Let's now look at some applications where
the relationships between prediction,
making decisions based on the predictions,
and then controlling when actions become
abundantly clear.
First example is Robo-soccer.
I don't know if many of you have tried
this but, you know, the regular
competitions of people building small
robots and getting them to play soccer.
And this is a little photograph of one
such competition.
Imagine what this robot has to do.
It needs to predict where the ball will
be.
Decide the best path to get to the ball
and navigate there.
And that's, that's the basic thing that it
needs to do if it doesn't have position of
the ball.
Prediction is very critical, the ball may
be moving, with a certain velocity, and
it's and the robot need to figure out
where it's going to go.
Much like our very first example of the
paddle game in the zeroth lecture of this
course.
Another example is self-driving cars.
You might have heard that Google, for
example, is developing a self-driving car,
which has already driven over 100,000
kilometers both in rough terrain as well
as in urban environments.
Imagine what such a car has to do.
Needs to predict the path of a pedestrian
that it see.
Decide the path it needs to follow to
avoid that person.
And then steer the car appropriately.
Prediction, decision, that is
optimization, best path.
And then control or steering.
We've all heard about the intelligent
energy grid, which is going to change the
way we consume electricity and natural
resources.
These things happened already today.
Great predict of the energy demand, decide
and control the distribution, all using
techniques very similar to those that we
have discussed, and these are already in
place.
Think about a business of having to deal
with a supply chain of products being
built, produced based on raw materials
which have to be sourced from rather
different locations around the world.
In order to decide which products to
produce and what quantities, a supply
chain needs to predict the demand for
every product.
Based on that, decide the best production
plan based on the raw materials available
and which products require which raw
materials in which quantities,
An optimization problem and then execute
it.
A very complex control system indeed,
which, which involves many, many moving
parts and many pieces of a complex
organization.
Another example of supply chain.
Is.
When products are manufactured all around
the world, there could be a variety of
risks.
You could have floods or fires or unstable
political conditions in different parts of
the world which might risk, pose a risk to
one supply chain.
These kinds of events can be detected
increasingly from social media.
Sources such as Twitter, once such risks
are detected or potential risks are
detected they need to be evaluated whether
or not they pose significant risks and if
so, the production plan needs to be
re-planned and then re-executed or.
Amendments to that plan need to be put
into place.
Finally, in marketing.
When one is trying to sell products, one
might need to predict the demand for a
product, decide the promotion strategy per
region based on the demand and then
execute the promotion strategy.
In each case we have the predict, decide
and control piece.
These examples are not necessarily simple,
but they can get increasingly complicated
as we shall see in a minute.
Think about the examples below, which are
grayed out now.
And compare these to robo software self
driving car energy grid.
They'd, there does appear to be something
fundamentally different about the ones at
the bottom and the ones at top.
For example, the ones on the top are
possibly executable or implementable
using, say, a system which will predict
the ball.
Directly take that prediction.
Decide the best path using some
optimization techniques.
And then navigate there using a simple
control system.
Similarly for self-driving cars.
In fact, auto-pilots for airplanes are
already part and parcel of every day life.
Cars are a little bit more complicated,
because of things like pedestrians and
overtaking and lanes.
But still, one can imagine.
Which is, you know, a suitably implemented
system actually doing this, and simulate
the energy great as that said is already
there, and running already today.
However, consider some of these examples
below.
Detecting potential risks from say, tweets
is possible, but then evaluating the
impact requires certain degree of fairly
complicated reasoning.
Similarly, prediction demand might be easy
but deciding the right promotion strategy
by region might be a fairly complicated
exercise.
The problems above can also be made more
complicated.
For example in addition to predicting
where the ball will be, one might want to
predict how other players will move.
And this requires a certain degree of
reasoning.
Similarly, one might need to predict
traffic based on possible inputs from many
different people about where they want to
go that day.
Based on that, you might want to decide
the optimal roots for to destination for
everybody.
And in some sense, a self-driving car
might manifest itself into a network of
many, many self driving cars.
Along with people deciding where they want
to go through their cell phones.
And all the cars magically getting
everybody where they want to be with
minimal traffic.
A fairly utopian situation but, in
principle, possible.
Again, a little more complicated
decision-making needs to happen.
Reasoning to a certain extend, a lot of
planning for sure.
Similarly in the energy grid, you might
want to predict the supply by how green
that supply is.
Whether it's coming from a wind turbine or
from a coal fired power plant, and adjust
the prices optimally so that one could,
maximize the utility of the green power
when it's available, that is, let's say
when the wind is blowing, but, not,
necessarily, make people pay too much if
there's no alternative to a coal fired
power at a particular time.
Again, more complicated predict the site
control systems.
All these are, today being built or are
already in production to a certain extent.
And because of the number of devices, the
number of censors and information on the
web, on social media.
And the interconnection of almost
everything in general.
I definitely call all of these potential
web intelligence systems.
They incorporate many of the techniques
that we have talked about, as well as many
that we haven't such as, optimization and
control.
Well, having seen a number of applications
this lecture we are talking about
prediction and, and certainly learning, so
we will now briefly come back to the
techniques and try to see which techniques
one should use in which situations.