WEBVTT

00:00:09.746 --> 00:00:13.009
Okay, decisions in the age
of digitalization.

00:00:13.619 --> 00:00:16.957
To make one thing very clear
at the very beginning:

00:00:16.957 --> 00:00:21.018
Most important human decisions
can and will never be automated.

00:00:21.018 --> 00:00:24.876
They are done by gut feeling,
under extreme uncertainty.

00:00:25.146 --> 00:00:27.957
I'm not going to talk about these.

00:00:28.557 --> 00:00:31.171
Don't blame me for that.

00:00:31.451 --> 00:00:35.549
0% automation on decisions like,

00:00:35.549 --> 00:00:39.118
these are examples from my life:
marrying my wife,

00:00:39.118 --> 00:00:42.132
declining a tax-free,
permanent CERN job offer,

00:00:42.132 --> 00:00:45.122
instead of a KIT professorship,

00:00:45.125 --> 00:00:46.843
or founding Blue Yonder.

00:00:46.843 --> 00:00:49.881
In retrospect, these were
all very good decisions,

00:00:49.881 --> 00:00:52.169
but they weren't clear beforehand.

00:00:52.169 --> 00:00:54.515
So I'm not talking about these things.

00:00:54.785 --> 00:01:01.057
But I claim that 99% of all operational
decisions in enterprises

00:01:01.060 --> 00:01:02.813
can be automated.

00:01:02.822 --> 00:01:04.729
And they will be automated.

00:01:05.055 --> 00:01:09.264
They will also be simultaneously improved
by quite some margin.

00:01:10.566 --> 00:01:13.273
That is what I am going to report on.

00:01:13.605 --> 00:01:18.474
And these are especially repeated,
regularly repeated, similar decisions.

00:01:19.174 --> 00:01:22.095
Let's take an example from retail.

00:01:22.975 --> 00:01:26.222
Assume you are a store manager
and you have to decide:

00:01:26.222 --> 00:01:28.816
Of these articles,
how many do I have to order

00:01:28.823 --> 00:01:30.804
so that I have enough tomorrow,

00:01:30.814 --> 00:01:34.525
but not so many
that I have to throw them away

00:01:34.532 --> 00:01:36.050
once the shelf life is passed?

00:01:36.063 --> 00:01:40.055
Ten? Zero? 100?

00:01:40.055 --> 00:01:44.549
That's a decision many people
have to make every day, again and again.

00:01:44.549 --> 00:01:46.214
Or, again, you are store manager:

00:01:46.214 --> 00:01:49.954
Do you want to decrease
the price of this product today? Yes?

00:01:50.214 --> 00:01:53.511
By 5%? By 10%? By 20%?

00:01:53.511 --> 00:01:56.787
That's a decision that people
have to make very often.

00:01:56.787 --> 00:02:01.005
Or to send this expensive catalog 
to that customer or to that customer?

00:02:01.005 --> 00:02:03.672
Is it worthwhile? Will it be good for him?

00:02:03.672 --> 00:02:06.086
Will it be good for us?

00:02:06.086 --> 00:02:08.676
These are questions I'm talking about.

00:02:09.727 --> 00:02:12.740
Don't you believe the 99%?

00:02:13.850 --> 00:02:19.509
First, you have to know how most
common decisions are taken in real life.

00:02:19.590 --> 00:02:21.390
And that's about something like that,

00:02:21.390 --> 00:02:28.366
so 90% of these decisions are:
Do nothing, or do what we always do.

00:02:28.518 --> 00:02:30.039
What people always have done,

00:02:30.044 --> 00:02:32.050
even before I came here

00:02:32.053 --> 00:02:38.895
and wanted to bring into this business
everything I have learned at university.

00:02:38.895 --> 00:02:43.296
First, you are told, "No, we do it
this way - ever, always."

00:02:44.216 --> 00:02:48.534
About 9% apply business rules
in one way or another.

00:02:48.534 --> 00:02:52.187
But many of these business rules,
if you really look at them,

00:02:52.187 --> 00:02:53.345
are not very good.

00:02:53.345 --> 00:02:58.367
Almost none of them
has a real proof of value.

00:02:59.147 --> 00:03:03.119
[For] only about 1% of decisions,

00:03:04.149 --> 00:03:08.428
[does] somebody sit down,
use his brain, and think about it.

00:03:09.338 --> 00:03:13.196
And even this 1% is far from optimal.

00:03:13.196 --> 00:03:14.664
For this, you have to know

00:03:14.664 --> 00:03:17.650
that the human decision-making
system has two systems.

00:03:18.420 --> 00:03:21.630
The so-called system one
is fast and intuitive,

00:03:21.630 --> 00:03:23.820
but it has many biases.

00:03:25.100 --> 00:03:27.858
System number two is slow and rational,

00:03:27.858 --> 00:03:29.379
but it is hardly used.

00:03:29.379 --> 00:03:33.281
Why? Because it's work, it takes
a lot of energy to use it.

00:03:33.661 --> 00:03:39.130
Most of the time we don't use it,
we use system one rules for everyday life.

00:03:39.730 --> 00:03:46.329
And even, and especially, scientists
and experts use their "gut feeling."

00:03:46.594 --> 00:03:50.916
They are not always thinking
and deciding rationally.

00:03:52.067 --> 00:03:53.729
And it's very important to know,

00:03:53.732 --> 00:03:56.973
system one, the fast one,
cannot speak statistics.

00:03:57.546 --> 00:04:04.245
It cannot really judge risk and chance.

00:04:04.695 --> 00:04:06.437
That's a big problem.

00:04:07.217 --> 00:04:09.572
Daniel Kahneman, Nobel Prize winner,

00:04:11.512 --> 00:04:15.953
has done research his whole life
about how we think.

00:04:17.333 --> 00:04:21.454
His book, "Thinking Fast and Slow"
is really worth reading.

00:04:24.094 --> 00:04:26.216
He describes all of this.

00:04:26.216 --> 00:04:30.094
He describes how our system one
makes decisions,

00:04:30.094 --> 00:04:32.995
and it is all explained by evolution.

00:04:32.995 --> 00:04:37.333
So it was right: we are here now
because we did it that way.

00:04:37.513 --> 00:04:42.984
But there are many, many effects,
for example, the IKEA effect:

00:04:42.997 --> 00:04:45.453
We value something higher

00:04:45.462 --> 00:04:49.199
when we have worked
for it ourselves a bit.

00:04:49.216 --> 00:04:53.244
There are many
so called "cognitive biases."

00:04:53.244 --> 00:04:56.409
And it's very funny to see

00:04:56.425 --> 00:04:59.903
that somebody, like myself,
who thinks he is very rational,

00:04:59.916 --> 00:05:03.198
has all these biases every day.

00:05:03.578 --> 00:05:07.330
So, it's like that,
we can't do anything against it.

00:05:07.330 --> 00:05:08.816
So, our question was:

00:05:08.816 --> 00:05:12.484
How can we make more rational decisions
than we are actually doing?

00:05:12.484 --> 00:05:16.124
The origin of this idea
comes from CERN and KIT,

00:05:16.124 --> 00:05:19.024
so from where I do my research work.

00:05:19.444 --> 00:05:22.484
The development is now done
at Blue Yonder.

00:05:22.484 --> 00:05:26.536
The basis is really big data,
Bayesian statistics, machine learning,

00:05:26.536 --> 00:05:30.085
data science, stochastic programming,
causality reconstruction.

00:05:30.085 --> 00:05:32.285
In other words, it's really science,

00:05:32.285 --> 00:05:36.724
how scientists work
in order to get insights from data,

00:05:37.134 --> 00:05:43.479
and then optimize decisions on the basis
of these insights for the future.

00:05:44.009 --> 00:05:46.896
It's a purely scientific
thinking and ethos,

00:05:46.896 --> 00:05:50.785
that is also not true
in most enterprises, as I have learned.

00:05:51.465 --> 00:05:53.657
We are trying to do
this technology transfer

00:05:53.657 --> 00:05:57.205
to reinvent business processes
with scientific methods.

00:05:57.585 --> 00:05:58.727
This is a key.

00:05:58.727 --> 00:06:01.836
The key to better decisions
in the digitalization era is:

00:06:01.836 --> 00:06:04.146
data and algorithms.

00:06:04.406 --> 00:06:08.119
So, first, you need
relevant objective data.

00:06:08.119 --> 00:06:11.111
So, it's mostly your data
or your company's data.

00:06:11.301 --> 00:06:15.226
And you need good, 
mathematically correct predictions

00:06:15.226 --> 00:06:20.156
for near future events on the basis
of this data, including uncertainties.

00:06:20.156 --> 00:06:21.966
That's a very important point.

00:06:21.966 --> 00:06:26.286
And this is handled by an area
that is called "predictive analytics."

00:06:26.756 --> 00:06:31.417
Then you also want to know, if you decide
this or that, what does it mean?

00:06:31.417 --> 00:06:34.447
So you need cost functions,
utility functions:

00:06:34.447 --> 00:06:37.380
What is the consequence
of this or that decision?

00:06:38.900 --> 00:06:42.320
And then you want
to optimize this decision

00:06:43.040 --> 00:06:46.787
using the prediction
and the utility function

00:06:46.787 --> 00:06:48.878
to give a recipe for what you should do.

00:06:48.878 --> 00:06:51.172
That is called "prescriptive analytics."

00:06:51.172 --> 00:06:53.537
And finally, the King's discipline:

00:06:53.537 --> 00:06:56.999
If you [have got this] far,
then you can automate it,

00:06:56.999 --> 00:07:01.439
then it's completely automated.

00:07:01.439 --> 00:07:03.968
When can we use predictive analytics?

00:07:05.878 --> 00:07:10.170
In between purely random
processes, like the lottery,

00:07:10.800 --> 00:07:15.798
and completely deterministic processes,
like a mechanical pendulum.

00:07:16.078 --> 00:07:20.011
Most processes that we are interested in
are in the continuum between those.

00:07:20.021 --> 00:07:21.623
So they are partly predictable,

00:07:21.645 --> 00:07:24.542
but there's a lot of randomness
and uncertainty inside.

00:07:24.542 --> 00:07:28.652
Most people think that the world
is very deterministic. That's not true.

00:07:28.816 --> 00:07:31.223
There is some determinism inside,

00:07:31.230 --> 00:07:34.300
and it's our work
to get this component out.

00:07:34.329 --> 00:07:38.073
But there is always lots
of uncertainty inside,

00:07:38.090 --> 00:07:42.683
and in evaluating the value of a decision,

00:07:42.683 --> 00:07:45.122
we also need to take
into account the uncertainty.

00:07:45.122 --> 00:07:46.717
That's a very important point.

00:07:46.717 --> 00:07:49.716
So, that is what
predictive analytics can do.

00:07:50.036 --> 00:07:54.021
Going back to our store manager
and his problem:

00:07:54.021 --> 00:07:56.831
How many articles to order?

00:07:56.831 --> 00:07:59.161
There may be many influencing factors

00:07:59.161 --> 00:08:02.532
on how many articles
will be sold tomorrow,

00:08:02.532 --> 00:08:05.741
from weather, weather forecast,
price, promotion,

00:08:06.811 --> 00:08:10.632
day of the week, holidays,
seasonality, and so on.

00:08:10.694 --> 00:08:13.791
But the store manager has
to make his decision based on,

00:08:13.812 --> 00:08:17.483
or take his information from,
many individual events:

00:08:17.493 --> 00:08:21.815
namely many articles, many locations,
many points in time.

00:08:22.325 --> 00:08:26.655
And all together, there are many
such decisions to take.

00:08:26.949 --> 00:08:28.830
The output of predictive analytics

00:08:28.848 --> 00:08:32.265
should be something
which is not only a value -

00:08:32.279 --> 00:08:34.611
it will never be 80 pieces,

00:08:34.625 --> 00:08:37.679
because we don't know whether
80 pieces will be sold tomorrow -

00:08:37.696 --> 00:08:42.092
but a distribution
of all possible futures:

00:08:42.119 --> 00:08:45.094
79, 80, 81, 82, and so on.

00:08:45.106 --> 00:08:48.738
The output is a probability
for each of the possible futures,

00:08:48.738 --> 00:08:50.496
a probability distribution.

00:08:50.496 --> 00:08:53.256
If you have that for each
of the single cases,

00:08:53.256 --> 00:08:56.518
then you can calculate
what the value of a decision is.

00:08:56.518 --> 00:08:58.358
That's prescriptive analytics.

00:08:58.358 --> 00:09:01.867
Use the predictions to take decisions
on the basis of utility functions,

00:09:01.867 --> 00:09:04.597
which are somehow known
or have to be found.

00:09:08.257 --> 00:09:12.117
That is still a management decision:
What is a value of this or that?

00:09:14.857 --> 00:09:20.726
We have to decide what is a value,
what we want to optimize.

00:09:20.726 --> 00:09:25.259
Prescriptive applications give recipes,
prescriptions, for every single decision.

00:09:26.489 --> 00:09:31.331
The recipe might be:
Take alternative two, buy 38 tomorrow.

00:09:32.601 --> 00:09:34.432
For some very fine tasks,

00:09:34.441 --> 00:09:36.739
the mathematical and technical complexity

00:09:36.743 --> 00:09:39.192
of the whole process,
of prediction and decision,

00:09:39.192 --> 00:09:41.724
can already be outsourced completely,

00:09:41.724 --> 00:09:48.052
and you can buy it in some cases
as a solution available in the Cloud.

00:09:48.702 --> 00:09:54.042
With predictive analytics, we very often
can attain two mutually exclusive goals.

00:09:54.042 --> 00:09:55.581
That's very interesting,

00:09:55.583 --> 00:09:58.469
because usually in management decisions,
you cannot do it.

00:09:58.474 --> 00:10:01.552
For example, in purchase
order optimization,

00:10:01.552 --> 00:10:06.107
you can have an increased availability
during the season,

00:10:06.107 --> 00:10:10.343
and simultaneously reduce what remains
at the end of the season,

00:10:10.740 --> 00:10:12.501
for example, in fashion retail.

00:10:12.504 --> 00:10:17.686
Or you can reduce the out-of-stock rate
in your meat shelves,

00:10:17.686 --> 00:10:21.456
but simultaneously,
you can reduce the meat waste

00:10:21.456 --> 00:10:23.597
at the end of the shelf lifetime.

00:10:24.097 --> 00:10:28.537
Or you can save 35%
of your marketing budget,

00:10:28.587 --> 00:10:32.379
but still have more customers
to buy your products.

00:10:33.229 --> 00:10:38.503
These are examples that are possible
with such technologies.

00:10:38.503 --> 00:10:42.982
We have also automated
the work of scientists,

00:10:42.982 --> 00:10:47.553
namely of my own science
as a particle physicist.

00:10:47.640 --> 00:10:52.391
So we have seen that
usually a PhD student or researcher

00:10:52.394 --> 00:10:54.904
has to do very similar things,

00:10:54.911 --> 00:10:59.373
and once you understand it,
you can try to optimize all this.

00:10:59.393 --> 00:11:02.935
We have found that usually
a PhD thesis on particle physics

00:11:02.935 --> 00:11:05.746
can be done by 72 decisions,

00:11:05.786 --> 00:11:08.824
and they all can be optimized.

00:11:09.294 --> 00:11:14.345
This was done for one
of the most successful experiments

00:11:14.345 --> 00:11:19.103
in experimental particle physics,
the Belle experiment in Japan.

00:11:19.423 --> 00:11:22.961
All the data having been
collected over ten years

00:11:22.961 --> 00:11:25.374
have been reprocessed with such a program,

00:11:25.374 --> 00:11:27.664
and the output of it was twice as good

00:11:27.664 --> 00:11:32.944
as 400 physicists have done,
by hand together, in ten years.

00:11:33.714 --> 00:11:37.684
That was the work of machine learning
or artificial intelligence,

00:11:37.684 --> 00:11:43.013
and three PhD students corresponds
to about 500 normal PhD theses,

00:11:43.013 --> 00:11:46.065
and corresponds to another
ten years of data taking,

00:11:46.065 --> 00:11:49.188
which would have cost 700 million euros.

00:11:50.618 --> 00:11:54.923
Very often we have too many decisions
to be made by a human.

00:11:55.073 --> 00:11:58.435
For example, 130,000 articles,
five productizes,

00:11:58.435 --> 00:12:03.447
three storage locations, two predictions
40 days in advance, and so on to do,

00:12:03.458 --> 00:12:07.535
and immediately you have some
100 million decisions.

00:12:07.550 --> 00:12:11.189
so it's always good then to automate.

00:12:11.197 --> 00:12:14.062
Because even if you have
100 million recipes,

00:12:14.065 --> 00:12:17.018
who's going to read all the recipes
and then decide what to do?

00:12:17.158 --> 00:12:22.429
That is where we see that automation
is the ultimate goal of what we want.

00:12:22.759 --> 00:12:26.614
The impact of decision automation 
can, for example, be seen here:

00:12:26.614 --> 00:12:30.289
So, this is the out-of-stock rate
at a German supermarket chain

00:12:30.289 --> 00:12:33.131
at constant overall stock level

00:12:33.131 --> 00:12:34.561
as a function of time.

00:12:34.561 --> 00:12:39.112
So, the original out-of-stock level
was about 7%,

00:12:39.112 --> 00:12:44.811
and then when the project started,
you see that black line there,

00:12:44.811 --> 00:12:49.112
it got less, that's what we want,
it's good if it's small,

00:12:49.125 --> 00:12:52.623
and then suddenly, one day,
we really automated it,

00:12:52.623 --> 00:12:55.492
and no humans were allowed
to do anything anymore.

00:12:55.492 --> 00:12:59.054
Then you see, the result
was much better, and constantly better.

00:12:59.364 --> 00:13:02.914
So that's the good or bad thing.

00:13:02.914 --> 00:13:04.934
So, three examples at the end.

00:13:04.934 --> 00:13:08.094
Automatic fresh-meat replenishment
for supermarket chains

00:13:08.094 --> 00:13:10.014
does work very well,

00:13:10.014 --> 00:13:14.744
and in some German supermarkets,
it is done automatically.

00:13:15.024 --> 00:13:17.106
Also dynamic pricing:

00:13:18.456 --> 00:13:22.018
in retail, it's something
that is appearing more frequently.

00:13:22.018 --> 00:13:24.507
With the Internet, it's normal already.

00:13:24.507 --> 00:13:27.006
It's also coming
in brick-and-mortar retail.

00:13:27.596 --> 00:13:29.851
It's not bad even for the customer, right?

00:13:29.851 --> 00:13:33.027
Because a retailer can never fight
against the customer,

00:13:33.027 --> 00:13:37.335
the customer has to want
to come and to buy,

00:13:37.335 --> 00:13:40.165
so it does not mean
that the price is always higher;

00:13:40.165 --> 00:13:45.807
it just means the right price,
at the right time, for the right article.

00:13:46.137 --> 00:13:52.332
So most retailers do not fight against
the customers, but against competitors,

00:13:52.602 --> 00:13:57.072
and the competitors also use
methods like these nowadays.

00:13:58.602 --> 00:14:03.982
So finally, another example from research,

00:14:04.192 --> 00:14:07.422
from the collaboration
of Blue Yonder and KIT,

00:14:07.422 --> 00:14:11.113
we have built such
a decision system onto one chip,

00:14:11.343 --> 00:14:15.221
and this now holds
the world record in decisions,

00:14:15.241 --> 00:14:17.662
it does eight billion
decisions per second,

00:14:17.732 --> 00:14:21.704
and that is for the next generation
of particle accelerators.

00:14:21.704 --> 00:14:27.912
So much data is produced in the sensors, 
that we cannot read them out anymore.

00:14:27.925 --> 00:14:33.193
So directly on the sensors,
there's a chip which decides:

00:14:33.201 --> 00:14:37.093
"Okay, in this event,
this part of the response

00:14:37.115 --> 00:14:39.443
seems to be important so we save it,

00:14:39.459 --> 00:14:41.480
we put it in a computer,

00:14:41.498 --> 00:14:44.449
and the others are not
even read out anymore."

00:14:44.491 --> 00:14:46.464
So, these are things that are coming.

00:14:46.472 --> 00:14:48.860
And here, it's very clear,
here it's simply time,

00:14:48.880 --> 00:14:51.426
no human would have the time
to make the decision,

00:14:51.435 --> 00:14:53.996
it has to be extremely fast.

00:14:54.262 --> 00:14:58.478
What I'm talking about here is not
the far future, but it exists now,

00:14:58.478 --> 00:15:01.368
and for sure, it will evolve further.

00:15:01.722 --> 00:15:06.678
So, my prediction is that
quite some repetitive routine work,

00:15:06.678 --> 00:15:08.478
but also white-collar work,

00:15:08.478 --> 00:15:13.071
so it's not only the workers
whose jobs will change in the future,

00:15:13.071 --> 00:15:17.061
but also, like me, white-collar work
will be automated.

00:15:17.061 --> 00:15:18.953
Should we be afraid of it?

00:15:19.513 --> 00:15:22.923
I think not, it is normal innovation,
it makes us stronger,

00:15:22.930 --> 00:15:29.791
it makes our strategic decisions
more efficient and also sustainable.

00:15:30.156 --> 00:15:32.505
I think that's the way
towards a better future,

00:15:32.505 --> 00:15:36.690
and I am completely convinced
that it is unavoidable. It will come.

00:15:36.696 --> 00:15:38.093
We don't have to discuss it,

00:15:38.097 --> 00:15:40.617
it's only a question of time,
it will come, for sure.

00:15:40.617 --> 00:15:41.707
Thank you very much.

00:15:41.707 --> 00:15:43.629
(Applause)