offered load how much work arrives to
the course and there you must account for

those customers who abandon.

Right because they brought work with them
and they were not willing to wait for you.

So it's an interesting challenging
statistical question how

do you estimate the service time of
both of those who are not served.

And that's the question
that we actually have some.

Before and it has to do with the fact
that service times are censored I didn't

mention that before but the generate I
think this would be a good time to stop.

Thank you for your.

So.

And the gift of my favorite is my visit so
thank you very much.