Parkes CS 700 1 Parkes CS 700 CS 700 Computational Mechanism Design 2 Motivation Mechanism design what is the power of the market David C Parkes Internet both a market and enabler of markets design is not just for the few parkes eecs harvard edu Wireless commerce A Platform for AI Government policy Elections Lecture 1 Grid and cloud computing Fall 2008 EPFL Parkes CS 700 3 Parkes CS 700 4 Interdisciplinary Fast emerging field Microconomics multiple agents self interest game theoretic rationality often descriptive EconCS Economics Computer Science bidirectional Tenth ACM Electronic Commerce Conference June 2009 Computer Science computational and informational constraints trad one user or one user vs adversarial opponents often constructive Workshop on Internet and Network Economics December 2008 European SAGT 08 COMSOC 08 IJCAI AAAI AAMAS STOC FOCS INFORMS SODA Operations Research optimization deterministic and stochastic Yahoo research Google Labs Microsoft New England Lab CMD resolve tension between incentive and computational constraints new domains develop theories and computational methods Parkes CS 700 5 Parkes CS 700 6 Internet Advertising Example Slot auction Sponsored search Google AdSense Yahoo Panama Microsoft AdCenter Google 6B revenue 2005 10B revenue 2006 more 2007 Big and small advertisers Ecosystem of SEOs 2 slots to allocate 3 bidders Bids price per click 10 6 1 Quality Pr Click Impression 0 05 0 1 0 05 Outcome slot 1 A2 slot 2 A1 charge 5 1 per click Content monetization Keyword based affiliate networks Banner ads also sold through human negotiation Rank by revenue Second price online Offline split roughly 16B to 300B 2006 Online projected to grow 30 year Parkes CS 700 7 Parkes CS 700 8 Sponsored Search History GoTo selling search results in March 98 rank by bid first price Ingredients of CMD Became Overture switch to second price in 2002 acquired Yahoo in July 03 Algorithmic problem e g which advert to allocate to which slot Google rank by revenue and second price in feb 02 Incentive problem e g how should an advertiser bid Microsoft AdCenter introduced in June 06 Preference elicitation problem e g what is appropriate expressiveness of a bid Yahoo rank by revenue in Dec 07 Currently Google may serve some of Yahoo s ads emergence of exchanges e g RightMedia Parkes CS 700 9 Parkes CS 700 10 Outline of Course Introductory lectures until Oct 7 3 in total GT MD auction theory notes posted on WWW accompanied by 3 short homeworks Required Background Level of mathematical sophistication linear algebra complexity theory basic algorithms Class Discussion 9 classes 2 papers per class posted on WWW students will prepare summary everyone will read and submit comments Familiarity with rationality concepts e g an AI course or a microeconomics course Helpful Linear programming game theory microeconomics Final Project can be in pairs theory computational or survey short presentation on Dec 16 Parkes CS 700 11 Parkes CS 700 12 Grading Plan Problem sets 25 Projects 3 problem sets on introductory material Participation 20 Submitting comments Presentation of 20 short survey and critique Goal Develop a deep understanding of an important open research area and to the extent possible to work on an open research problem participation in discussion research papers Project 35 with a partner Projects may be theoretical or experimental Proposal class presentation Proposals due Nov 25 Presnetation Dec 16 final report can work in pairs if you use a reference then give a citation Parkes CS 700 13 Parkes CS 700 Today s Lecture Background reading Intro to Game theory strategic form games incomplete information games Mechanism Design for Computerized Agents by H Varian In Proc USENIX Workshop on Electronic Commerce July 11 12 1995 New York Intro to Mechanism design indirect and direct revelation incentive compatibility single peaked Groves VCG Sponsored Search A Brief History by D C Fain and J O Pedersen In Proc Second Workshop on Sponsored Search Auctions ACM EC 06 Putting Auction Theory to Work by P Milgrom Chapter 1 Cambridge University Press 2004 A Combinatorial Auction Mechanism for Airport Time Slot Allocation by S J Rassenti V L Smith and R L Bulfin in The Bell Journal of Economics Vol 13 No 2 Autumn 1982 pp 402 417 Introduction to Noncooperative Game Theory by Y Shoham and K Leyton Brown Chapter 3 in Multiagent Systems MAS Cambridge University Press to appear October 2008 Scientific Background for 2007 Nobel In Economic Sciences for mechanism design theory Mechanism Theory M O Jackson Chapter in Encyclopedia of Life Support Stystems U Derigs ed 2003 Parkes CS 700 14 15 Parkes CS 700 16 Historical Perspective Preliminaries Nash 1950 General definition of equilibrium for a large class of games proof of existence the analytical structure or studying all situations of conflict and cooperation Agents N 1 n strategy si Si Def A strategy si is a complete contingent plan defines an action in all states of the world Myerson 99 Vickrey 61 birth of auction theory Nobel 96 later Myerson applied game theory to study optimal auction design Strategy profile s s1 sn determines the outcome of the game Def The utility function ui S1 Sn R FCC Spectrum Auction 100B by end of 2001 mandate was to put license in hands of those that value it the most Model agents as rational expected utility maximizers see p 49 54 in MAS for axiomatic justification Def A strategic form or normal form game G N Si ui In grad micro textbooks from the early 90s now used to design multi billion dollar markets and within CS to study networked systems Parkes CS 700 17 Parkes CS 700 18 Example 1 Prisoner s Dilemma Two people are arrested for a crime Cooperate Confess Defect Deny also TCP game C D C 1 1 4 0 D 0 4 3 3 Pareto Optimality Def Strategy profile s Pareto dominates s if ui s ui s for all i and there exists some agent i for which ui s ui s Def Strategy profile s is Pareto optimal if no other s Pareto dominates s Def Strategy profile s is a weak dominant strategy equilibrium if for all i ui s i s i ui si s i si Si s i S i Notation s i s1 si 1 si 1 sn Parkes CS 700 19 Parkes CS 700 Example 3 Example 2 L M R U 0 4 4 0 5 3 3 6 M 4 0 0 4 5 3 2 8 B 3 5 3 5 6 6 L M R U 4 3 5 1 6 2 M 2 1 8 4 D 3 0 9 6 Reading this player 1 is the row player player 2 is the column player entry x y denotes payoff x to player 1 and y to player 2 20 Def Strategy profile s is a Nash equilibrium of a game if for all i ui s i s i ui si s i si Si Parkes CS 700