Top Banner
Game Theory for Humans Matt Hawrilenko MIT: Poker Theory and Analytics 1
58

Game Theory for Humans - MIT OpenCourseWare · Game Theory for Humans Matt Hawrilenko MIT: Poker Theory and Analytics 1

Aug 04, 2018

Download

Documents

vantuong
Welcome message from author
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Page 1: Game Theory for Humans - MIT OpenCourseWare · Game Theory for Humans Matt Hawrilenko MIT: Poker Theory and Analytics 1

Game Theory for Humans

Matt Hawrilenko MIT: Poker Theory and Analytics

1

Page 2: Game Theory for Humans - MIT OpenCourseWare · Game Theory for Humans Matt Hawrilenko MIT: Poker Theory and Analytics 1

Game Theoretic Approach

Computer Scientists Humans

Theory Practice

Toy Games Real Hand

Play Good Poker

Read-Based Approach

Rules/How-To Training Principles

Why Game Theory?

Audience

Bridge Theory and Practice

Strategy for Continual

Development

2

Page 3: Game Theory for Humans - MIT OpenCourseWare · Game Theory for Humans Matt Hawrilenko MIT: Poker Theory and Analytics 1

For example . . .

3

Page 4: Game Theory for Humans - MIT OpenCourseWare · Game Theory for Humans Matt Hawrilenko MIT: Poker Theory and Analytics 1

Pot: $119,000

4

Page 5: Game Theory for Humans - MIT OpenCourseWare · Game Theory for Humans Matt Hawrilenko MIT: Poker Theory and Analytics 1

Pot: $160,000

5

Page 6: Game Theory for Humans - MIT OpenCourseWare · Game Theory for Humans Matt Hawrilenko MIT: Poker Theory and Analytics 1

Pot: $310,000

6

Page 7: Game Theory for Humans - MIT OpenCourseWare · Game Theory for Humans Matt Hawrilenko MIT: Poker Theory and Analytics 1

Pot: $720,000

7

Page 8: Game Theory for Humans - MIT OpenCourseWare · Game Theory for Humans Matt Hawrilenko MIT: Poker Theory and Analytics 1

$310,000 $720,000

$160,000 $119,000

? 8

Page 9: Game Theory for Humans - MIT OpenCourseWare · Game Theory for Humans Matt Hawrilenko MIT: Poker Theory and Analytics 1

Screenshot of Daniel Negreanu © ESPN/WSOP. All rights reserved. This content is excluded fromour Creative Commons license. For more information, see http://ocw.mit.edu/help/faq-fair-use/.

9

Page 10: Game Theory for Humans - MIT OpenCourseWare · Game Theory for Humans Matt Hawrilenko MIT: Poker Theory and Analytics 1

Screenshot of Phil Hellmuth © ESPN. All rights reserved. This content is excluded from ourCreative Commons license. For more information, see http://ocw.mit.edu/help/faq-fair-use/.

10

Page 11: Game Theory for Humans - MIT OpenCourseWare · Game Theory for Humans Matt Hawrilenko MIT: Poker Theory and Analytics 1

Screenshot of Matt Mattusow © ESPN. All rights reserved. Thiscontent is excluded from our Creative Commons license. Formore information, see http://ocw.mit.edu/help/faq-fair-use/. 12

Page 12: Game Theory for Humans - MIT OpenCourseWare · Game Theory for Humans Matt Hawrilenko MIT: Poker Theory and Analytics 1

The trouble is, the other side can do magic too

-Cornelius Fudge, Minister for Magic

14

Page 13: Game Theory for Humans - MIT OpenCourseWare · Game Theory for Humans Matt Hawrilenko MIT: Poker Theory and Analytics 1

RULE #1

1. Forget her hand1 (a). Forget her range

15

Page 14: Game Theory for Humans - MIT OpenCourseWare · Game Theory for Humans Matt Hawrilenko MIT: Poker Theory and Analytics 1

1,000,000,000,000,000,000

Pump up your puny strategy with toy

games

*Apologies to John Nash and to the tank top I ripped off16

Page 15: Game Theory for Humans - MIT OpenCourseWare · Game Theory for Humans Matt Hawrilenko MIT: Poker Theory and Analytics 1

The Clairvoyance Game What would you do if you lived in a world where you

always knew your opponent’s hand . . .

Film still of Harry Potter and Ron Weasley staring into a crystal ballduring Divination class. Image removed due to copyright restrictions.

. . . And he knew that you knew?

17

Page 16: Game Theory for Humans - MIT OpenCourseWare · Game Theory for Humans Matt Hawrilenko MIT: Poker Theory and Analytics 1

Coin Flip Clairvoyance 1. Each player antes $1.2. You flip a coin. Heads, you win. Tails, your opponent wins.

HOWEVER

3. Only you see the coin after the flip, then you can bet.4. You choose to bet $1 or check. Your opponent can only

check/call or fold.

18

Page 17: Game Theory for Humans - MIT OpenCourseWare · Game Theory for Humans Matt Hawrilenko MIT: Poker Theory and Analytics 1

Scenario 1 Scenario 2

Images are in the public domain.

19

Page 18: Game Theory for Humans - MIT OpenCourseWare · Game Theory for Humans Matt Hawrilenko MIT: Poker Theory and Analytics 1

Two Questions 1. How often should she call?2. How often should you bluff?

20

Page 19: Game Theory for Humans - MIT OpenCourseWare · Game Theory for Humans Matt Hawrilenko MIT: Poker Theory and Analytics 1

How often to call? Enough to make opponent indifferent to bluffing or giving up

E(Bluffing) = E(Giving Up)

Pot*(1 - % Call) = (amount bluffed)*(% Call)

P (1 - C) = 1*C Amount won by bluffing

Amount lost by bluffing C = P / (P + 1)

21

Page 20: Game Theory for Humans - MIT OpenCourseWare · Game Theory for Humans Matt Hawrilenko MIT: Poker Theory and Analytics 1

How often to bluff? Enough to make opponent indifferent to calling or folding

E(calling) = E(folding)

(Ratio bluffs/value bets) (pot + 1) – value bets = 0

b (P+1) – 1 = 0

b = 1 / (P + 1)

22

Page 21: Game Theory for Humans - MIT OpenCourseWare · Game Theory for Humans Matt Hawrilenko MIT: Poker Theory and Analytics 1

Generalizes to variable bet sizes

Calling % = 1 / (1 + S)

Bluff ratio = S / (1 + S)

S = proportion of the pot bet

23

Page 22: Game Theory for Humans - MIT OpenCourseWare · Game Theory for Humans Matt Hawrilenko MIT: Poker Theory and Analytics 1

Q: What if it’s not a repeated game?

A: It’s a repeated game.

24

Page 23: Game Theory for Humans - MIT OpenCourseWare · Game Theory for Humans Matt Hawrilenko MIT: Poker Theory and Analytics 1

Observations It’s not about value-betting or bluffing, it’s about the

combination of the two

We’re trying to maximize the value of our entire set of hands,not just the hand we’re currently playing

25

Page 24: Game Theory for Humans - MIT OpenCourseWare · Game Theory for Humans Matt Hawrilenko MIT: Poker Theory and Analytics 1

The Ace-King-Queen Game

RULES: 1. Each player antes $1 and is dealt 1 card2. Player 1 can check or bet3. Player 2 can only check/call, or fold

26

Page 25: Game Theory for Humans - MIT OpenCourseWare · Game Theory for Humans Matt Hawrilenko MIT: Poker Theory and Analytics 1

Case 1 YOU OPPONENT

CHECK OR BET?

27

Page 26: Game Theory for Humans - MIT OpenCourseWare · Game Theory for Humans Matt Hawrilenko MIT: Poker Theory and Analytics 1

Case 2 OPPONENT YOU

CALL OR FOLD?

28

Page 27: Game Theory for Humans - MIT OpenCourseWare · Game Theory for Humans Matt Hawrilenko MIT: Poker Theory and Analytics 1

Case 3 OPPONENT YOU

CALL OR FOLD?

29

Page 28: Game Theory for Humans - MIT OpenCourseWare · Game Theory for Humans Matt Hawrilenko MIT: Poker Theory and Analytics 1

Case 4 YOU OPPONENT

CHECK OR BET?

30

Page 29: Game Theory for Humans - MIT OpenCourseWare · Game Theory for Humans Matt Hawrilenko MIT: Poker Theory and Analytics 1

Case 5 YOU OPPONENT

CHECK OR BET?

31

Page 30: Game Theory for Humans - MIT OpenCourseWare · Game Theory for Humans Matt Hawrilenko MIT: Poker Theory and Analytics 1

Case 6 OPPONENT YOU

CALL OR FOLD?

32

Page 31: Game Theory for Humans - MIT OpenCourseWare · Game Theory for Humans Matt Hawrilenko MIT: Poker Theory and Analytics 1

How often to call?

Calling ratio = 1 / ( 1 + 0.5) = 2/3 of hands that beat a bluff

beats a bluff

Aces represent 50% of hands that beat a bluff

All aces + 1/3 of Kings = 2/3 of hands that beat a bluff

50%

16.67%

33

Page 32: Game Theory for Humans - MIT OpenCourseWare · Game Theory for Humans Matt Hawrilenko MIT: Poker Theory and Analytics 1

Observations It’s not about value-betting or bluffing, it’s about the combination

of the two

We’re trying to maximize the value of our entire set of hands,not just the hand we’re currently playing

Useful to map hands as value, bluff catchers, and bluffs

You strategy for one hand determines your strategy for other hands

34

Page 33: Game Theory for Humans - MIT OpenCourseWare · Game Theory for Humans Matt Hawrilenko MIT: Poker Theory and Analytics 1

1. Know Thyself

2. Nothing in Excess

3. Make a Pledge andMischief is Nigh

Temple of Apollo at Delphi where people would go to visitthe oracle. Image courtesy of Pilar Torres on Flickr.License: CC BY-NC-SA.

35

Page 34: Game Theory for Humans - MIT OpenCourseWare · Game Theory for Humans Matt Hawrilenko MIT: Poker Theory and Analytics 1

$310,000 $720,000

$160,000$119,000

? 36

Page 35: Game Theory for Humans - MIT OpenCourseWare · Game Theory for Humans Matt Hawrilenko MIT: Poker Theory and Analytics 1

Three (exploitive) Strategies 1. My hand vs. your hand2. My hand vs. your distribution

Distribution: the frequency distribution of hands a player might hold, given all the action that has occurred

3. My distribution vs. your distribution

$720,000 ? 37

Page 36: Game Theory for Humans - MIT OpenCourseWare · Game Theory for Humans Matt Hawrilenko MIT: Poker Theory and Analytics 1

Mapping the AKQ Game + + + E.V. at

showdown

Call/Raise % 1/(1+S)

Bluff-to-Value Ratio S/(1+S)

- - - E.V. at showdown

99%

1%

Value Betting

Beats a Bluff

38

Page 37: Game Theory for Humans - MIT OpenCourseWare · Game Theory for Humans Matt Hawrilenko MIT: Poker Theory and Analytics 1

Reading Your Own Hand What you do with one hand depends on what you’d do with

your other hands

Most important skill in poker

Two updates for each street: Account for card removal Account for your action

39

Page 38: Game Theory for Humans - MIT OpenCourseWare · Game Theory for Humans Matt Hawrilenko MIT: Poker Theory and Analytics 1

Hand Combos Hand Combos pairs 22 6 no gap AK 16Preflop 33 6 KQ 16

44 6 QJ 16Distribution 55 6 JT 16 66 6 one gap 86s 4 77 6 97s 4 88 6 T8s 4 99 6 J9s 4 TT 6 QT 16 JJ 6 KJ 16

QQ 6 AQ 16 KK 6 2 gaps KT 16 AA 6 AJ 16

suited conn T9s 4 3 gaps K9s 4 98s 4 AT 16 87s 4 A2s-A9s 32 76s 4 65s 4 total 310

$119,000

40

Page 39: Game Theory for Humans - MIT OpenCourseWare · Game Theory for Humans Matt Hawrilenko MIT: Poker Theory and Analytics 1

Flop Card Removal Update

Hand Combos 22 6 33 6 44 6 55 6 66 6 77 6 88 3 99 6 TT 6 JJ 3

QQ 6 KK 3 AA 6 T9s 4 98s 3 87s 3 76s 4 65s 4

Hand Combos AK 12 KQ 12 QJ 12 JT 12

86s 3 97s 4 T8s 3 J9s 3 QT 16 KJ 9 AQ 16 KT 12 AJ 12

K9s 3 AT 16

A2s-A9s 31

total 263

$160,000

41

Page 40: Game Theory for Humans - MIT OpenCourseWare · Game Theory for Humans Matt Hawrilenko MIT: Poker Theory and Analytics 1

Flop Action Update

Villain checks, hero bets 75,000

Hand Combos 22 6 33 6 44 6 55 6 66 0 77 0 88 3 99 0 TT 0 JJ 3

QQ 6 KK 3 AA 6 T9s 4 98s 0 87s 0 76s 4 65s 4

Hand Combos AK 12 KQ 12 QJ 12 JT 12

86s 0 97s 0 T8s 0 J9s 3 QT 16 KJ 9 AQ 0 KT 12 AJ 12

K9s 3 AT 0

A2-A9s 0 total 160

$160,000

42

Page 41: Game Theory for Humans - MIT OpenCourseWare · Game Theory for Humans Matt Hawrilenko MIT: Poker Theory and Analytics 1

Turn Card Removal Update

Hand Combos 22 6 33 6 44 6 55 3 88 3 JJ 3

QQ 6 KK 3 AA 6 T9s 4 76s 4 65s 3

Hand Combos AK 12 KQ 12 QJ 12 JT 12 J9s 3 QT 16 KJ 9 KT 12 AJ 12

K9s 3

total 156

Villain checks, hero bets 205,000,

Villain calls

$310,000

43

Page 42: Game Theory for Humans - MIT OpenCourseWare · Game Theory for Humans Matt Hawrilenko MIT: Poker Theory and Analytics 1

Turn Action Update

Hand Combos 22 0 33 0 44 0 55 3 88 3 JJ 3

QQ 6 KK 3 AA 6 T9s 4 76s 4 65s 0

Hand Combos AK 12 KQ 12 QJ 12 JT 0 J9s 0 QT 16 KJ 9 KT 12 AJ 12

K9s 3

total 120

Villain checks, hero bets 205,000,

Villain calls

$310,000

How’s our proportion of bluffs here?

44

Page 43: Game Theory for Humans - MIT OpenCourseWare · Game Theory for Humans Matt Hawrilenko MIT: Poker Theory and Analytics 1

Hand Combos

88 3River Card 55 3

Removal Update JJ 3 QQ 6 KK 1 AA 6 T9s 4 76s 4 AK 8 KQ 8 QJ 12 QT 16 KJ 6 KT 8 AJ 12

K9s 2 Total 102

$720,000

45

Page 44: Game Theory for Humans - MIT OpenCourseWare · Game Theory for Humans Matt Hawrilenko MIT: Poker Theory and Analytics 1

River Decision

S = 1,080,000 / 720,000 = 1.5

Call = 1.5 / (1 + 1.5) = 40%

of hands that beat a bluff

$720,000 ? 46

Page 45: Game Theory for Humans - MIT OpenCourseWare · Game Theory for Humans Matt Hawrilenko MIT: Poker Theory and Analytics 1

Call 1/(1+S)

Know Thyself!

Nothing in

excess!

KK 1% KJ 9% JJ 13% 88 17% 55 21% AK 31% KQ 41% KT 51% K9s 54% AA 62% QQ 69% AJ 85% QJ 100% QT T9s 76s

beats a bluff

47

Page 46: Game Theory for Humans - MIT OpenCourseWare · Game Theory for Humans Matt Hawrilenko MIT: Poker Theory and Analytics 1

Recap Solved! Fold AA, Even fold KT

Gut check: Do we want a distribution where we have to fold

trips?

48

Page 47: Game Theory for Humans - MIT OpenCourseWare · Game Theory for Humans Matt Hawrilenko MIT: Poker Theory and Analytics 1

1% 9%

13% 17% 21% 31% 41% 51% 54% 62% 69% 85%

100%

KK KJ JJ 88 55 AK KQ KT K9s AA QQ AJ QJ QT T9s 76s

Rule of thumb: If you’d bet it for value, you

want a distribution where you don’t have to fold it

49

Page 48: Game Theory for Humans - MIT OpenCourseWare · Game Theory for Humans Matt Hawrilenko MIT: Poker Theory and Analytics 1

KK KJ

We can add some JJ 88 Call

hands in . . . 55 1/(1+S) AK KQ KT K9s AA QQ AJ QJ JT

Beats a bluff QT T9s 76s

1% 8%

11% 14% 17% 26% 34% 43% 45% 52% 58% 71% 84% 97%

J9s 100%

50

Page 49: Game Theory for Humans - MIT OpenCourseWare · Game Theory for Humans Matt Hawrilenko MIT: Poker Theory and Analytics 1

KK KJ

We can also 88JJ

construct a 55 Call AK 1/(1+S) distribution where KQ

we call with AA KT K9s AA QQ AJ QJ JT

J9o, J8o, J7s beats a bluff QT

T9s 76s

1% 6% 9%

11% 14% 21% 27% 34% 36% 41% 46% 56% 67% 77%

100%

51

Page 50: Game Theory for Humans - MIT OpenCourseWare · Game Theory for Humans Matt Hawrilenko MIT: Poker Theory and Analytics 1

RYOH Redux Check for balance on all streets

Don’t overthink it: focus on the glaring errors

Don’t needlessly bifurcate your distribution

Identify situations where you tend to become imbalanced, then watch opponents for the same tendency

52

Page 51: Game Theory for Humans - MIT OpenCourseWare · Game Theory for Humans Matt Hawrilenko MIT: Poker Theory and Analytics 1

INSINCERE APOLOGY

+

BRIEF MONOLOGUE

53

Page 52: Game Theory for Humans - MIT OpenCourseWare · Game Theory for Humans Matt Hawrilenko MIT: Poker Theory and Analytics 1

Exploitive Play: Foundations

99% + + + E.V. at

showdown

- - - E.V. at showdown

Beats a Bluff

Value Betting

If they bluff too much, expand the

marginal calls

If they fold too much, expand the marginal

bluffs and contract the marginal bets

If they don’t value bet enough, contract

the marginal calls

If they call too much, expand the marginal

bets and contract the marginal bluffs

1%

54

Page 53: Game Theory for Humans - MIT OpenCourseWare · Game Theory for Humans Matt Hawrilenko MIT: Poker Theory and Analytics 1

With our Example Hand

Call 1/(1+S)

Beats a Bluff

KK 1% KJ 8% JJ 11% 88 14% 55 17% AK 26% KQ 34% KT 43% K9s 45% AA 52% QQ 58% AJ 71% QJ 84% JT 97% J9s 100% QT T9s 76s

55

Page 54: Game Theory for Humans - MIT OpenCourseWare · Game Theory for Humans Matt Hawrilenko MIT: Poker Theory and Analytics 1

With my Shamefully Exploitive Hand

+ + + E.V. atshowdown

42 o

50% vs. random

33% vs. random

Know Thyself! Bet

Fold

56

Page 55: Game Theory for Humans - MIT OpenCourseWare · Game Theory for Humans Matt Hawrilenko MIT: Poker Theory and Analytics 1

Advanced Exploitive Play

Don’t forget about this part of the equation!

Bluff-to-Value Ratio = S / ( 1 + S )

57

Page 56: Game Theory for Humans - MIT OpenCourseWare · Game Theory for Humans Matt Hawrilenko MIT: Poker Theory and Analytics 1

Four Principles

1. Know thyself2. Nothing in excess3. Mischief4. Exploit at the margins

Screenshot of Matt Mattusow © ESPN. All rights reserved. Thiscontent is excluded from our Creative Commons license. Formore information, see http://ocw.mit.edu/help/faq-fair-use/.

60

Page 58: Game Theory for Humans - MIT OpenCourseWare · Game Theory for Humans Matt Hawrilenko MIT: Poker Theory and Analytics 1

MIT OpenCourseWarehttp://ocw.mit.edu

15.S50 Poker Theory and AnalyticsJanuary IAP 2015

For information about citing these materials or our Terms of Use, visit: http://ocw.mit.edu/terms.