19
9/2/10 1 CS286r: Topics at the Interface between Computer Science and Economics Fall 2010 Information, Prediction, and Collective Intelligence Yiling Chen [email protected] Course website: hHp://www.eecs.harvard.edu/cs286r/ Computer Science is the study and the science of the theoreIcal foundaIons of informaIon and computaIon and their implementaIon and applicaIon in computer systems. [Wikipedia] Building systems F(x) =? How fast can we get the answer? Focus on computaIonal and informaIonal constrains. 9/1/10 2 CS 286r Fall'10

Computer Science9/2/10 2 Economics • is the social science that studies the producon, distribuon, and consumpon of goods and services. [Wikipedia]

  • Upload
    others

  • View
    0

  • Download
    0

Embed Size (px)

Citation preview

Page 1: Computer Science9/2/10 2 Economics • is the social science that studies the producon, distribuon, and consumpon of goods and services. [Wikipedia]

9/2/10 

CS286r: Topics at the Interface between  Computer Science and Economics 

Fall 2010 

Information, Prediction, and Collective Intelligence

Yiling Chen [email protected] 

Course website: hHp://www.eecs.harvard.edu/cs286r/ 

Computer Science 

•  is  the  study  and  the  science  of  the  theoreIcal foundaIons of informaIon and computaIon and their  implementaIon  and  applicaIon  in computer systems. [Wikipedia] 

–  Building systems 

–  F(x) =? – How fast can we get the answer? –  Focus on computaIonal and informaIonal constrains. 

9/1/10  2 CS 286r Fall'10 

Page 2: Computer Science9/2/10 2 Economics • is the social science that studies the producon, distribuon, and consumpon of goods and services. [Wikipedia]

9/2/10 

Economics 

•  is the social science that studies the producIon, distribuIon, and consumpIon of goods and services. [Wikipedia] 

–  Economies (systems) – Many self‐interested agents – Agents’ preferences/uIliIes over outcomes – Agents’ informaIon and beliefs – Agents’ decision making – Game‐theoreIc interacIons of agents 

9/1/10  3 CS 286r Fall'10 

The Interface 

•  Computer systems are increasingly being developed and used by mulIple parIes with different preferences –  Predict system outcomes –  Design systems to achieve desired outcomes 

•  Economic problems someImes are (hard) computaIonal problems –  Resource allocaIon –  Price discovery 

Theories, algorithms, and systems that satisfy both economic and computational constraints.

9/1/10  4 CS 286r Fall'10 

Page 3: Computer Science9/2/10 2 Economics • is the social science that studies the producon, distribuon, and consumpon of goods and services. [Wikipedia]

9/2/10 

Lots of Compelling ApplicaIons 

9/1/10  5 

•  Internet MoneIzaIon:  Google, Yahoo!, Microsof are using aucIons to sell ads 

•  Markets are used for informaIon aggregaIon –  Google, Yahoo!, Microsof, GE, etc. have internal predicIon 

markets 

•  Social network and Social Tagging:  Facebook, MySpace, LinkedIn, Flickr, LibraryThing 

 … 

CS 286r Fall'10 

This Course 

•  RotaIng topic course •  Previous 

–  Fall 2009. Assignment, Matching, and Dynamics –  Fall 2008. Social CompuIng –  Spring 2008. ComputaIonal Finance –  Spring 2007. ComputaIonal Mechanism Design –  Spring 2006. MulI‐agent Learning and ImplementaIon 

…  

•  Seminar style 

9/1/10  6 CS 286r Fall'10 

Page 4: Computer Science9/2/10 2 Economics • is the social science that studies the producon, distribuon, and consumpon of goods and services. [Wikipedia]

9/2/10 

Course Goals 

•  Provide an introducIon to an emerging, interdisciplinary literature  

•  Develop a level of comfort with both economic and computaIonal thinking  

•  Develop general skills related to reading papers, idenIfying research quesIons  

•  Provide a basis for conInued research. 

9/1/10  7 CS 286r Fall'10 

Fall 2010 

•  InformaIon, PredicIon, and CollecIve Intelligence 

•  Algorithmic, game theoreIc, and conceptual quesIons related to obtaining informaIon, making predicIons, and gejng tasks done by the crowds.  

•  Focus on elici%ng and aggrega%ng probabilis%c informa%on, bridging from economic theory to theory of online learning.  

9/1/10  8 CS 286r Fall'10 

Page 5: Computer Science9/2/10 2 Economics • is the social science that studies the producon, distribuon, and consumpon of goods and services. [Wikipedia]

9/2/10 

Crowds Are Smarter… 

•  Who wants to be a millionaire? – Fify‐Fify   Correct 50% of the Ime 

– Phone‐A‐Friend  Correct 65% of the Ime 

– Ask the Audience  Correct 91% of the Ime 

9/1/10  9 CS 286r Fall'10 

Crowds Are Smarter… •  Jelly‐Beans‐in‐the‐Jar Experiment 

– Professor Jack Treynor ran the experiment in his class  

– with a jar that held 850 beans  –  the group esImate was 871  – only one of the 56 people in the class made a beHer guess 

9/1/10  10 CS 286r Fall'10 

Page 6: Computer Science9/2/10 2 Economics • is the social science that studies the producon, distribuon, and consumpon of goods and services. [Wikipedia]

9/2/10 

Are Crowds Smarter? 

•  No always – Bad commiHee decisions 

– Endless group meeIngs 

•  In this course, we focus on mechanisms that intend to make crowds smarter.  

9/1/10  11 CS 286r Fall'10 

Structure of the Course 

•  Introductory lectures  (5 lectures) –  This one, informaIon theory and Kelly criterion, and game theory 

•  Research Papers  –  IncenIvizing experts   –  Peer predicIon       –  PredicIon markets     – Online learning – Analysis of exisIng collecIve intelligence systems        

9/1/10  12 CS 286r Fall'10 

Page 7: Computer Science9/2/10 2 Economics • is the social science that studies the producon, distribuon, and consumpon of goods and services. [Wikipedia]

9/2/10 

Enrollment & Prerequisites •  Enrollment is limited to about 20 students. Complete Survey at end of class! 

•  Prerequisites – Math background is important! At least a basic course in linear algebra (such as M 21b, AM 21b, or equivalent)   

–  A course on probabiliIes and staIsIcs (STAT 110 or equivalent) 

–  An algorithm course (CS 124, or equivalent) –  Familiarity with the concept of raIonality. An AI course or an economics/game theory course.  

 Advanced course in algorithms, microeconomics, game theory, or linear programming are helpful but not required.  

9/1/10  13 CS 286r Fall'10 

Grading 

Problem sets  25%  2‐3 homework problem sets 

ParIcipaIon  25%  Reading papers, submijng short summaries and quesIons before class, and parIcipaIon in class discussion. (Note: Absent students rarely contribute to discussions.) 

PresentaIon of one or two research papers 

15%  A short survey and criIque of the papers. Lead class discussion. 

Project  35%  Project proposal, class presentaIon, and final report. 

9/1/10  14 CS 286r Fall'10 

Page 8: Computer Science9/2/10 2 Economics • is the social science that studies the producon, distribuon, and consumpon of goods and services. [Wikipedia]

9/2/10 

Project 

•  Goal: develop a deep understanding of a specific research area and to the extend possible to work on an open research problem.  

•  Can be theoreIcal, computaIonal, experimental, or empirical. 

•  Can write an exposiIon paper, but needs novelty! •  A list of high‐level project topics will be provided. You are 

encouraged to propose your own topic for approval! •  TentaIve project due dates: 

–  Wednesday 11/3: project proposal due –  Monday 12/6: brief project presentaIon –  Friday 12/10: project report due 

9/1/10  15 CS 286r Fall'10 

LogisIcs 

•  TF – Alice Gao 

•  Office Hours –  Yiling: Monday 2:30 – 3:30, MD 339 –  Later will add office hours likely on Thursdays to meet with students in advance of presenIng papers 

–  Today, 2:30 – 3:30 – Alice: Tue 2:30‐4, MD 242 

 Missed course materials from the TF 

9/1/10  16 CS 286r Fall'10 

Page 9: Computer Science9/2/10 2 Economics • is the social science that studies the producon, distribuon, and consumpon of goods and services. [Wikipedia]

9/2/10 

InformaIon, PredicIon, and CollecIve Intelligence 

•  MoIvaIng examples – Disease surveillance – Business forecasIng – MulI‐agent systems – Almost any decision making under uncertainty 

– Neulix Prize – DARPA Red Balloon Challenge 

9/1/10  17 CS 286r Fall'10 

Examples of Course Topics 

•  IncenIvizing experts   •  Peer predicIon •  PredicIon Markets 

•  Online learning •    

9/1/10  18 CS 286r Fall'10 

Page 10: Computer Science9/2/10 2 Economics • is the social science that studies the producon, distribuon, and consumpon of goods and services. [Wikipedia]

9/2/10 

10 

IncenIvize Experts 

•  Suppose I’d like to get informaIon about tomorrow’s weather (sunny or rainy?) 

•  How can I ensure that an expert will tell me his/her true probability assessment of the event?  

9/1/10  19 CS 286r Fall'10 

What If We Won’t Know the Outcome? 

•  Eg. CondiIonal events, subjecIve informaIon •  Surveys 

– Eg. How many hours per week you spent on assignments?  •  Less than 5 hours •  5‐10 hours •  10‐20 hours •  Above 20 hours 

9/1/10  20 

Peer PredicIon and Bayesian Truth Serum  CS 286r Fall'10 

Page 11: Computer Science9/2/10 2 Economics • is the social science that studies the producon, distribuon, and consumpon of goods and services. [Wikipedia]

9/2/10 

11 

Combining informaIon is hard! 

•  If we have mulIple experts, how can we combine their informaIon?  

•  Some impossibility results on combining probability distribuIons. – T(f1, f2, f3, …, fn) – External Bayesianity –  Independent of irrelevant alternaIves => dictatorship 

9/1/10  21 CS 286r Fall'10 

Bet = Credible Opinion •  Q: Will Obama win the PresidenIal elecIon? 

•  Bejng intermediaries –  Las Vegas, Wall Street, Beuair, Intrade,... 

9/1/10  22 

McCain will win the elec;on 

Info 

I bet $1000 Obama will win the elec;on. 

Info 

CS 286r Fall'10 

Page 12: Computer Science9/2/10 2 Economics • is the social science that studies the producon, distribuon, and consumpon of goods and services. [Wikipedia]

9/2/10 

12 

PredicIon Markets •  A predicIon market is a futures market (bejng intermediary) that is designed for informaIon aggregaIon and predicIon.  

•  Payoffs of the traded item is associated with outcomes of future events.  

9/1/10  23 

$1 if Obama Wins 

$0 Otherwise 

$1×Percentage of Vote Share That

Obama Wins

$1 if Patriots win

$0 Otherwise $f(x)

CS 286r Fall'10 

Example: Iowa Electronic Market 

Popular vote in 2008 presidenIal elecIon 

9/1/10  24 CS 286r Fall'10 

Page 13: Computer Science9/2/10 2 Economics • is the social science that studies the producon, distribuon, and consumpon of goods and services. [Wikipedia]

9/2/10 

13 

A Combinatorial Bejng Example 

•  251 outcomes, 2251 combinaIons 

•  Allow parIcipants to bet on logical formulas – Create contracts on the fly:   $1 if Ohio AND Florida OR New York, $0 otherwise 

– Specify buy price and quanIty •  ComputaIonally hard!  

9/1/10  25 CS 286r Fall'10 

Learning from Expert Advice

Page 14: Computer Science9/2/10 2 Economics • is the social science that studies the producon, distribuon, and consumpon of goods and services. [Wikipedia]

9/2/10 

14 

Learning from Expert Advice •  The algorithm maintains weights over N experts

w1,t w2,t

wN,t

Slide Source: J.W. Vaughan

Learning from Expert Advice •  The algorithm maintains weights over N experts

•  At each time step t, the algorithm… •  Observes the instantaneous loss li,t of each expert i

w1,t w2,t

wN,t

Slide Source: J.W. Vaughan

Page 15: Computer Science9/2/10 2 Economics • is the social science that studies the producon, distribuon, and consumpon of goods and services. [Wikipedia]

9/2/10 

15 

Learning from Expert Advice •  The algorithm maintains weights over N experts

•  At each time step t, the algorithm… •  Observes the instantaneous loss li,t of each expert i

w1,t w2,t

… l1,t = 0 l2,t = 1

wN,t

lN,t = 0.5

Slide Source: J.W. Vaughan

Learning from Expert Advice •  The algorithm maintains weights over N experts

•  At each time step t, the algorithm… •  Observes the instantaneous loss li,t of each expert i •  Receives its own instantaneous loss lA,t = Σi wi,t li,t

w1,t w2,t

… l1,t = 0 l2,t = 1

wN,t

lN,t = 0.5

Slide Source: J.W. Vaughan

Page 16: Computer Science9/2/10 2 Economics • is the social science that studies the producon, distribuon, and consumpon of goods and services. [Wikipedia]

9/2/10 

16 

Learning from Expert Advice •  The algorithm maintains weights over N experts

•  At each time step t, the algorithm… •  Observes the instantaneous loss li,t of each expert i •  Receives its own instantaneous loss lA,t = Σi wi,t li,t

•  Selects new weights wi,t+1

w1,t w2,t

… l1,t = 0 l2,t = 1

wN,t

lN,t = 0.5

Slide Source: J.W. Vaughan

Learning from Expert Advice •  The algorithm maintains weights over N experts

•  At each time step t, the algorithm… •  Observes the instantaneous loss li,t of each expert i •  Receives its own instantaneous loss lA,t = Σi wi,t li,t

•  Selects new weights wi,t+1

w1,t+1 w2,t+1 wN,t+1

… l1,t = 0 l2,t = 1 lN,t = 0.5

Slide Source: J.W. Vaughan

Page 17: Computer Science9/2/10 2 Economics • is the social science that studies the producon, distribuon, and consumpon of goods and services. [Wikipedia]

9/2/10 

17 

Learning from Expert Advice •  Classic result: There exist algorithms such that on any

(bounded) sequence of losses,

loss of the best performing expert

< O((T logN)1/2) algorithm’s cumulative loss

Slide Source: J.W. Vaughan

Learning from Expert Advice •  Classic result: There exist algorithms such that on any

(bounded) sequence of losses,

•  Holds even in a fully adversarial setting with no statistical assumptions about the sequence of losses

loss of the best performing expert

< O((T logN)1/2) algorithm’s cumulative loss

Slide Source: J.W. Vaughan

Page 18: Computer Science9/2/10 2 Economics • is the social science that studies the producon, distribuon, and consumpon of goods and services. [Wikipedia]

9/2/10 

18 

Learning from Expert Advice •  Classic result: There exist algorithms such that on any

(bounded) sequence of losses,

•  Holds even in a fully adversarial setting with no statistical assumptions about the sequence of losses

loss of the best performing expert

< O((T logN)1/2) algorithm’s cumulative loss

“regret”

Slide Source: J.W. Vaughan

Learning from Expert Advice •  Classic result: There exist algorithms such that on any

(bounded) sequence of losses,

•  Holds even in a fully adversarial setting with no statistical assumptions about the sequence of losses

•  Algorithm is said to have “no regret” since the average regret per trial approaches 0 as T grows

loss of the best performing expert

< O((T logN)1/2) algorithm’s cumulative loss

“regret”

Slide Source: J.W. Vaughan

Page 19: Computer Science9/2/10 2 Economics • is the social science that studies the producon, distribuon, and consumpon of goods and services. [Wikipedia]

9/2/10 

19 

For Wed. 9/8 •  Submit comments on Chapter 2 (2.1 – 2.8) of Elements of

Information Theory •  Reading is posted on the class schedule. •  Check course website later this week on how to submit

your comments •  What is unclear? What would you like to hear about in

class? What did you enjoy?

•  Please hand in the survey now.