Location via proxy:   [ UP ]  
[Report a bug]   [Manage cookies]                
SlideShare a Scribd company logo

How AlphaGo Works

•
•

The slides go through the implementation details of Google Deepmind's AlphaGo, a computer Go AI that defeated the European champion. The slides are targeted for beginners in the machine learning area. Korean version (한국어 버젼): http://www.slideshare.net/ShaneSeungwhanMoon/ss-59226902

1 of 54
Download to read offline

1

Presenter: Shane  (Seungwhan)  Moon PhD  student Language  Technologies  Institute,  School  of  Computer  Science Carnegie  Mellon  University 3/2/2016 How  it  works

2

AlphaGo vs  European  Champion  (Fan  Hui 2-­‐Dan) October  5  – 9,  2015 <Official  match> -­‐ Time  limit:  1  hour -­‐ AlphaGo Wins (5:0) * rank

3

AlphaGo vs  World  Champion  (Lee  Sedol 9-­‐Dan) March  9  – 15,  2016 <Official  match> -­‐ Time  limit:  2  hours Venue:  Seoul,  Four  Seasons  Hotel Image  Source: Josun  Times Jan  28th 2015

4

Lee  Sedol Photo  source: Maeil  Economics 2013/04 wiki

5

Computer  Go  AI?

6

Computer  Go  AI – Definition s (state) d  =  1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 = (e.g.  we  can  represent  the  board  into  a  matrix-­‐like  form) *  The  actual  model  uses  other  features  than  board  positions  as  well

7

Computer  Go  AI  – Definition s (state) d  =  1 d  =  2 a (action) Given  s,  pick  the  best  a Computer  Go Artificial   Intelligence s a s'

8

Computer  Go  AI – An  Implementation  Idea? d  =  1 d  =  2 … How  about  simulating  all  possible  board  positions?

9

Computer  Go  AI  – An  Implementation  Idea? d  =  1 d  =  2 … d  =  3 … … … …

10

Computer  Go  AI  – An  Implementation  Idea? d  =  1 d  =  2 … d  =  3 … … … … … d  =  maxD Process  the  simulation  until  the  game  ends, then  report  win  /  lose  results

11

Computer  Go  AI  – An  Implementation  Idea? d  =  1 d  =  2 … d  =  3 … … … … … d  =  maxD Process  the  simulation  until  the  game  ends, then  report  win  /  lose  results e.g. it  wins  13  times  if  the  next  stone  gets  placed  here 37,839  times 431,320  times Choose  the  “next  action  /  stone” that  has  the  most  win-­‐counts  in  the  full-­‐scale  simulation

12

This  is  NOT  possible;  it  is  said  the  possible  configurations  of  the  board  exceeds  the  number   of  atoms  in  the  universe

13

Key: To  Reduce Search  Space

14

Reducing  Search  Space 1.  Reducing  “action  candidates”  (Breadth  Reduction) d  =  1 d  =  2 … d  =  3 … … … … d  =  maxD Win? Loss? IF  there  is  a  model  that  can  tell  you  that  these  moves are  not  common  /  probable  (e.g.  by  experts,  etc.)  …

15

Reducing  Search  Space 1.  Reducing  “action  candidates”  (Breadth  Reduction) d  =  1 d  =  2 … d  =  3 … … d  =  maxD Win? Loss? Remove  these  from  search  candidates  in  advance (breadth  reduction)

16

Reducing  Search  Space 2.  Position  evaluation  ahead  of  time  (Depth  Reduction) d  =  1 d  =  2 … d  =  3 … … d  =  maxD Win? Loss? Instead  of  simulating  until  the  maximum  depth ..

17

Reducing  Search  Space 2.  Position  evaluation  ahead  of  time  (Depth  Reduction) d  =  1 d  =  2 … d  =  3 … V  =  1 V  =  2 V  =  10 IF  there  is  a  function  that  can  measure: V(s):  “board  evaluation  of  state  s”

18

Reducing  Search  Space 1. Reducing  “action  candidates”  (Breadth  Reduction) 2. Position  evaluation  ahead  of  time  (Depth  Reduction)

19

1.  Reducing  “action  candidates” Learning:  P  (  next  action  |  current  state  ) =  P  (  a  |  s  )

20

1.  Reducing  “action  candidates” (1) Imitating  expert  moves  (supervised  learning) Current  State Prediction   Model Next  State s1 s2 s2 s3 s3 s4 Data:  Online  Go experts (5~9  dan) 160K games, 30M  board  positions

21

1.  Reducing  “action  candidates” (1) Imitating  expert  moves  (supervised  learning) Prediction  Model Current  Board Next  Board

22

1.  Reducing  “action  candidates” (1) Imitating  expert  moves  (supervised  learning) Prediction  Model Current  Board Next  Action There  are  19  X  19  =  361 possible  actions (with  different  probabilities)

23

1.  Reducing  “action  candidates” (1) Imitating  expert  moves  (supervised  learning) Prediction  Model 0 0   0 0 0   0 0 0 0 0 0   0 0 0 1 0 0 0 0 -­‐1 0 0 1 -­‐1 1 0 0 0 1 0 0 1 -­‐1 0 0 0 0 0   0 0 -­‐1 0 0 0 0 0 0   0 0 0   0 0 0 0 0 -­‐1 0 0 0   0 0 0 0 0 0   0 0 0   0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 s af:  s à a Current  Board Next  Action

24

1.  Reducing  “action  candidates” (1) Imitating  expert  moves  (supervised  learning) Prediction   Model 0 0   0 0 0   0 0 0 0 0 0   0 0 0 1 0 0 0 0 -­‐1 0 0 1 -­‐1 1 0 0 0 1 0 0 1 -­‐1 0 0 0 0 0   0 0 -­‐1 0 0 0 0 0 0   0 0 0   0 0 0 0 0 -­‐1 0 0 0   0 0 0 0 0 0   0 0 0   0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 s g:  s à p(a|s) 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0             0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.2 0.1 0 0 0 0 0 0 0 0.4  0.2 0 0 0 0 0 0 0 0.1       0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 p(a|s) aargmax Current  Board Next  Action

25

1.  Reducing  “action  candidates” (1) Imitating  expert  moves  (supervised  learning) Prediction   Model 0 0   0 0 0   0 0 0 0 0 0   0 0 0 1 0 0 0 0 -­‐1 0 0 1 -­‐1 1 0 0 0 1 0 0 1 -­‐1 0 0 0 0 0   0 0 -­‐1 0 0 0 0 0 0   0 0 0   0 0 0 0 0 -­‐1 0 0 0   0 0 0 0 0 0   0 0 0   0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 s g:  s à p(a|s) p(a|s) aargmax Current  Board Next  Action

26

1.  Reducing  “action  candidates” (1) Imitating  expert  moves  (supervised  learning) Deep  Learning (13  Layer  CNN) 0 0   0 0 0   0 0 0 0 0 0   0 0 0 1 0 0 0 0 -­‐1 0 0 1 -­‐1 1 0 0 0 1 0 0 1 -­‐1 0 0 0 0 0   0 0 -­‐1 0 0 0 0 0 0   0 0 0   0 0 0 0 0 -­‐1 0 0 0   0 0 0 0 0 0   0 0 0   0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 s g:  s à p(a|s) 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0             0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.2 0.1 0 0 0 0 0 0 0 0.4  0.2 0 0 0 0 0 0 0 0.1       0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 p(a|s) aargmax Current  Board Next  Action

27

Convolutional  Neural  Network  (CNN) CNN  is  a  powerful  model  for  image  recognition  tasks;  it  abstracts  out  the  input  image  through  convolution  layers Image  source

28

Convolutional  Neural  Network  (CNN) And  they  use  this  CNN  model  (similar  architecture)  to  evaluate  the  board  position;  which learns  “some”  spatial  invariance

29

Go: abstraction  is  the  key  to  win CNN:  abstraction  is  its  forte

30

1.  Reducing  “action  candidates” (1) Imitating  expert  moves  (supervised  learning) Expert  Moves  Imitator  Model (w/  CNN) Current  Board Next  Action Training:

31

1.  Reducing  “action  candidates” (2) Improving  through  self-­‐plays  (reinforcement  learning) Expert  Moves   Imitator  Model (w/  CNN) Expert  Moves   Imitator  Model (w/  CNN) VS Improving  by  playing  against  itself

32

1.  Reducing  “action  candidates” (2) Improving  through  self-­‐plays  (reinforcement  learning) Expert  Moves   Imitator  Model (w/  CNN) Expert  Moves   Imitator  Model (w/  CNN) VS Return:  board  positions, win/lose info

33

1.  Reducing  “action  candidates” (2) Improving  through  self-­‐plays  (reinforcement  learning) Expert  Moves  Imitator  Model (w/  CNN) Board  position win/loss Training: Loss z  =  -­‐1

34

1.  Reducing  “action  candidates” (2) Improving  through  self-­‐plays  (reinforcement  learning) Expert  Moves  Imitator  Model (w/  CNN) Training: z  =  +1 Board  position win/loss Win

35

1.  Reducing  “action  candidates” (2) Improving  through  self-­‐plays  (reinforcement  learning) Updated  Model ver 1.1 Updated  Model ver 1.3VS Return:  board  positions, win/lose info It  uses  the  same  topology  as  the  expert  moves  imitator  model,  and  just  uses  the  updated parameters Older  models  vs.  newer  models

36

1.  Reducing  “action  candidates” (2) Improving  through  self-­‐plays  (reinforcement  learning) Updated  Model   ver 1.3 Updated  Model   ver 1.7VS Return:  board  positions, win/lose info

37

1.  Reducing  “action  candidates” (2) Improving  through  self-­‐plays  (reinforcement  learning) Updated  Model   ver 1.5 Updated  Model   ver 2.0VS Return:  board  positions, win/lose info

38

1.  Reducing  “action  candidates” (2) Improving  through  self-­‐plays  (reinforcement  learning) Updated  Model   ver 3204.1 Updated  Model   ver 46235.2VS Return:  board  positions, win/lose info

39

1.  Reducing  “action  candidates” (2) Improving  through  self-­‐plays  (reinforcement  learning) Updated  Model   ver 1,000,000VS The  final  model  wins 80%  of  the time when  playing  against  the  first  model Expert  Moves   Imitator  Model

40

2.  Board  Evaluation

41

2.  Board  Evaluation Updated  Model ver 1,000,000 Board  Position Training: Win  /  Loss Win (0~1) Value   Prediction   Model (Regression) Adds  a regression  layer  to  the  model Predicts  values  between  0~1 Close  to  1:  a  good  board  position Close  to  0:  a  bad  board  position

42

Reducing  Search  Space 1. Reducing  “action  candidates” (Breadth  Reduction) 2. Board  Evaluation (Depth  Reduction) Policy  Network Value  Network

43

Looking  ahead  (w/  Monte  Carlo  Search  Tree) Action  Candidates  Reduction (Policy  Network) Board  Evaluation (Value  Network) (Rollout):  Faster  version  of  estimating  p(a|s) à uses shallow  networks  (3  ms à 2µs)

44

Results Elo rating  system Performance  with  different  combinations  of  AlphaGo components

45

Takeaways Use  the  networks  trained  for  a  certain  task  (with  different  loss  objectives)  for  several  other  tasks

46

Lee  Sedol 9-­‐dan vs  AlphaGo

47

Lee  Sedol 9-­‐dan vs  AlphaGo Energy  Consumption Lee  Sedol AlphaGo -­‐ Recommended calories  for  a man per  day : ~2,500 kCal -­‐ Assumption: Lee consumes  the  entire  amount  of   per-­‐day calories  in  this  one  game 2,500  kCal *  4,184  J/kCal ~=  10M  [J] -­‐ Assumption: CPU:  ~100  W,  GPU:  ~300 W -­‐ 1,202 CPUs, 176 GPUs 170,000  J/sec  *  5  hr *  3,600  sec/hr ~=  3,000M  [J] A  very,  very  rough  calculation  ;)

48

AlphaGo is  estimated  to  be  around  ~5-­‐dan =  multiple  machines European  champion

49

Taking  CPU  /  GPU resources  to  virtually  infinity? But  Google  has  promised  not  to  use  more  CPU/GPUs than  they  used  for  Fan  Hui  for  the  game  with  Lee No  one  knows how it  will  converge

50

AlphaGo learns  millions  of  Go  games  every  day AlphaGo will  presumably  converge  to  some  point  eventually. However,  in  the  Nature  paper  they  don’t  report  how  AlphaGo’s performance  improves as  a  function  of  times  AlphaGo plays  against  itself  (self-­‐plays).

51

What  if  AlphaGo learns  Lee’s  game  strategy Google  said  they  won’t  use  Lee’s  game  plays  as  AlphaGo’s training  data   Even  if  it  does,  it  won’t  be  easy  to  modify  the  model  trained  over  millions  of data  points  with  just  a  few  game  plays  with  Lee (prone  to  over-­‐fitting,  etc.)

52

AlphaGo’s Weakness?

53

AlphaGo – How  It  Works Presenter: Shane  (Seungwhan)  Moon PhD  student Language  Technologies  Institute,  School  of  Computer  Science Carnegie  Mellon  University me@shanemoon.com 3/2/2016

54

Reference • Silver,  David,  et  al.  "Mastering  the  game  of  Go  with  deep  neural   networks  and  tree  search." Nature 529.7587  (2016):  484-­‐489.

More Related Content

How AlphaGo Works

  • 1. Presenter: Shane  (Seungwhan)  Moon PhD  student Language  Technologies  Institute,  School  of  Computer  Science Carnegie  Mellon  University 3/2/2016 How  it  works
  • 2. AlphaGo vs  European  Champion  (Fan  Hui 2-­‐Dan) October  5  – 9,  2015 <Official  match> -­‐ Time  limit:  1  hour -­‐ AlphaGo Wins (5:0) * rank
  • 3. AlphaGo vs  World  Champion  (Lee  Sedol 9-­‐Dan) March  9  – 15,  2016 <Official  match> -­‐ Time  limit:  2  hours Venue:  Seoul,  Four  Seasons  Hotel Image  Source: Josun  Times Jan  28th 2015
  • 4. Lee  Sedol Photo  source: Maeil  Economics 2013/04 wiki
  • 6. Computer  Go  AI – Definition s (state) d  =  1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 = (e.g.  we  can  represent  the  board  into  a  matrix-­‐like  form) *  The  actual  model  uses  other  features  than  board  positions  as  well
  • 7. Computer  Go  AI  – Definition s (state) d  =  1 d  =  2 a (action) Given  s,  pick  the  best  a Computer  Go Artificial   Intelligence s a s'
  • 8. Computer  Go  AI – An  Implementation  Idea? d  =  1 d  =  2 … How  about  simulating  all  possible  board  positions?
  • 9. Computer  Go  AI  – An  Implementation  Idea? d  =  1 d  =  2 … d  =  3 … … … …
  • 10. Computer  Go  AI  – An  Implementation  Idea? d  =  1 d  =  2 … d  =  3 … … … … … d  =  maxD Process  the  simulation  until  the  game  ends, then  report  win  /  lose  results
  • 11. Computer  Go  AI  – An  Implementation  Idea? d  =  1 d  =  2 … d  =  3 … … … … … d  =  maxD Process  the  simulation  until  the  game  ends, then  report  win  /  lose  results e.g. it  wins  13  times  if  the  next  stone  gets  placed  here 37,839  times 431,320  times Choose  the  “next  action  /  stone” that  has  the  most  win-­‐counts  in  the  full-­‐scale  simulation
  • 12. This  is  NOT  possible;  it  is  said  the  possible  configurations  of  the  board  exceeds  the  number   of  atoms  in  the  universe
  • 13. Key: To  Reduce Search  Space
  • 14. Reducing  Search  Space 1.  Reducing  “action  candidates”  (Breadth  Reduction) d  =  1 d  =  2 … d  =  3 … … … … d  =  maxD Win? Loss? IF  there  is  a  model  that  can  tell  you  that  these  moves are  not  common  /  probable  (e.g.  by  experts,  etc.)  …
  • 15. Reducing  Search  Space 1.  Reducing  “action  candidates”  (Breadth  Reduction) d  =  1 d  =  2 … d  =  3 … … d  =  maxD Win? Loss? Remove  these  from  search  candidates  in  advance (breadth  reduction)
  • 16. Reducing  Search  Space 2.  Position  evaluation  ahead  of  time  (Depth  Reduction) d  =  1 d  =  2 … d  =  3 … … d  =  maxD Win? Loss? Instead  of  simulating  until  the  maximum  depth ..
  • 17. Reducing  Search  Space 2.  Position  evaluation  ahead  of  time  (Depth  Reduction) d  =  1 d  =  2 … d  =  3 … V  =  1 V  =  2 V  =  10 IF  there  is  a  function  that  can  measure: V(s):  “board  evaluation  of  state  s”
  • 18. Reducing  Search  Space 1. Reducing  “action  candidates”  (Breadth  Reduction) 2. Position  evaluation  ahead  of  time  (Depth  Reduction)
  • 19. 1.  Reducing  “action  candidates” Learning:  P  (  next  action  |  current  state  ) =  P  (  a  |  s  )
  • 20. 1.  Reducing  “action  candidates” (1) Imitating  expert  moves  (supervised  learning) Current  State Prediction   Model Next  State s1 s2 s2 s3 s3 s4 Data:  Online  Go experts (5~9  dan) 160K games, 30M  board  positions
  • 21. 1.  Reducing  “action  candidates” (1) Imitating  expert  moves  (supervised  learning) Prediction  Model Current  Board Next  Board
  • 22. 1.  Reducing  “action  candidates” (1) Imitating  expert  moves  (supervised  learning) Prediction  Model Current  Board Next  Action There  are  19  X  19  =  361 possible  actions (with  different  probabilities)
  • 23. 1.  Reducing  “action  candidates” (1) Imitating  expert  moves  (supervised  learning) Prediction  Model 0 0   0 0 0   0 0 0 0 0 0   0 0 0 1 0 0 0 0 -­‐1 0 0 1 -­‐1 1 0 0 0 1 0 0 1 -­‐1 0 0 0 0 0   0 0 -­‐1 0 0 0 0 0 0   0 0 0   0 0 0 0 0 -­‐1 0 0 0   0 0 0 0 0 0   0 0 0   0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 s af:  s à a Current  Board Next  Action
  • 24. 1.  Reducing  “action  candidates” (1) Imitating  expert  moves  (supervised  learning) Prediction   Model 0 0   0 0 0   0 0 0 0 0 0   0 0 0 1 0 0 0 0 -­‐1 0 0 1 -­‐1 1 0 0 0 1 0 0 1 -­‐1 0 0 0 0 0   0 0 -­‐1 0 0 0 0 0 0   0 0 0   0 0 0 0 0 -­‐1 0 0 0   0 0 0 0 0 0   0 0 0   0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 s g:  s à p(a|s) 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0             0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.2 0.1 0 0 0 0 0 0 0 0.4  0.2 0 0 0 0 0 0 0 0.1       0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 p(a|s) aargmax Current  Board Next  Action
  • 25. 1.  Reducing  “action  candidates” (1) Imitating  expert  moves  (supervised  learning) Prediction   Model 0 0   0 0 0   0 0 0 0 0 0   0 0 0 1 0 0 0 0 -­‐1 0 0 1 -­‐1 1 0 0 0 1 0 0 1 -­‐1 0 0 0 0 0   0 0 -­‐1 0 0 0 0 0 0   0 0 0   0 0 0 0 0 -­‐1 0 0 0   0 0 0 0 0 0   0 0 0   0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 s g:  s à p(a|s) p(a|s) aargmax Current  Board Next  Action
  • 26. 1.  Reducing  “action  candidates” (1) Imitating  expert  moves  (supervised  learning) Deep  Learning (13  Layer  CNN) 0 0   0 0 0   0 0 0 0 0 0   0 0 0 1 0 0 0 0 -­‐1 0 0 1 -­‐1 1 0 0 0 1 0 0 1 -­‐1 0 0 0 0 0   0 0 -­‐1 0 0 0 0 0 0   0 0 0   0 0 0 0 0 -­‐1 0 0 0   0 0 0 0 0 0   0 0 0   0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 s g:  s à p(a|s) 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0             0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.2 0.1 0 0 0 0 0 0 0 0.4  0.2 0 0 0 0 0 0 0 0.1       0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 p(a|s) aargmax Current  Board Next  Action
  • 27. Convolutional  Neural  Network  (CNN) CNN  is  a  powerful  model  for  image  recognition  tasks;  it  abstracts  out  the  input  image  through  convolution  layers Image  source
  • 28. Convolutional  Neural  Network  (CNN) And  they  use  this  CNN  model  (similar  architecture)  to  evaluate  the  board  position;  which learns  “some”  spatial  invariance
  • 29. Go: abstraction  is  the  key  to  win CNN:  abstraction  is  its  forte
  • 30. 1.  Reducing  “action  candidates” (1) Imitating  expert  moves  (supervised  learning) Expert  Moves  Imitator  Model (w/  CNN) Current  Board Next  Action Training:
  • 31. 1.  Reducing  “action  candidates” (2) Improving  through  self-­‐plays  (reinforcement  learning) Expert  Moves   Imitator  Model (w/  CNN) Expert  Moves   Imitator  Model (w/  CNN) VS Improving  by  playing  against  itself
  • 32. 1.  Reducing  “action  candidates” (2) Improving  through  self-­‐plays  (reinforcement  learning) Expert  Moves   Imitator  Model (w/  CNN) Expert  Moves   Imitator  Model (w/  CNN) VS Return:  board  positions, win/lose info
  • 33. 1.  Reducing  “action  candidates” (2) Improving  through  self-­‐plays  (reinforcement  learning) Expert  Moves  Imitator  Model (w/  CNN) Board  position win/loss Training: Loss z  =  -­‐1
  • 34. 1.  Reducing  “action  candidates” (2) Improving  through  self-­‐plays  (reinforcement  learning) Expert  Moves  Imitator  Model (w/  CNN) Training: z  =  +1 Board  position win/loss Win
  • 35. 1.  Reducing  “action  candidates” (2) Improving  through  self-­‐plays  (reinforcement  learning) Updated  Model ver 1.1 Updated  Model ver 1.3VS Return:  board  positions, win/lose info It  uses  the  same  topology  as  the  expert  moves  imitator  model,  and  just  uses  the  updated parameters Older  models  vs.  newer  models
  • 36. 1.  Reducing  “action  candidates” (2) Improving  through  self-­‐plays  (reinforcement  learning) Updated  Model   ver 1.3 Updated  Model   ver 1.7VS Return:  board  positions, win/lose info
  • 37. 1.  Reducing  “action  candidates” (2) Improving  through  self-­‐plays  (reinforcement  learning) Updated  Model   ver 1.5 Updated  Model   ver 2.0VS Return:  board  positions, win/lose info
  • 38. 1.  Reducing  “action  candidates” (2) Improving  through  self-­‐plays  (reinforcement  learning) Updated  Model   ver 3204.1 Updated  Model   ver 46235.2VS Return:  board  positions, win/lose info
  • 39. 1.  Reducing  “action  candidates” (2) Improving  through  self-­‐plays  (reinforcement  learning) Updated  Model   ver 1,000,000VS The  final  model  wins 80%  of  the time when  playing  against  the  first  model Expert  Moves   Imitator  Model
  • 41. 2.  Board  Evaluation Updated  Model ver 1,000,000 Board  Position Training: Win  /  Loss Win (0~1) Value   Prediction   Model (Regression) Adds  a regression  layer  to  the  model Predicts  values  between  0~1 Close  to  1:  a  good  board  position Close  to  0:  a  bad  board  position
  • 42. Reducing  Search  Space 1. Reducing  “action  candidates” (Breadth  Reduction) 2. Board  Evaluation (Depth  Reduction) Policy  Network Value  Network
  • 43. Looking  ahead  (w/  Monte  Carlo  Search  Tree) Action  Candidates  Reduction (Policy  Network) Board  Evaluation (Value  Network) (Rollout):  Faster  version  of  estimating  p(a|s) à uses shallow  networks  (3  ms à 2µs)
  • 44. Results Elo rating  system Performance  with  different  combinations  of  AlphaGo components
  • 45. Takeaways Use  the  networks  trained  for  a  certain  task  (with  different  loss  objectives)  for  several  other  tasks
  • 46. Lee  Sedol 9-­‐dan vs  AlphaGo
  • 47. Lee  Sedol 9-­‐dan vs  AlphaGo Energy  Consumption Lee  Sedol AlphaGo -­‐ Recommended calories  for  a man per  day : ~2,500 kCal -­‐ Assumption: Lee consumes  the  entire  amount  of   per-­‐day calories  in  this  one  game 2,500  kCal *  4,184  J/kCal ~=  10M  [J] -­‐ Assumption: CPU:  ~100  W,  GPU:  ~300 W -­‐ 1,202 CPUs, 176 GPUs 170,000  J/sec  *  5  hr *  3,600  sec/hr ~=  3,000M  [J] A  very,  very  rough  calculation  ;)
  • 48. AlphaGo is  estimated  to  be  around  ~5-­‐dan =  multiple  machines European  champion
  • 49. Taking  CPU  /  GPU resources  to  virtually  infinity? But  Google  has  promised  not  to  use  more  CPU/GPUs than  they  used  for  Fan  Hui  for  the  game  with  Lee No  one  knows how it  will  converge
  • 50. AlphaGo learns  millions  of  Go  games  every  day AlphaGo will  presumably  converge  to  some  point  eventually. However,  in  the  Nature  paper  they  don’t  report  how  AlphaGo’s performance  improves as  a  function  of  times  AlphaGo plays  against  itself  (self-­‐plays).
  • 51. What  if  AlphaGo learns  Lee’s  game  strategy Google  said  they  won’t  use  Lee’s  game  plays  as  AlphaGo’s training  data   Even  if  it  does,  it  won’t  be  easy  to  modify  the  model  trained  over  millions  of data  points  with  just  a  few  game  plays  with  Lee (prone  to  over-­‐fitting,  etc.)
  • 53. AlphaGo – How  It  Works Presenter: Shane  (Seungwhan)  Moon PhD  student Language  Technologies  Institute,  School  of  Computer  Science Carnegie  Mellon  University me@shanemoon.com 3/2/2016
  • 54. Reference • Silver,  David,  et  al.  "Mastering  the  game  of  Go  with  deep  neural   networks  and  tree  search." Nature 529.7587  (2016):  484-­‐489.