智能技术应 - Nanjing University › zhangl › ai10.pdf · 2019-04-30 · 神经络 •...

⼈人⼯工智能技术应⽤用计算博弈

�1

神经⽹网络

• 前向神经⽹网络

• 卷积神经⽹网络

• 循环神经⽹网络

• 递归神经⽹网络

• …

!2

• 过程

• 函数

• 循环

• 递归

神经⽹网络 程序设计

Compiler, Linker, Debugger, Testing…

Understanding Black Box Predictions via Influence Functions

Pang Wei Koh and Percy Liang. 2017. Understanding black-box predictions via influence functions. In Proceedings of the 34th International Conference on Machine Learning - Volume 70 (ICML'17), Doina Precup and Yee Whye Teh (Eds.), Vol. 70. JMLR.org 1885-1894.

!3

4

Free Services for Free Data

• The digital economy encompasses many of the most popular products ever. It creates surplus for end users and tech companies.

Data

Services

The High Cost of Free Data

• Despite these advances, there has been backlash against tech companies.

!6

An Anxious Time For Labor

(Autor et al. 2017)

Fears about future of AI and job displacement

Growing inequality

labor share has been declining since 1980s.

!7

AI and Productivity

• The “free-data” model has failed to induce the “AI-revolution” in productivity.

!8

The High Cost of Free Data

privacy, job displacement, and no large productivity gains

!9

The High Cost of Free Data

• Root of all of these problems is that returns to data are being treated as payments to capital rather than to labor.

!10

Data as Capital vs. Data as Labor

1. Data is exhaust to be picked up by whoever is smart enough to make useful

2. AI success mostly driven by algorithms, computational power, brilliant programmers

3. AI will flourish if there are big payoffs to entrepreneurs and innovators

4. AI will displace workers, so either they will work in other areas or we will give them a basic income now that they are useless

5. People will have to find meaning in areas of life other than work and “get over themselves”

6. Individuals are happy to consent to surveillance in exchange for “free stuff”, so companies deserve to use their data

1. Data created by user actions and efforts, so fruits belong first to individual contributors

2. AI is really collective human intelligence, driven by innate knowledge/data in ordinary people

3. AI will flourish if the individual contributors who make it possible have a real economic stake

4. AI is just another production technology that raises individual productivity by allowing people to amplify the value of their data

5. The value of data will support sense of meaning and “digital dignity”

6. Individuals on their own have little bargaining power in face of digital monopolies, must be protected by countervailing power (competition, unions or governments)

Arrieta Ibarra, Imanol and Goff, Leonard and Jiménez Hernández, Diego and Lanier, Jaron and Weyl, Eric Glen, Should We Treat Data as Labor? Moving Beyond 'Free' (December 27, 2017). American Economic Association Papers & Proceedings, Vol. 1, No. 1, Forthcoming.

11

Economics and computation

• Problems• at the interface of Economics and Computer science• can’t be tackled alone by economics or CS

• put economics to work (e.g., efficient execution of a mechanism)• "As far as the laws of mathematics refer to reality, they are not certain; and as far

as they are certain, they do not refer to reality.“ - Albert Einstein• design algorithms that take strategic inputs (incentive issues)

• encountered in internet economics• 互联⽹网 + 经济学

• at foundations of electronic commerce• A synthesis of • Artificial intelligence (MAS, optimization, learning)• Economics (game theory, mechanism design)• Algorithms and analysis (aka. AGT)

!12

Applications

• Examples• How to design online auctions?• Baidu, Google advertisement auctions• Facebook, Weibo advertisement auctions

• How to design algorithms for strategic inputs?• How to design Taobao reputation system? • How to design trust network?

• How to design matching and pricing?• Match drivers and passengers on Uber and DiDi

• Sharing economy, Community (Fans) economy

!13

Game theory

• Decision making by multiple, selfish agents• Single-agent decision = AI• Multi-agent decision = Game Theory

• Themes• How to formulate multi-agent interactions• Model

• How rational agents should (would) behave• Analysis (prediction)

• How to incentivize rational agents to well behave• Design

!14

Examples

• Equilibrium and commitment

• Selfish routine

• Matching

• Facility location game

• Auctions

!15

Prisoner’s Dilemma

-2, -2 0, -3

-3, 0 -1, -1

defect

• Two criminals has been caught• District attorney has evidence to convict them of a minor

crime (1 year in jail); knows that they committed a major crime together (3 years in jail) but cannot prove it

• Offers each of them a deal:– If both defect (背叛), they each get a 1 year reduction– If only one defects, that one gets 3 years reduction

cooperate

cooperate

defect

!16

What if one player can make a commitment?

-2, -2 0, -3

-3, 0 -1, -1

defect

• I promise that I will defect/cooperate• I have a reputation of being defective/cooperative

• What will I promise?cooperate

cooperate

defect

!17

What if the game is played repeatedly?

-2, -2 0, -3

-3, 0 -1, -1

defect

• How would you play?

cooperate

cooperate

defect

!18

A modified game

1, 0 3, 2

2, 1 4, 0

L

• Equilibrium play?

R

B

T

!19

A modified game

1, 0 3, 2

2, 1 4, 0

L

• I promise that I will play T/B• What will I promise?

R

B

T

!20

Matching pennies

1, -1 -1, 1

-1, 1 1, -1

H

• I promise that I will play H/T• What will I promise?

T

T

H

!21

P2P file sharing

• What peer-to-peer softwares have you used?

• Pros and cons?

!22

Client-server paradigm

• Download apps from Apple store

• Stream videos from Youtube

• Buy ebooks from Amazon

• Key features• Client and server are clearly separated• For large scale, lots of local servers needed, costly

!23

Peer to peer paradigm

• P2P network: server = client = peers

!24

Client-server vs. P2P

!25

!26

!27

!28

!29

Files shared by few hosts!

!30

!31

!32

Game theory?

• Another try: the Kazaa Client

• Client of the FastTrack protocol

• Peer maintains upload/download stats, shares stats with others to gain priority.

• The BitTorrent Network

• Approx 85% of P2P traffic in US (2016)

• BT used for illegal purposes (e.g., movies), but also legal purposes (e.g., Linux software distribution).

!33

Key innovation

• Break file into pieces: A repeated game!

• Allows for tit-for-tat like behavior

• “If you let me download, I’ll reciprocate.”

!34

#2 Selfish route

!35

#2 Selfish route

!36

#2 Selfish route

!37

Network flow

!38

Network flow

!39

Braess’ Paradox: Better network, worse outcome!

!40

#3. Matching: Stable marriage problem

• PlayersSet of men M, with man m in M Set of women W, with woman w in W|M|=|W|One-to-one matching: each man matched to one woman, and

vice-versa

• PreferencesEach man has strict preferences over women, and vice versa

!41

Matching

• A matching is a set of pairs (m,w) such that each individual has one partner.

• A matching is stable if• Every individual is matched with a partner.• There is no man-woman pair, each of whom would prefer to match

with each other rather than their assigned partner

• If such pair exist, the match is unstable

!42

Examples

• Two men m,m’ and two women w,w’

• Example 1

• m prefers w to w’ and m’ prefers w’ to w

• w prefers m to m’ and w’ prefers m’ to m

• Unique stable match: (m,w) and (m’,w’)• Example 2

• m prefers w to w’ and m’ prefers w’ to w

• w prefers m’ to m and w’ prefers m to m’

• Two stable matches {(m,w),(m’,w’)} and {(m,w’),(m’,w)}• First match is better for the men, second for the women.

• Is there always a stable match?

!43

The Gale-Shapley Algorithm

• Input: Men and women rank all potential partners• Algorithm• Each man proposes to highest woman on his list• Women make a “tentative match” based on their preferred offer,

and reject other offers• Each rejected man removes woman from his list, and makes a new

offer• Continue until no more rejections or offers, at which point finalize

tentative matches

• This is the man-proposing algorithm; there is also a “woman proposing” version.

!44

GS in pictures

!45

Properties of men-propose matching

• Stable• O(n2)• Worst for all women, among all stable matchings• Truthful for men, not truthful for women• Find an example where woman can benefit from lying

• Further interesting things about stable matching:• All stable matchings form a lattice• Men and women have strictly opposite preferences among the

stable matchings• Men get their best stable matching in men-proposing algo.

!46

Importance of stability

• Roth (2000) surveys 17 major matching markets, 10 of which are stable

• All stable markets survive over time

• For the remaining 7 unstable markets, 2 survive

!47

Application to Kidney Exchanges--Nobel Prize 2012

• More than 75,000 people in the United States are waiting to receive a kidney transplant.

• There is a shortage of donors• Deceased donors• Living donors• In 2005, 4200 patients died on the wait list, 10,000 in 2010

• Problem is not just straight supply and demand• Donor kidney needs to be compatible with the patient.• So sometimes patient has a living donor, but can’t use the kidney

because of incompatibility.• Maybe two patients could trade donor kidneys, or several patients

could engage in a kidney exchange.• Matching theory can also help us understand this problem, and make

optimal use of a limited pool of donors.

!48

#4. Facility location game –Where to build a library?

Single-peaked preference

!49

Minimizing social cost: Median mechanism

5

1

2

3

44

!50

Another objective: minimizing the maximum cost

• Algorithmic solution:• Given x=(x1,…xn)• Output: OPT(x)=(leftmost(x)+rightmost(x))/2• Not truthful:• x=(0,1), player 2 will misreport: 1 !2

• An algorithmic and game-theoretical solution:• Given x=(x1,…xn)• Output: leftmost(x)• Truthful• 2-approximation: for all x, no more than twice of OPT

!51

#5. Auctions: Dutch flower auction (Aalsmeer)

!52

Japan Bluefin Tuna Auction

•34usd (Spain) vs 46 usd (Japan)•Record price: 1.7 M USD, 220 KG•Fish market = 43 football courts

•Largest seafood market in the word

!53

# 5. 1st price and 2nd price auctions

• 1st price auction:• Everyone secretly bids a price, the highest bidder wins,

paying his bid

• 2nd price auction:• Everyone secretly bids a price, the highest bidder wins,

paying the 2nd highest bid

!54

Which one is better?

• Definitions of good:• Simple to play• Higher revenue for the seller• Higher utility for the buyer

!55

Proof of truthfulness in 2nd -price auction

!56

Which auction makes more money?

• 2nd price• Truthful• revenue=2nd highest value

• 1st price is not truthful• Bidder value drawn iid from a uniform [0,1]• Bidder with value v will bid (n-1)v/n• Revenue= (n-1)/n *highest value

• Fact:• E(highest among n~uniform [0,1])=n/(n+1)• E(2nd highest among n~uniform [0,1])=(n-1)/(n+1)

• Thus:• E(2nd highest value)=E((n-1)/n *highest value)• Expected revenue equivalence

!57

Failure of 2nd price auction

• Second price is arbitrarily far from optimal in the worst case

• Why?

!58

2nd price auction fails

• Example

• 1 strong buyer + several weak buyers

Claude and Paloma, Picasso, $ 28M

!59

Which auction is revenue-optimal then?

• Solved by Myerson, 1981• Nobel Prize 2007• n~uniform[0,1], second price auction with 0.5 reserve• In general:• Does not necessarily sell to the highest bidder• Does not necessarily sell the item at all

!60

Application: Selling advertisement

!61

总结

• Economics and computation

• Example

• Equilibrium and commitment

• Selfish routine

• Matching

• Facility location game

• Auctions

!62