WEBVTT

00:00.330 --> 00:02.610
-: Hello, and welcome to the final level

00:02.610 --> 00:04.710
of the self-driving car challenge.

00:04.710 --> 00:08.880
So this time my goal will be to beat the self-driving car

00:08.880 --> 00:11.610
not the level four, because I want to challenge you

00:11.610 --> 00:14.760
into improving the code, improving the AI

00:14.760 --> 00:16.590
or maybe improving the strategy

00:16.590 --> 00:20.220
so that you can beat this challenge yourself.

00:20.220 --> 00:24.150
So I'm going to make a highly complex road

00:24.150 --> 00:27.240
that will still go from the airport to the downtown

00:27.240 --> 00:31.290
but maybe too complicated for the car to find its way.

00:31.290 --> 00:33.330
So that will be my challenge, so

00:33.330 --> 00:36.510
that I hope you can have fun with a challenging homework.

00:36.510 --> 00:37.680
So let's do this.

00:37.680 --> 00:41.160
Let's approach this level four.

00:41.160 --> 00:46.160
So let's select everything, then let's load the brain.

00:46.440 --> 00:47.490
And there we go.

00:47.490 --> 00:48.930
Now it's doing the round trips

00:48.930 --> 00:51.900
between the airport and the downtown.

00:51.900 --> 00:53.070
So it's gonna take some time.

00:53.070 --> 00:54.300
So I'm just gonna put

00:54.300 --> 00:56.943
on some music and here we go with the road.

00:59.612 --> 01:02.279
(gentle music)

01:32.383 --> 01:33.390
And there we go.

01:33.390 --> 01:34.530
Here is the road.

01:34.530 --> 01:37.320
And as you can see, the car seems to have a lot

01:37.320 --> 01:40.140
of trouble so that's exactly what I wanted.

01:40.140 --> 01:43.860
It's, you know, doing this little round trips

01:43.860 --> 01:45.470
in the same part of the road.

01:45.470 --> 01:46.680
So that's a problem.

01:46.680 --> 01:50.520
It doesn't find its way right now to the downtown.

01:50.520 --> 01:52.380
So how can we fix this?

01:52.380 --> 01:54.690
Can we, do we need to change the strategy?

01:54.690 --> 01:57.000
Do we need to change the parameters

01:57.000 --> 02:00.240
of the neural networks that has changed the synapses

02:00.240 --> 02:02.010
in the brain of the car?

02:02.010 --> 02:05.280
Do we need to do something with the rewards?

02:05.280 --> 02:09.100
You know, maybe get a worse reward

02:10.043 --> 02:12.720
when it's not finding the goal as it is the case right now?

02:12.720 --> 02:15.510
So so I don't know.

02:15.510 --> 02:17.640
That's for you to find out.

02:17.640 --> 02:21.840
As you can see, the rewards is not very severe

02:21.840 --> 02:23.517
when it's not finding the goal, you know

02:23.517 --> 02:26.370
and it's not getting closer to the goal.

02:26.370 --> 02:29.160
The reward is just minus oh 0.2.

02:29.160 --> 02:31.950
So that's not a severe bad reward.

02:31.950 --> 02:35.760
You could try to decrease the reward even more

02:35.760 --> 02:37.710
like setting a reward equals

02:37.710 --> 02:40.680
to minus oh 0.5, maybe that will work.

02:40.680 --> 02:43.860
I'm just throwing some suggestions to help you.

02:43.860 --> 02:47.310
So that's a change you can make with the reward strategy.

02:47.310 --> 02:48.143
And then of course

02:48.143 --> 02:53.143
you can make some other changes with your neural network.

02:53.850 --> 02:57.900
So you know, in this section we create the architecture

02:57.900 --> 02:58.733
of the neural network,

02:58.733 --> 03:03.150
we choose to have 30 hidden neurons in one hidden layer.

03:03.150 --> 03:06.480
Maybe you can try to change the architecture

03:06.480 --> 03:10.050
by trying some more layers or some more hidden neurons,

03:10.050 --> 03:11.040
I don't know.

03:11.040 --> 03:13.740
So that's another suggestion for improvement.

03:13.740 --> 03:17.070
And then you can also try to change something

03:17.070 --> 03:20.280
in a deep key learning algorithm, or simply you can move

03:20.280 --> 03:24.120
on to the course and find out about the other algorithms.

03:24.120 --> 03:25.420
Maybe there is another one

03:26.283 --> 03:28.830
that will manage to handle this situation.

03:28.830 --> 03:30.120
So good luck.

03:30.120 --> 03:32.520
I look forward to seeing your solutions.

03:32.520 --> 03:35.160
Try to draw that same road, that's perfect,

03:35.160 --> 03:38.310
that's really pretty exciting enigma.

03:38.310 --> 03:41.850
And if you want me to give you a hint, or

03:41.850 --> 03:44.730
if you want me to explain quickly what the problem is,

03:44.730 --> 03:47.100
well you can see that right here.

03:47.100 --> 03:49.740
When it reaches that point, it's going back.

03:49.740 --> 03:51.060
And that's because when it's reaching

03:51.060 --> 03:53.250
that point it's getting too far away

03:53.250 --> 03:55.830
from the goal which right now is the downtown.

03:55.830 --> 03:59.160
You know, when it's reaching that point, if it goes further,

03:59.160 --> 04:02.430
well it will go further away from the goal.

04:02.430 --> 04:04.920
So that's why it's going back here.

04:04.920 --> 04:07.243
And so you have to change something in a code,

04:07.243 --> 04:11.820
something in a strategy to maybe punish it less

04:11.820 --> 04:14.700
when it's getting further away from the goal.

04:14.700 --> 04:16.140
Maybe that's a solution.

04:16.140 --> 04:18.120
I'm just throwing some suggestions.

04:18.120 --> 04:19.740
I don't want to give you the solution too fast

04:19.740 --> 04:22.620
but that would be a really good exercise.

04:22.620 --> 04:25.350
And that's typically a problem that can be encountered

04:25.350 --> 04:28.650
by the engineers when they're making a self-driving car.

04:28.650 --> 04:30.930
But anyway, I think that's a pretty cool enigma.

04:30.930 --> 04:33.180
So I hope you will have fun.

04:33.180 --> 04:35.460
And before solving this homework

04:35.460 --> 04:38.520
maybe you can do another very efficient homework

04:38.520 --> 04:42.240
which would be simply to try to implement all this AI again

04:42.240 --> 04:44.190
you know, just implement the exact same one.

04:44.190 --> 04:46.170
That will be excellent practice

04:46.170 --> 04:48.810
because so far you've just listened to my explanations

04:48.810 --> 04:51.390
but there is a huge difference between listening

04:51.390 --> 04:54.270
to something and try to do the thing ourself.

04:54.270 --> 04:56.790
Oh, oh, did you see what happened?

04:56.790 --> 04:58.050
It just found its way.

04:58.050 --> 04:59.310
That's amazing.

04:59.310 --> 05:01.800
It's funny how it's got out of it

05:01.800 --> 05:03.720
but I'm sure there is a better strategy

05:03.720 --> 05:06.840
to get out of it more efficiently.

05:06.840 --> 05:08.280
And now it seems to be pretty good.

05:08.280 --> 05:10.410
Maybe I'm speaking too fast.

05:10.410 --> 05:11.730
Maybe it's going to solve this.

05:11.730 --> 05:12.720
No, here we go.

05:12.720 --> 05:13.883
It's getting stuck here again.

05:13.883 --> 05:16.183
That's pretty challenging for the car, right?

05:16.183 --> 05:18.720
But I'm sure this can be solved.

05:18.720 --> 05:20.511
That's your goal, so good luck,

05:20.511 --> 05:23.880
and please let me know about your solutions.

05:23.880 --> 05:25.863
Until then, enjoy AI.

05:27.021 --> 05:29.688
(gentle music)