﻿1
00:00:00,233 --> 00:00:01,300
All right, let's do this.

2
00:00:01,300 --> 00:00:03,700
Let's implement Thompson sampling.

3
00:00:03,700 --> 00:00:06,600
So we already initialize,
you know, the main parameters

4
00:00:06,600 --> 00:00:10,766
which are part of the common framework
between Thompson sampling and UCB.

5
00:00:10,833 --> 00:00:14,666
So this is n the total number of rounds
or the total number of users

6
00:00:14,666 --> 00:00:17,700
to whom
we're going to show successively the ads.

7
00:00:17,966 --> 00:00:20,000
And then we have the number of ads
which is ten.

8
00:00:20,000 --> 00:00:22,566
Here we're dealing
with ten different designs of ads,

9
00:00:22,566 --> 00:00:26,733
which we are experimenting to figure out
which one has the highest conversion rate.

10
00:00:27,100 --> 00:00:30,600
And then we have this list
which is initialized as an empty list, but

11
00:00:30,600 --> 00:00:34,733
which will be populated over the rounds
with the different ads that we select.

12
00:00:34,733 --> 00:00:35,300
So at the end

13
00:00:35,300 --> 00:00:39,100
it will contain all the different ads
that were selected at each round.

14
00:00:39,466 --> 00:00:44,166
And now now what we're about to code
is specific to Thompson sampling.

15
00:00:44,300 --> 00:00:45,300
So there we go.

16
00:00:45,300 --> 00:00:47,366
We're going to start with the first step.

17
00:00:47,366 --> 00:00:50,900
Step one which says that at each round n

18
00:00:51,066 --> 00:00:55,966
we actually consider two numbers for each
ad I and I1N,

19
00:00:55,966 --> 00:00:58,966
which is the number of times two ad
I get reward one

20
00:00:59,033 --> 00:01:02,033
up to round n and an I0N

21
00:01:02,066 --> 00:01:05,600
the number of times that I got reward
zero up to round them.

22
00:01:05,833 --> 00:01:08,966
So basically what we have to do in this
step one is create

23
00:01:09,100 --> 00:01:12,066
two variables here for these two numbers.

24
00:01:12,066 --> 00:01:16,400
And we have to initialize them
of course as zero you know zero for both.

25
00:01:16,400 --> 00:01:19,400
Because of course before we begin,
you know the first round.

26
00:01:19,400 --> 00:01:23,133
Well these two numbers are equal to zero
because the ad didn't

27
00:01:23,133 --> 00:01:24,766
get any reward at all.

28
00:01:24,766 --> 00:01:26,966
So that's what I would like you to do.

29
00:01:26,966 --> 00:01:30,766
And since you know
these numbers are for each ad, I.

30
00:01:30,900 --> 00:01:34,800
Well, what we're going to do
is we will create a list of ten elements

31
00:01:35,000 --> 00:01:38,800
where each element is indeed this number
for each of the ads.

32
00:01:38,833 --> 00:01:40,500
Right. And same for this one.

33
00:01:40,500 --> 00:01:42,333
So now your turn.

34
00:01:42,333 --> 00:01:45,900
Please press pause on the video
and create two variables

35
00:01:45,900 --> 00:01:50,233
which will initialize
these two lists of numbers.

36
00:01:50,400 --> 00:01:52,500
And you can call these two lists first.

37
00:01:52,500 --> 00:01:55,800
For this one,
you can call it numbers of rewards, one

38
00:01:55,800 --> 00:02:00,300
you know with underscores and for this one
you can call it numbers of rewards.

39
00:02:00,333 --> 00:02:01,566
Zero okay.

40
00:02:01,566 --> 00:02:04,566
So please press pause and I'll give you
the solution in a second.

41
00:02:05,466 --> 00:02:07,766
Okay. Good. Let's do this together.

42
00:02:07,766 --> 00:02:11,933
So as we said we want to introduce
two new variables which will be two lists.

43
00:02:12,200 --> 00:02:17,400
First, for the numbers of times
each ad gets the reward one.

44
00:02:17,400 --> 00:02:21,900
And we're going to
call that numbers of rewards

45
00:02:23,033 --> 00:02:24,366
one. All right.

46
00:02:24,366 --> 00:02:28,933
And as we said we want to initialize it
as a list of ten elements.

47
00:02:29,300 --> 00:02:32,100
And all these elements
should be initialized to zeros

48
00:02:32,100 --> 00:02:35,266
because at the beginning
no AD got any reward.

49
00:02:35,266 --> 00:02:35,933
Right.

50
00:02:35,933 --> 00:02:39,366
And remember the trick to initialize
a list of ten zeros.

51
00:02:39,566 --> 00:02:42,400
It is to put a zero here
in square brackets,

52
00:02:42,400 --> 00:02:46,166
and then multiply this by the okay.

53
00:02:46,533 --> 00:02:53,066
And then very simply for the other list
you know corresponding to these numbers.

54
00:02:53,266 --> 00:02:54,400
Well same.

55
00:02:54,400 --> 00:02:57,400
We're going to call
that numbers of rewards

56
00:02:57,500 --> 00:03:00,433
not one but zero and same.

57
00:03:00,433 --> 00:03:03,500
We initialize it as a list of ten zeros.

58
00:03:03,500 --> 00:03:06,900
And over the rounds the elements
that these lists will be populated

59
00:03:06,966 --> 00:03:07,733
for this one.

60
00:03:07,733 --> 00:03:12,600
As soon as an ad receives
a reward, one in which case we increment

61
00:03:12,833 --> 00:03:16,466
the element corresponding
to the add by one and same for this one.

62
00:03:16,466 --> 00:03:20,733
Each time an ad gets a reward zero,
we increment the element

63
00:03:20,733 --> 00:03:22,733
corresponding to this ad in the list.

64
00:03:22,733 --> 00:03:24,300
Okay, so very simply.

65
00:03:24,300 --> 00:03:28,066
And that's the two parameters
specific to Thompson sampling.

66
00:03:28,200 --> 00:03:29,366
So if you got this two

67
00:03:29,366 --> 00:03:32,566
well congratulations you just implemented
step one.

68
00:03:33,100 --> 00:03:33,666
All right.

69
00:03:33,666 --> 00:03:36,233
So now according to you
what do we have to do.

70
00:03:36,233 --> 00:03:38,900
Well not yet the big four loop.

71
00:03:38,900 --> 00:03:41,700
Remember that we also want to keep
the accumulated

72
00:03:41,700 --> 00:03:44,700
reward that we, well,
you know, accumulate over time.

73
00:03:44,766 --> 00:03:48,233
And we're going to introduce
the same variable as in UCB.

74
00:03:48,266 --> 00:03:50,633
You know we could have even kept it
in the framework.

75
00:03:50,633 --> 00:03:53,866
But remember that variable
is total reward.

76
00:03:53,866 --> 00:03:56,100
So here we introduce the same new variable

77
00:03:56,100 --> 00:03:59,133
which of course
we initialize still to zero.

78
00:03:59,400 --> 00:04:01,733
And now now we can start the for loop.

79
00:04:01,733 --> 00:04:04,733
You know that main for loop
which will iterate

80
00:04:04,933 --> 00:04:07,800
through all the 10,000 rounds in each.

81
00:04:07,800 --> 00:04:12,733
We will show an ad to a user, which
will decide to click yes or no on the add.

82
00:04:12,933 --> 00:04:14,933
And so there we go. Let's stop this loop.

83
00:04:14,933 --> 00:04:18,600
So it's exactly the same as in the UCB
we still was for.

84
00:04:18,600 --> 00:04:22,900
Then we choose n for the iterated variable
because same and represents the rounds.

85
00:04:23,133 --> 00:04:28,500
And then and will go in the range
from same zero.

86
00:04:28,500 --> 00:04:32,833
Because indexes in Python
start from zero up to n, right.

87
00:04:32,833 --> 00:04:36,366
The total number of rounds,
or the total number of customers of users

88
00:04:36,566 --> 00:04:40,466
to whom we show successively the yhat
and then colon.

89
00:04:40,633 --> 00:04:41,600
And there we go.

90
00:04:41,600 --> 00:04:44,600
Here we are inside
this big first for loop.

91
00:04:44,700 --> 00:04:45,033
All right.

92
00:04:45,033 --> 00:04:48,033
So according to you,
what will be the first step here?

93
00:04:48,100 --> 00:04:50,566
You know,
there is a second full loop coming up.

94
00:04:50,566 --> 00:04:53,766
It is the full loop that will iterate
through the different ads.

95
00:04:54,000 --> 00:04:57,333
But before this
we need to introduce a new variable,

96
00:04:57,333 --> 00:05:00,433
which we already did
in the UCB implementation.

97
00:05:00,633 --> 00:05:03,100
Remember
we need to introduce that variable,

98
00:05:03,100 --> 00:05:06,700
which will be the
ad that we select at each front end.

99
00:05:06,966 --> 00:05:12,266
And as in the UCB implementation we will
introduce that variable as ad right

100
00:05:12,466 --> 00:05:17,600
ad which will be the index of the ad
that will be selected at each round end.

101
00:05:17,966 --> 00:05:20,700
And we initialize this to zero

102
00:05:20,700 --> 00:05:23,733
because you know we're going to have that
second for loop that will iterate through

103
00:05:23,733 --> 00:05:27,266
different ads starting from zero
up to the last one of index nine.

104
00:05:27,366 --> 00:05:28,200
Okay.

105
00:05:28,200 --> 00:05:31,800
Then let's have a look at the algorithm
and let's see

106
00:05:31,800 --> 00:05:33,866
if we need to introduce a new variable.

107
00:05:33,866 --> 00:05:36,366
You know,
before we start this second for loop,

108
00:05:36,366 --> 00:05:41,233
I remind that, you know, for UCB we indeed
had to introduce a new variable here,

109
00:05:41,233 --> 00:05:44,400
which was max up about,
you know, we introduced Max upper bound

110
00:05:44,400 --> 00:05:48,300
and initialized it to zero
before we started the second loop.

111
00:05:48,433 --> 00:05:50,366
Okay. So here that's kind of the same.

112
00:05:50,366 --> 00:05:53,700
Let's see if we have to introduce
a new variable other than add one

113
00:05:53,900 --> 00:05:56,800
before starting this second for loop.

114
00:05:56,800 --> 00:06:00,266
And well you know,
when we have a look at step three here,

115
00:06:00,466 --> 00:06:03,600
we see that indeed
we select the ad that has the highest

116
00:06:03,833 --> 00:06:07,500
of these random draws, you know,
taken from the beta distributions.

117
00:06:07,766 --> 00:06:11,366
And therefore there is still
a maximum value to consider here.

118
00:06:11,700 --> 00:06:15,233
And that's exactly what we will introduce
as a new variable

119
00:06:15,233 --> 00:06:17,666
just before starting this second for loop.

120
00:06:17,666 --> 00:06:21,933
That new variable will be exactly
that maximum of the random draws.

121
00:06:21,933 --> 00:06:25,500
And we're going to call it
max underscore random

122
00:06:25,766 --> 00:06:30,466
okay max random which we will initialize
of course to zero.

123
00:06:30,700 --> 00:06:33,700
And now now we can start the second loop.

124
00:06:33,700 --> 00:06:36,000
All right. So let's do this.
Try to do it faster than me.

125
00:06:36,000 --> 00:06:38,966
You know exactly how to do it.
We start with four.

126
00:06:38,966 --> 00:06:42,000
Then you know
the iterated variable will still be I,

127
00:06:42,000 --> 00:06:47,000
which will iterate through the different
ads from 0 to 9, because are the indexes.

128
00:06:47,233 --> 00:06:51,833
So for I in the range from indeed zero to

129
00:06:51,900 --> 00:06:57,800
well d, because we had to divide
both for the number of ads and then colon.

130
00:06:57,833 --> 00:06:59,066
And there we go.

131
00:06:59,066 --> 00:07:04,466
Now is the time we implement the next step
in the Thompson sampling algorithm.

132
00:07:04,766 --> 00:07:08,833
I'm talking of course about step two
where for each ad I.

133
00:07:08,866 --> 00:07:12,433
Well, that's
exactly what this second loop is about.

134
00:07:12,433 --> 00:07:15,666
You know, it will implement
step two for each of the ad.

135
00:07:15,666 --> 00:07:17,733
I from 0 to 9.

136
00:07:17,733 --> 00:07:20,733
And so for each of these ads,
we take a random draw

137
00:07:20,800 --> 00:07:25,166
from the data distribution
given below of parameters.

138
00:07:25,166 --> 00:07:29,300
And I n plus one where an n is
of course number of times

139
00:07:29,500 --> 00:07:33,466
that I got reward one up to run n
and the second parameter

140
00:07:33,466 --> 00:07:36,700
and i0 and plus one where NI0N is

141
00:07:36,700 --> 00:07:40,333
of course a number of times
I got reward zero up to running.

142
00:07:40,700 --> 00:07:45,633
So that means you can totally implement
this step to on your own,

143
00:07:45,633 --> 00:07:47,700
because you have
all the parameters, right.

144
00:07:47,700 --> 00:07:50,300
We already introduce
these two new variables.

145
00:07:50,300 --> 00:07:51,966
We initialize them the right way.

146
00:07:51,966 --> 00:07:55,400
And so now you can totally implement
this step two on your own

147
00:07:55,400 --> 00:07:58,033
before we do it together,
I will just give you a hint.

148
00:07:58,033 --> 00:07:59,566
Unless you don't want to listen to it

149
00:07:59,566 --> 00:08:03,366
and look into the documentation online,
which is even better.

150
00:08:03,566 --> 00:08:08,033
But I'll give you that hint
anyway, the way to get the random draw

151
00:08:08,033 --> 00:08:10,666
from a beta distribution
of these parameters

152
00:08:10,666 --> 00:08:15,700
is given by a certain function of well,
this random library, which we, by the way,

153
00:08:15,700 --> 00:08:19,166
already imported in our implementation,
that random library.

154
00:08:19,333 --> 00:08:22,200
Well, this random library contains
a certain function

155
00:08:22,200 --> 00:08:25,766
which can give you exactly
what you want here, meaning a random

156
00:08:25,766 --> 00:08:28,966
draw from the data distribution
of any parameters.

157
00:08:29,133 --> 00:08:32,700
And that function is called
beta var it in one word

158
00:08:32,733 --> 00:08:35,733
betta var iat.

159
00:08:36,200 --> 00:08:37,500
Okay, so that's the function.

160
00:08:37,500 --> 00:08:38,900
That's all I'm going to tell you.

161
00:08:38,900 --> 00:08:41,833
And so now you can really implement
this step to on your own.

162
00:08:41,833 --> 00:08:43,966
So please press pause on this video

163
00:08:43,966 --> 00:08:47,133
and will implement this solution
together in a few seconds.

164
00:08:47,600 --> 00:08:48,233
All right.

165
00:08:48,233 --> 00:08:48,966
Are you ready?

166
00:08:48,966 --> 00:08:50,433
I hope you got it correct.

167
00:08:50,433 --> 00:08:53,366
Really there is no trap
so it should be totally fine.

168
00:08:53,366 --> 00:08:54,733
But let's do this.

169
00:08:54,733 --> 00:08:55,533
So here we are

170
00:08:55,533 --> 00:08:58,800
at the beginning of the second full loop,
iterating through the different ads.

171
00:08:59,000 --> 00:09:02,433
And so what we have to do
is collect for each of these ads.

172
00:09:02,700 --> 00:09:05,733
This particular number taken
as a random draw of this

173
00:09:05,733 --> 00:09:09,200
beta distribution
called with these parameters.

174
00:09:09,200 --> 00:09:10,333
So let's do this.

175
00:09:10,333 --> 00:09:11,000
First we're going

176
00:09:11,000 --> 00:09:14,833
to introduce a new variable
which will be exactly that random draw.

177
00:09:14,833 --> 00:09:18,600
And we're going to call it
random underscore data.

178
00:09:18,866 --> 00:09:20,433
So that's exactly the random draw

179
00:09:20,433 --> 00:09:23,766
taken from the beta distribution
of these two parameters.

180
00:09:24,000 --> 00:09:25,033
So random beta.

181
00:09:25,033 --> 00:09:29,033
And then as I told you,
the way to get the random draw from a

182
00:09:29,066 --> 00:09:31,266
beta distribution of certain parameters.

183
00:09:31,266 --> 00:09:34,933
Well we first need to call the random
library

184
00:09:34,933 --> 00:09:39,733
from which we're going to call
the beta variate function.

185
00:09:39,933 --> 00:09:41,733
All right so that's the function.

186
00:09:41,733 --> 00:09:45,666
And inside this function well of course
we input these two parameters.

187
00:09:45,666 --> 00:09:49,300
And I'm plus one and I and zero plus one.

188
00:09:49,666 --> 00:09:50,733
And since right now

189
00:09:50,733 --> 00:09:55,233
we're dealing with a specific ad
you know the ad I in this second for loop.

190
00:09:55,233 --> 00:09:58,400
So it starts from zero
and then goes to 1 to 3 etc.

191
00:09:58,400 --> 00:09:59,400
up to nine.

192
00:09:59,400 --> 00:10:02,266
Well these two parameters
that we have to enter are

193
00:10:02,266 --> 00:10:06,400
for this specific
ad I meaning for the specific index

194
00:10:06,400 --> 00:10:10,600
I in these two numbers, and I want n
and i0 n okay.

195
00:10:10,900 --> 00:10:14,033
And therefore the way to get these
two numbers

196
00:10:14,033 --> 00:10:19,200
is by calling our list
of numbers of rewards one.

197
00:10:19,200 --> 00:10:22,366
You know, the number of times
each ad, and especially our ad,

198
00:10:22,366 --> 00:10:25,633
I we're dealing with right now
got the reward one.

199
00:10:25,833 --> 00:10:27,366
So we take this list.

200
00:10:27,366 --> 00:10:32,500
We paste it inside this beta void function
as the first parameter.

201
00:10:32,766 --> 00:10:35,500
And then of course we take the index I.

202
00:10:35,500 --> 00:10:35,833
Right.

203
00:10:35,833 --> 00:10:40,200
Because we're dealing right now with this
specific ad I in this second for loop.

204
00:10:40,533 --> 00:10:43,566
And then it's going to forget
the plus one.

205
00:10:44,266 --> 00:10:47,133
And then same for the second parameter.

206
00:10:47,133 --> 00:10:53,033
That's going to be the number of times
this particular ad I we're dealing with

207
00:10:53,033 --> 00:10:57,166
right now in the second
for loop has received the reward zero.

208
00:10:57,500 --> 00:10:59,966
So here I'm basing that.

209
00:10:59,966 --> 00:11:02,866
But I don't forget to take the index

210
00:11:02,866 --> 00:11:07,833
I of this ad I
and then of course I ad plus one okay.

211
00:11:07,833 --> 00:11:08,633
And that's it.

212
00:11:08,633 --> 00:11:11,533
That's all you had to do for this.
Step two.

213
00:11:11,533 --> 00:11:15,100
So you see it's actually a bit simpler
than the UCB algorithm.

214
00:11:15,300 --> 00:11:19,933
But now I trust that you're going to 
use the same right trick to implement

215
00:11:19,933 --> 00:11:20,500
step three.

216
00:11:20,500 --> 00:11:22,100
Because indeed in this step three

217
00:11:22,100 --> 00:11:26,400
we have to select the ad that has
the highest of these random draws.

218
00:11:26,400 --> 00:11:30,366
You know, we took ten successive random
draws for each of these ad,

219
00:11:30,400 --> 00:11:34,366
but inside the same for loop
that is iterating through the ad.

220
00:11:34,466 --> 00:11:39,466
We need to use a trick to keep in memory
the highest of these random draws.

221
00:11:39,466 --> 00:11:43,233
And at the end of the full loop, we'll get
the final highest of these random draws.

222
00:11:43,633 --> 00:11:48,100
So the trick that you have to use
is exactly the same as in UCB.

223
00:11:48,366 --> 00:11:53,200
You're going to use, of course, this
variable max random, which we introduced

224
00:11:53,433 --> 00:11:56,433
and initialized to zero
right before this second full loop.

225
00:11:56,600 --> 00:12:01,100
And you're going to use it to keep
in memory over this second for loop.

226
00:12:01,300 --> 00:12:03,633
Well the maximum of these random draws.

227
00:12:03,633 --> 00:12:06,000
So that's the new exercise for you.

228
00:12:06,000 --> 00:12:10,200
Your new exercise is basically to
implement step three with that same trick.

229
00:12:10,200 --> 00:12:13,200
We already learned
in the UCB implementation.

230
00:12:13,233 --> 00:12:17,100
And now I'll see you in next tutorial
to implement this solution with you.

231
00:12:17,466 --> 00:12:19,366
And until then enjoy machine learning.
