Presentation at CPA Australia Brisbane

1. AI and Probabilistic Reasoning
Dr. Jin-Song Dong
Professor, National University of Singapore
Professor and Director
Institute for Integrated Intelligent Systems (IIIS),
Griffith University
j.dong@griffith.edu.au
IIIS has about 40 professors/lectures and 70 PhD students across ICT, Eng,
Sci and Biz and has expertise in AI, Computer Vision, Speech, Robotic,
Data Analytics, Cybersecurity, Formal Verification & Analysis, etc.
1

2. Microsoft is focusing on AI
July 17, 2017 - July 18, 2017, Redmond, WA
Topics: Machine learning, Human language
technologies, Perception and sensing, Cyber-
physical systems and robotics

3. Acting humanly: Turing Test
• Turing (1950) "Computing machinery and intelligence":
• "Can machines think?"  "Can machines behave intelligently?"
• Operational test for intelligent behavior: the Imitation Game
• Suggested major components of AI:
- knowledge representation
- reasoning,
- language/image understanding,
- learning

4. Early History of AI
• 1940—50: Early days
1943: McCulloch & Pitts: Boolean circuit model of brain
1950: Turing's “Computing Machinery and Intelligence”
• 1950—70: Excitement:
1950s: Early AI programs, including Samuel's checkers program,
Newell & Simon's Logic Theorist, Gelernter's Geometry Engine
1956: Dartmouth meeting: “Artificial Intelligence” adopted
1965: Robinson's complete algorithm for logical reasoning
• 1970—88: Knowledge-based approaches
1969—79: Early development of knowledge-based systems
1980—88: Expert systems industry booms
• 1988—: Probabilistic & Statistical approaches
1988—93: Expert systems industry busts: “AI Winter”
Resurgence of probability, focus on uncertainty
General increase in technical depth
Agents and learning systems… “AI Spring”
4

5. Recent Achievements (Some)
• IBM’s Deep Blue defeated the reigning world chess
champion Garry Kasparov in 1997
• 2007: Checkers is solved by a research team at the
University of Alberta.
• No hands across America (CMU, driving autonomously
98% of the time from Pittsburgh to San Diego)
• NASA's on-board autonomous planning program
controlled the scheduling of operations for a spacecraft
• 2011: IBM’s Watson computer defeated TV game show
Jeopardy! champions. (probability based)
• 2011: Apple’s Siri
• 2015-6: Google DeepMind’s AlphaGo defeated 3 time
European Go champion and World Chamption Lee Sedol.
(probability based)
• 2016-7: Google driverless car Waymo
5

6. Intelligent Systems in Your Everyday Life
• Post Office
– automatic address recognition and sorting of mail
• Banks
– automatic check readers, signature verification systems
– automated loan application classification
• Telephone Companies
– automatic voice recognition for directory inquiries
• Credit Card Companies
– automated fraud detection
• Computer Companies
– automated diagnosis for help-desk applications
• Netflix:
– movie recommendation
• Google:
– Search Technology

7. AI Applications: Predicting the Stock Market
• The Prediction Problem
– given the past, predict the future
– very difficult problem!
– we can use learning algorithms to learn a predictive model from
historical data
– such models are routinely used by banks and financial traders to
manage portfolios worth millions of dollars
?
?
time in days
Value of
the Stock

8. AI and Accounting
partly according to: Eleanor O'Neill

9. KPMG with IBM’s Watson
• Watson Analytics read many thousands of
pages of contracts or agreements and quickly
summarise those based on given criteria and
it will learn over time and improving.
• It will take time for the tools to improve its
precision and it will take time for the
regulators to accept its service.

10. Deloitte with Kira Systems to aid in
contract and document reviewing
• Kira works by using examples uploaded by the user as
reference for accuracy, style and terminology in both English
and Latin.
• Deloitte-customised instances of the Kira platform for audit
processes and consulting have already been rolled out with
further applications being explored for tax and advisory
practices.
• This deployment helped Deloitte receive the Audit Innovation
of the Year award from the International Accounting Bulletin.

11. Machine Learning at Xero
• AI-based accounting systems help accountant make
fewer mistakes.
• Use machine learning to educate the accounting system
each time the accountant corrects an error.
• Next time the accountant creates an invoice, the account
code is automatically suggested.
• Xero chatbot uses machine learning to analyse financial
info and transection data, enabling queries about
– Who owns you money.
– When your next bill is due.
– How much money is in your account.

12. But Should We Be Worried?
• Computers and machines can now do many things that
used to require humans resources.
• 80% of accounting and finance tasks will be delivered
with automation in the next few years. -
Accenture
• But accountants can still provide more.
– Machines will have a larger role in accounting and finance.
– But it’s hard for machines to match human insights and
shower thoughts.
– Machines help accountants become more proficient and
productive.

13. The Future
AI and Data Analytics Integration
• Use AI to categorise expenditure and send accounts
to be checked.
• Accountants will be able to have streamlined
reporting doing all the heavy tax data lifting for
them.
• Finance sector will look into new powerful AI and Data
Analytics to help them to make smart decisions

14. AI/Data-Analytics is a major industry but
can lead to wrong decisions/predications

15. IIIS Long term focus direction/goals
Cyber Security
Trusted Intelligence
Data Analytics
Automated verification systems for making smart trusted decisions

16. J.Sun, SUTD
Y.Liu, NTU
J.S.Dong, NUS, Griffith
 1Million lines of C# code, 20+ verification systems, 200+ build in
examples, 10+ PhDs, 100+ publications.
 3700+ registered users from 900+ organizations in 89 countries,
e.g. Microsoft, HP, Sony, Mitsubishi, NTT, Toyota, JAXA …
Commercialised in Japan and thanks to CATS, NII and others

17. Probabilistic Reasoning [CAV’12]
• Syntax
– Hierarchical concurrent systems with probabilistic
choices
• Semantics
– Markov decision processes
• Given a property, probabilistic model checking
returns, instead of true or false
– the maximum and minimum probability of
satisfying the property.
17

18. • In search of a new car, the player picks a door, say 1. The game host
then opens one of the other doors, say 3, to reveal a goat and
offers to let the player pick door 2 instead of door 1. Should the
player take the offer?
• What if the host is dishonest, e.g., place car after 1st guess or host
do a switch 33% time after the guess?
18
Monty Hall Problem
The Monty Hall problem is based on
the American television game show
Let's Make a Deal and named after
the show's original host, Monty Hall.
The problem was originally posed in
a letter by Steve Selvin to the
American Statistician in 1975.

19. PAT Model
19

20. What if the host is Dishonest?
20

21. A New Application Domain –
Sports Analytics

22. Germany World Cup
2014 Success Secret?

23. Germany World Cup
success:
SAP Match Insights
Runs analytics and allows
coaches to target performance
metrics for specific players and
give them feedback –
mainly database queries with
a visualization tool.

24. Roger Federer
• 17 Grand Slam
• 302 weeks at #1
• 6 ATP Year End Titles
Rafael Nadal
• 14 Grand Slam
• 141 weeks at #1
• 0 ATP Year End Titles

25. Roger Federer Rafael Nadal
Strategy Analytics

26. Nadal’s strategy

27. Model Description

28. Building a PAT Location based MDP model
Federer
de_ct ad_ct
------+------ baseline
| 1 | 2 |
|-----|-----| service line
| 3 | 4 |
|===========| net
| 5 | 6 |
|-----|-----| service line
| 7 | 8 |
------+------ baseline
ad_ct de_ct
Nadal
"9" represents net error or
hit outside
28
enum{f_ad_ct, n_ad_ct, f_de_ct, n_de_ct};
//serve position: ad court or deuce court
enum{federer, nadal, na};
var turn = na; //serve turn;
var fscore = 0;
var nscore = 0;
var won = na;
var ball = 9;
WhoServe1st = []i:{f_de_ct,n_de_ct}@
TossCoin{turn = i} -> Skip;
TieBreakGame = WhoServe1st;
(FedererServe [] NadalServe);

29. Full Description !

30. Page Source

31. Page Source

32. Page Source
Extracting Information
and Events Probability
Software to automate the
process
Input the data to the
model

33. Great Analytics Correct Prediction for top 200 players
Bet on MarketProfit = testing our model

34. 2014 betting simulation: 827/3700 = 22.35%

36. Finding the best action to improve: vs
Right-handed players

37. Probabilistic Model Checking + Learning
Case Study: 4-Player Kuhn Poker
• Simplified Poker Game
• Probabilistic in Nature
• Requires Opponent Learning
38

38. 4-Player Kuhn Poker: Rules
• 4 Players
• 5 Cards
– TEN (lowest value) to ACE (highest value)
39
OPP
1
Play
er
OPP
2
OPP
3

39. 4-Player Kuhn Poker: Rules
• Each hand
– every player puts a chip down
– each player is dealt a card
– each player can choose to bet or not bet during their turn
40
OPP
1
PATt
y
OPP
2
OPP
3

40. 4-Player Kuhn Poker: Rules
• Game ends when:
– All players don’t Bet
– Some player has Bet, and all players have responded to the Bet
• Player who bets with highest card wins
41

41. 4-Player Kuhn Poker: Example
OPP1
Player
OPP2
OPP3
42

42. 4-Player Kuhn Poker: Example
OPP1
Player
OPP2
OPP3
43

43. 4-Player Kuhn Poker: Example
OPP1
Player
OPP2
OPP3
44

44. 4-Player Kuhn Poker: Example
OPP1
Player
OPP2
OPP3
45

45. 4-Player Kuhn Poker: Example
OPP1
Player
OPP2
OPP3
46

46. 4-Player Kuhn Poker: Example
OPP1
Player
OPP2
OPP3
47

47. 4-Player Kuhn Poker: Example
OPP1
Player
OPP2
OPP3
48

48. 4-Player Kuhn Poker: Example
OPP1
Player
OPP2
OPP3
49

49. 4-Player Kuhn Poker: Example
OPP1
Player
OPP2
OPP3
50

50. 4-Player Kuhn Poker: Example
OPP1
Player
OPP2
OPP3
51

51. 4-Player Kuhn Poker: Example
OPP1
Player
OPP2
OPP3
52

52. 4-Player Kuhn Poker: Example
OPP1
Player
OPP2
OPP3
53

53. 4-Player Kuhn Poker: Example
OPP1
Player
OPP2
OPP3
54

54. 4-Player Kuhn Poker: Example
OPP1
Player
OPP2
OPP3
55

55. 4-Player Kuhn Poker: Example
OPP1
Player
OPP2
OPP3
56

56. 4-Player Kuhn Poker: Example
OPP1
Player
OPP2
OPP3
57

57. 4-Player Kuhn Poker: Example
OPP1
Player
OPP2
OPP3
58

58. 4-Player Kuhn Poker: Example
OPP1
Player
OPP2
OPP3
59

59. 4-Player Kuhn Poker: Example
OPP1
Player
OPP2
OPP3
60

60. 4-Player Kuhn Poker: Playing Method
• There are only 3 ways players play in
– 00 - Do not bet on both first and second turns
– 01 - Do not bet on first turn, but bet on second turn
– 1X - Bet on first turn, and will not have second turn
61
OPP1
Playe
r
OPP2
OPP3
01
01
1X
00

61. 4-Player Kuhn Poker: Implementation
62
PATty
Observations
Table Learning
PAT Model
Learn Action
Input

62. 4-Player Kuhn Poker: Experimentation
Different configurations
X-Bluffer-Conservative-Bluffer (X-BCB)
X-Conservative-Bluffer-Conservative (X-CBC)
Where X can be either PATty, Bluffer or Conservative
30 games each
1500 hands per game
63

63. 4-Player Kuhn Poker: Findings
64
(PATty)
(PATty)

64. 4-Player Kuhn Poker: Findings
65
(PATty)
(PATty)

65. Event/Strategy Analytics
66