evaluation through conflict martin zinkevich yahoo! inc

36
Evaluation Through Conflict Martin Zinkevich Yahoo! Inc. http://martin.zinkevich.org/lemonade

Upload: ebony-hellyer

Post on 15-Dec-2015

214 views

Category:

Documents


0 download

TRANSCRIPT

Evaluation Through Conflict

Martin ZinkevichYahoo! Inc.

http://martin.zinkevich.org/lemonade

Who was I

• Worked with U Alberta Computer Poker Research Group– Designed Counterfactual Regret Algorithm– Theory behind DIVAT

• Worked on AAAI Computer Poker Competition– 2006 as lead programmer, 2007 as chair

• Work used in Man Vs Machine

Who am I

• Run the Lemonade Stand Game Competition• Work with Yahoo Anti-Abuse Team

AAAI Computer Poker Competition

• 5 years running• Now the ANNUAL Computer Poker

Competition• Latest-11 universities et al

Competitions:Science vs Entertainment

AAAI Computer Poker Competition

May The Best Program Win!And Win Again IF WE PLAYED AGAIN!

Head to Head

VS

for 1000 hands

Head to Head

VS

for 1000 hands

All Combinations

7,-7 10,-10

-7,7 5,-5

-10,10 -5,5

OK, But Who Won?

• Online: Maximize total winnings• Equilibrium: Maximize number of people I can

win money from (or don’t lose against)

Why a New Competition?

ComputingEquilibria

✓Choosing Equilibria

?

Bach or Stravinsky

2,1 0,0

0,0 1,2

Big Question: How Do (or Would) People Get to Nash Equilibria?

Solvable Games

$

Unsolvable Games

$ ?

An Old Idea

• Think about learning in the presence of other intelligent agents.

• Prove cool stuff about your learning algorithm given:– constraints about the adversary– constraints about the game

Solving the Unsolvable

• In current competitions, people are often applying techniques that are effective in solvable games, even when the game is not solvable.

• In what competitions is it useless to approximate the game as solvable?

Axelrod’s Iterated Prisoner’s Dilemma

• A competition between many competitors.• One entry: tit-for-tat (Anatol Rapaport)

– Nice (initially)– Retaliating– Forgiving– Non-envious

• Learned that cooperation has value, but:– Cooperate with whom?– How do we cooperate?

The Lemonade Stand Game

The Lemonade Stand Game

The Lemonade Stand Game

What Is The Lemonade Stand Game?

• Every round for 100 rounds:– each person selects an action privately– then, the actions are revealed

• The score of a player is the distance clockwise to the next player plus the distance counterclockwise.

Key Observations• A constant-sum game between 3 players.

– For every gain, someone has to lose.• Possibilities For Cooperation

– Opposite sides of the circle, “sandwiching”• Not a “Solvable Game” (Nash, 1951)

– Playing equilibrium strategies is not advisable• Easy To Set “Table Image”

– The constant strategy often evokes cooperative behavior• Existing Techniques Fail

– Experts algorithms lose to constant strategy

Strategy #1: Play Constant

Strategy #2: Play Opposite

Strategy #3: Sandwich

Competition Structure

• Every set of three players played 100 rounds 180 times (1.5 million rounds total)

• Highest Total Score Wins• Mean, Standard Error can be calculated

Competitors

• 28 players, 9 teams– University of Southampton/Imperial College London

(Soton)– Yahoo! Inc. (Pujara)– Rutgers University (RL3)– Brown University (Brown)– Carnegie Mellon (2 teams-Waugh, ACTR)– University of Michigan (FrozenPontiac)– Princeton University (Schapire)– (Greg Kuhlmann)

Competition Results

0123456789

10

Soton

PujaraRL3

Waugh

ACTR

Schapire

Brown

Froze

nPontiac

Kuhlmann

Competitor

Scor

e Pe

r Rou

nd

Results

-1.2

-1

-0.8

-0.6

-0.4

-0.2

0

0.2

0.4

0.6

0.8

Soton

PujaraRL3

Waugh

ACTR

Schapire

Brown

Froze

nPontiac

Kuhlmann

Competitor

Scor

e Pe

r Rou

nd-8

Modified Constant Uniformly Random

Restricting to Top 6

-1.5

-1

-0.5

0

0.5

1

Pujara Soton RL3 Waugh ACTR Schapire

Competitor

Scor

e Pe

r Rou

nd-8

Restricting to Top 4

-1.5

-1

-0.5

0

0.5

1

Pujara RL3 Soton Waugh

Teach Simply!EQUILIBRIUM

FREE

=

Learn

=

=

= ?

Learn

=

=

10 7

The High Level

• Phenomenal Intelligence: the observed behavior used by a set of people at a point in time for some task.

Lofty Goals

• Phenomenal Intelligence: the observed behavior used by a set of people at a point in time for some task.

• behavior: a fully specified strategy.• used: actually leveraged

Practical Concessions

• Phenomenal Intelligence: the observed behavior used by a set of people at a point in time for some task.

• Not any intelligent agent• Not any time (people change)• Not any task (context matters)

Thank You

http://martin.zinkevich.org/lemonade