Share
Embed
Slide 1
Slide 1 text
March 2016
Slide 2
Slide 2 text
Mainstream Media
Slide 3
Slide 3 text
1997
Slide 4
Slide 4 text
Ing cup 1985 – 2000
(up to 1,400,000$)
(1985-2000)
Slide 5
Slide 5 text
5d win 1998
Slide 6
Slide 6 text
October 2015
Slide 7
Slide 7 text
This is the first time that a
computer program has defeated
a human professional player in the
full-sized game of Go, a feat
previously thought to be at
least a decade away.
Silver, D. et al., 2016. Mastering the game of Go with deep neural
networks and tree search. Nature, 529(7587), p.484-489.
January 2016
Slide 8
Slide 8 text
What did AlphaGo do to beat the
strongest human Go player?
Tobias Pfeiffer
@PragTob
pragtob.info
Slide 9
Slide 9 text
Go
Slide 10
Slide 10 text
Computational Challenge
Slide 11
Slide 11 text
Monte Carlo Method
Slide 12
Slide 12 text
Neural Networks
Slide 13
Slide 13 text
Revolution with Neural Networks
Slide 14
Slide 14 text
What did we learn?
Slide 15
Slide 15 text
Go
Slide 16
Slide 16 text
No content
Slide 17
Slide 17 text
No content
Slide 18
Slide 18 text
No content
Slide 19
Slide 19 text
No content
Slide 20
Slide 20 text
No content
Slide 21
Slide 21 text
No content
Slide 22
Slide 22 text
No content
Slide 23
Slide 23 text
No content
Slide 24
Slide 24 text
No content
Slide 25
Slide 25 text
No content
Slide 26
Slide 26 text
No content
Slide 27
Slide 27 text
No content
Slide 28
Slide 28 text
No content
Slide 29
Slide 29 text
No content
Slide 30
Slide 30 text
No content
Slide 31
Slide 31 text
No content
Slide 32
Slide 32 text
No content
Slide 33
Slide 33 text
No content
Slide 34
Slide 34 text
No content
Slide 35
Slide 35 text
Computational Challenge
Slide 36
Slide 36 text
Go vs. Chess
Slide 37
Slide 37 text
Complex vs. Complicated
Slide 38
Slide 38 text
„While the Baroque rules of chess could only
have been created by humans, the rules of
go are so elegant, organic, and rigorously
logical that if intelligent life forms exist
elsewhere in the universe, they almost
certainly play go.“
Edward Lasker (chess grandmaster)
Slide 39
Slide 39 text
Larger board
19x19 vs. 8x8
Slide 40
Slide 40 text
Almost every move is legal
Slide 41
Slide 41 text
Average branching factor:
250 vs 35
Slide 42
Slide 42 text
State Space Complexity:
10171 vs 1047
Slide 43
Slide 43 text
1080
Slide 44
Slide 44 text
Global impact of moves
Slide 45
Slide 45 text
6
8
9
5
7
9
6
6
3
5
4
7
6
5
6
8
5
7
6
6
3
4
5
8
5
7
6
3
5
5
6
3
6
MAX
MIN
MAX
MIN
MAX
Traditional Seach
Slide 46
Slide 46 text
6
8
9
5
7
9
6
6
3
5
4
7
6
5
6
8
5
7
6
6
3
4
5
8
5
7
6
3
5
5
6
3
6
MAX
MIN
MAX
MIN
MAX
Evaluation Function
Slide 47
Slide 47 text
No content
Slide 48
Slide 48 text
Monte Carlo Method
Slide 49
Slide 49 text
What is Pi?
Slide 50
Slide 50 text
How do you determine Pi?
Slide 51
Slide 51 text
No content
Slide 52
Slide 52 text
2006
Slide 53
Slide 53 text
Browne, Cb, and Edward Powley. 2012. A survey of monte
carlo tree search methods. Intelligence and AI 4, no. 1: 1-49
Slide 54
Slide 54 text
2/4
1/1 0/1 1/1 0/1
A1
D5
F13
C7
Slide 55
Slide 55 text
2/4
1/1 0/1 1/1 0/1
A1
D5
F13
C7
Selection
Slide 56
Slide 56 text
2/4
1/1 0/1 1/1 0/1
A1
D5
F13
C7
0/0
B5
Expansion
Slide 57
Slide 57 text
2/4
1/1 0/1 1/1 0/1
A1
D5
F13
C7
0/0
B5
Simulation
Slide 58
Slide 58 text
Random
Slide 59
Slide 59 text
Not Human like?
Slide 60
Slide 60 text
3/5
2/2 0/1 1/1 0/1
A1
D5
F13
C7
1/1
B5
Backpropagation
Slide 61
Slide 61 text
3/5
2/2 0/1 1/1 0/1
A1
D5
F13
C7
1/1
B5
Perspective
Slide 62
Slide 62 text
2/5
1/2 0/1 1/1 0/1
A1
D5
F13
C7
1/1
B5
Perspective
Slide 63
Slide 63 text
Multi Armed Bandit
Slide 64
Slide 64 text
Multi Armed Bandit
Exploitation vs Exploration
Slide 65
Slide 65 text
wins
visits
+explorationFactor
√ln(totalVisits)
visits
Slide 66
Slide 66 text
15042
86/193
0/1 1/2 0/2
36/1116
2/2
58/151
1/2 0/2
3/3
Slide 67
Slide 67 text
15042
86/193
0/1 1/2 0/2
36/1116
2/2
58/151
1/2 0/2
3/3
Slide 68
Slide 68 text
15042
86/193
0/1 1/2 0/2
36/1116
2/2
58/151
1/2 0/2
3/3
Slide 69
Slide 69 text
Generate a valid random move
Slide 70
Slide 70 text
Who has won?
Slide 71
Slide 71 text
No content
Slide 72
Slide 72 text
General Game Playing
Slide 73
Slide 73 text
Anytime
Slide 74
Slide 74 text
Lazy
Slide 75
Slide 75 text
No content
Slide 76
Slide 76 text
Expert Knowledge
Slide 77
Slide 77 text
Neural Networks
Slide 78
Slide 78 text
2014
Slide 79
Slide 79 text
What does this even mean?
Slide 80
Slide 80 text
Neural Networks
Slide 81
Slide 81 text
Input
“Hidden” Layer
Output
Neural Networks
Slide 82
Slide 82 text
Weights
Slide 83
Slide 83 text
Bias/Threshold
Slide 84
Slide 84 text
4
2
-3
3.2
Activation
Slide 85
Slide 85 text
5.2 >= 4
2
-3
3.2
Activation
Slide 86
Slide 86 text
2.2 <= 4
2
-3
3.2
Activation
Slide 87
Slide 87 text
Activation
Slide 88
Slide 88 text
Training
Slide 89
Slide 89 text
Adjust parameters
Slide 90
Slide 90 text
Supervised Learning
Input
Expected
Output
Slide 91
Slide 91 text
Backpropagation
Slide 92
Slide 92 text
Data set
Slide 93
Slide 93 text
Training data + test data
Slide 94
Slide 94 text
Training
Slide 95
Slide 95 text
Verify
Slide 96
Slide 96 text
Overfitting
Slide 97
Slide 97 text
Deep Neural Networks
Slide 98
Slide 98 text
Convolutional Neural Networks
Slide 99
Slide 99 text
Local Receptive Field
Slide 100
Slide 100 text
Feature Map
Slide 101
Slide 101 text
Stride
Slide 102
Slide 102 text
Shared weights and biases
Slide 103
Slide 103 text
19 x 19
3 x 17 x 17
Multiple Feature maps/filters
Slide 104
Slide 104 text
Architecture
...
Input Features
12 layers with 64 – 192
filters
Output
Slide 105
Slide 105 text
Architecture
...
Input Features
12 layers with 64 – 192
filters
Output
Slide 106
Slide 106 text
Architecture
...
Input Features
12 layers with 64 – 192
filters
Output
Slide 107
Slide 107 text
2.3 million parameters
630 million connections
Slide 108
Slide 108 text
●
Stone Colour x 3
●
Liberties x 4
●
Liberties after move played x 6
●
Legal Move x 1
●
Turns since x 5
●
Capture Size x 7
●
Ladder Move x 1
●
KGS Rank x 9
Input Features
Slide 109
Slide 109 text
Training on game data predicting
the next move
Slide 110
Slide 110 text
55% Accuracy
Slide 111
Slide 111 text
Mostly beats GnuGo
Slide 112
Slide 112 text
Combined with MCTS in the
Selection
Slide 113
Slide 113 text
Asynchronous GPU Power
Slide 114
Slide 114 text
Revolution
Slide 115
Slide 115 text
Silver, D. et al., 2016. Mastering the game of Go with deep neural
networks and tree search. Nature, 529(7587), p.484-489.
Networks in Training
Slide 116
Slide 116 text
Silver, D. et al., 2016. Mastering the game of Go with deep neural
networks and tree search. Nature, 529(7587), p.484-489.
Networks in Training
Slide 117
Slide 117 text
Silver, D. et al., 2016. Mastering the game of Go with deep neural
networks and tree search. Nature, 529(7587), p.484-489.
AlphaGo Search
Slide 118
Slide 118 text
Selection
Slide 119
Slide 119 text
Action Value
Prior Probability
Visit Count
Selection
Slide 120
Slide 120 text
Action Value
Prior Probability
Visit Count
Slide 121
Slide 121 text
Action Value
Prior Probability
Visit Count
Selection
Slide 122
Slide 122 text
Action Value
Prior Probability
Visit Count
Selection
Slide 123
Slide 123 text
0.8
1.2 0.5 1.1 0.9
Action Value + Bonues
Slide 124
Slide 124 text
0.8
1.2 0.5 1.1 0.9
Expansion
Slide 125
Slide 125 text
0.8
1.2 0.5 1.1 0.9
Expansion
Slide 126
Slide 126 text
0.8
1.2 0.5 1.1 0.9
Prior Probability
Slide 127
Slide 127 text
0.8
1.2 0.5 1.1 0.9
Evalution
Slide 128
Slide 128 text
0.8
1.2 0.5 1.1 0.9
Evalution
Slide 129
Slide 129 text
0.8
1.2 0.5 1.1 0.9
Rollout
Slide 130
Slide 130 text
0.8
1.2 0.5 1.1 0.9
Value Network
Slide 131
Slide 131 text
0.81
1.3 0.5 1.1 0.9
Backup
1.6
Slide 132
Slide 132 text
1202 CPUs 176 GPUS
0.8
1.2 0.5 1.1 0.9
Slide 133
Slide 133 text
Tensor
Slide 134
Slide 134 text
Human Instinct Policy Network
Reading Capability Search
Positional Judgement Value Network
3 Strengths of AlphaGo
Slide 135
Slide 135 text
Human Instinct Policy Network
Reading Capability Search
Positional Judgement Value Network
Most Important Strength
Slide 136
Slide 136 text
More Natural
Slide 137
Slide 137 text
Lee Sedol match
Slide 138
Slide 138 text
Style
Slide 139
Slide 139 text
So when AlphaGo plays a slack looking move,
we may regard it as a mistake,
but perhaps it should more accurately be viewed
as a declaration of victory?
An Younggil 8p
Slide 140
Slide 140 text
Game 2
Slide 141
Slide 141 text
Game 4
Slide 142
Slide 142 text
No content
Slide 143
Slide 143 text
Game 4
Slide 144
Slide 144 text
Game 4
Slide 145
Slide 145 text
What can we learn?
Slide 146
Slide 146 text
Making X faster
vs
Doing less of X
Slide 147
Slide 147 text
Benchmark everything
Slide 148
Slide 148 text
Solving problems the human way
vs
Solving problems the computer
way
Slide 149
Slide 149 text
Don't blindly dismiss approaches
as infeasible
Slide 150
Slide 150 text
One Approach
vs
Combination of Approaches
Slide 151
Slide 151 text
Joy of Creation
Slide 152
Slide 152 text
PragTob/Rubykon
Slide 153
Slide 153 text
pasky/michi
Slide 154
Slide 154 text
What did AlphaGo do to beat the
strongest human Go player?
Tobias Pfeiffer
@PragTob
pragtob.info
Slide 155
Slide 155 text
Sources
●
Maddison, C.J. et al., 2014. Move Evaluation in Go Using Deep Convolutional Neural Networks.
●
Silver, D. et al., 2016. Mastering the game of Go with deep neural networks and tree search.
Nature, 529(7587), p.484-489.
●
Michael A. Nielsen, "Neural Networks and Deep Learning", Determination Press, 2015
http://neuralnetworksanddeeplearning.com
●
Gelly, S. & Silver, D., 2011. Monte-Carlo tree search and rapid action value estimation in
computer Go. Artificial Intelligence, 175(11), p.1856-1876.
●
I. Althöfer, “On the Laziness of Monte-Carlo Game Tree Search In Non-tight Situations,”
Friedrich-Schiller Univ., Jena, Tech. Rep., 2008.
●
Browne, C. & Powley, E., 2012. A survey of monte carlo tree search methods. IEEE Transactions
on Intelligence and AI in Games, 4(1), p.1-49.
●
Gelly, S. & Silver, D., 2007. Combining online and offline knowledge in UCT. Machine Learning,
p.273-280.
●
https://www.youtube.com/watch?v=LX8Knl0g0LE&index=9&list=WL
Slide 156
Slide 156 text
Photo Credit
●
http://www.computer-go.info/events/ing/2000/images/bigcup.jpg
●
https://en.wikipedia.org/wiki/File:Kasparov-29.jpg
●
http://www.geforce.com/hardware/desktop-gpus/geforce-gtx-titan-black/product-images
●
http://giphy.com/gifs/dark-thread-after-lCP95tGSbMmWI
●
https://cloudplatform.googleblog.com/2016/05/Google-supercharges-machine-learning-tasks-with-custom-chi
p.html
●
https://gogameguru.com/i/2016/01/Fan-Hui-vs-AlphaGo-550x364.jpg
●
http://makeitstranger.com/
●
CC BY 2.0
– https://en.wikipedia.org/wiki/File:Deep_Blue.jpg
– https://www.flickr.com/photos/luisbg/2094497611/
●
CC BY-SA 3.0
– https://en.wikipedia.org/wiki/Alpha%E2%80%93beta_pruning#/media/File:AB_pruning.svg
●
CC BY-SA 2.0
– https://flic.kr/p/cPUtny
– https://flic.kr/p/dLSKTQ
– https://www.flickr.com/photos/83633410@N07/7658272558/
Slide 157
Slide 157 text
Photo Credit
●
CC BY-NC-ND 2.0
– https://flic.kr/p/q15pzb
– https://flic.kr/p/bHSj7D
– https://flic.kr/p/ixSsfM
– https://www.flickr.com/photos/waxorian/4228645447/
– https://www.flickr.com/photos/pennstatelive/8972110324/
– https://www.flickr.com/photos/dylanstraub/6428496139/
●
https://en.wikipedia.org/wiki/Alphabet_Inc.#/media/File:Alphabet_Inc_Logo_2015.svg
●
CC BY 3.0
– https://en.wikipedia.org/wiki/File:Pi_30K.gif