CS 15-892 Foundations of Electronic Marketplaces

1. CS 15-892 Foundations of Electronic Marketplaces Tuomas Sandholm & John Dickerson Computer Science Department Carnegie Mellon University Course web page: www.cs.cmu.edu/~sandholm/cs15-892F15/cs15-892.htm

2. Example applications

3. Example markets with money

4. Combinatorial sourcing (live auctions & RFPs/RFQs)[Screen shots from Sandholm “Very-Large-Scale Generalized Combinatorial Multi-Attribute Auctions: Lessons from Conducting $60 Billion of Sourcing”] Bidder side Bid-taker (buyer) side

5. Spectrum auctions[Screen shots from 2015 AWS-3 auction: $45 billion, 341 rounds]

6. Incentive auction Need enough, but not all spectrum is equivalent Incentive Auction Spectrum Reverse Auction and Repacking Forward Auction Spectrum Spectrum Future Wi-Fi broadcasters Existing TV broadcasters Payment Payment Payment Minimize Stations currently allocated to channel 13 Current channel allocation statistics

7. Sponsored search

8. Display advertising

9. Example markets with “funny” money

10. Combinatorial course allocation[Screen shots from Budish et al. working paper 2015] Bidding Dynamic exclusions

11. Example markets withoutmoney

12. Patient 1 Patient 2 Kidney exchange Donor 1 Donor 2 Pair 1 Pair 2

13. Example applications… • B2B (business-to-business) • Sourcing • Also buying consortia (e.g., healthcare GPOs, Covisint, Trade-ranger) • IntercontinentalExchange, Inc. (acquired ChemConnect 6/2007) • B2C (business-to-consumer), e.g. goods, debt • C2C (consumer-to-consumer), e.g. eBay • Task and resource allocation in computer systems (networks, computational grids, storage systems…) • Electricity markets • Transportation exchanges • Stock markets • Collaborative filtering

14. Motivation

15. Perspectives • Leverage increasing computing (and communication) power to increase economic efficiency • E.g. running more complex mechanisms • Combinatorial auctions [S. ICE-98, IJCAI-99, AIJ-02, AIJ-03, AAAI-04, MgmtSci, …] • Expressive competition [S. Interfaces-06, IAAI-06, IJCAI-07, AI Mag-07, Market Design Handbook 13, …] • Expressive charity donations [Conitzer & S. EC-04, AIJ-11] • Voting ... • Capitalize on piles of info - unlike traditional market mechanisms • E.g. automated mechanism design (AMD) [Conitzer & S. UAI-02,…] • Revenue-maximizing combinatorial auctions? [Likhodedov & S. AAAI-04, AAAI-05, Operations Research 2016; Tang & S. IJCAI-11, AAMAS-12] • Sponsored search, display ads, TV ads, …

16. Automated negotiation systems • Agents search & make contracts • Through peer-to-peer negotiation or a mediated marketplace • Agents can be real-world parties or software agents that work on behalf of real-world parties • Increasingly important from a practical perspective

17. Fertile, timely, important research area • Deep theories from game-theory & CS merge • Started together in the 1940’s [Morgenstern & von Neumann] • There were a few decades of little interplay • Upswing of interplay in the last few years • In this setting the prescriptive power of game theory really comes into play • Market rules need to be explicitly specified • Software agents designed to act optimally, unlike humans • Computational capabilities can be quantitatively characterized, and prescriptions can be made about how the agents should use their computation optimally • Optimization has recently become scalable enough to make these things practical • Custom integer programs for clearing problems • Custom (e.g., convex) optimization for computing strategies • The applications change the world

18. This course • Covers • The most relevant classic results from game theory • The state-of-the-art through recent research papers • Many of them have not even been published yet • Covers • game-theoretic aspects • computational aspects • and most importantly, the intersection

19. Systems with self-interested agents (computational or human) • Mechanism (e.g., rules of an auction) specifies legal actions for each agent & how the outcome is determined as a function of the agents’ strategies • Strategy (e.g., bidding strategy) = Agent’s mapping from known history to action • Rational self-interested agent chooses its strategy to maximize its own expected utility given the mechanism => strategic analysis required for robustness => noncooperative game theory • But … computational complexity • In executing the mechanism • E.g. combinatorial auctions NP-complete & inapproximable to clear • In executing a strategy • How should computationally bounded agents play strategically? • Costly or limited computing [S. ICMAS-96, IJEC-00; Larson&S. AIJ-01, AAMAS-02, TARK-01, AGENTS-01 WS, EC-04, AAMAS-05…] • Optimal mechanism design for such agents? • (Voting) mechanisms that are hard to manipulate [Conitzer&S. AAAI-02, IJCAI-03, TARK-03, JACM-07, …] • No voting mechanism is usually hard to manipulate [Conitzer&S. AAAI-06, …] • What about multistage mechanisms and/or adding a bit of randomization in the mechanism? • Mechanisms that do better if the agents cannot solve their computational problems [Conitzer & S. LOFT-04; Othman & S. GAMES-08, COMSOC-08, SAGT-09]

20. How these techniques can/could play a role in different stages of an ecommerce transaction

21. Automated negotiation techniques in different ecommerce stages • 1. Interest generation (vendors compete for customers’ attention) • Sponsored search • Search keyword auctions (Google, Baidu, Yahoo!, Bing) • Bid optimization vendors • Display ad markets (Yahoo!, DoubleClick (now part of Google), Right Media (now part of Yahoo!), adECN (now part of Microsoft), Baidu, …) • Funded adlets that coordinate • Avatars for choosing which ads to read • Customer models for choosing who to send ads and how much $ to offer • 2. Finding • Simple early systems: BargainFinder, Jango • Meta-data, XML • Standardized feature lists on goods to allow comparison • How do these get (re)negotiated • Different vendors prefer different feature lists • Shopper agents need to understand the new lists • How do algorithms cope with new features? • Want to get a bundle => need to find many vendors

22. dynamic Pricing static nondiscriminatory discriminatory Automated negotiation techniques in different ecommerce stages... • 3. Negotiating • Advantages of dynamic pricing • Right things sold to (and bought from) right parties at right time • World becomes a better place (social welfare increases) • Further advantages from discriminatory pricing • Can increase social welfare (e.g., if production increases) • Fixed-menu take-it-or-leave-it offers -> negotiation • Cost of generating & disseminating catalogs? • Other customers see the price? • Negotiation overhead? • Personalized menus • Could check customer’s web page, links to & from it, what other similar customers did, customer profiles • Generating/printing the menu may be intractable, e.g. mortgages 530 • Negotiation can focus the generation, but vendor may bias prices & offerings based on path • Preferences over bundles • Coalition formation

23. Automated negotiation techniques in different ecommerce stages... • 4. Contract execution • Digital payment schemes • Safe exchange • Third party escrow companies (1- or 2-sided) • Sometimes an exchange can be carried out without enforcement by dividing it into chunks [Sandholm&Lesser IJCAI-95, Sandholm96,97, Sandholm&Ferrandon ICMAS-00, Sandholm&Wang AAAI-02] • 5. After sales

24. Agenthood, utility function, rationality & bounded rationality, evaluation criteria of multiagent systems

25. u i 1 Risk averse Risk neutral 0.5 Risk seeking 0 M$ 0 0.5 1 Agenthood • We use economic definition of agent as locus of self-interest • Could be implemented e.g. as several mobile “agents” … • Agent attempts to maximize its expected utility • Utility function ui of agent i is a mapping from outcomes to reals • Can be over a multi-dimensional outcome space • Incorporates agent’s risk attitude (allows quantitative tradeoffs) • E.g. outcomes over money Lottery 1: $0.5M w.p. 1 Lottery 2: $1M w.p. 0.5 $0 w.p. 0.5 Agent’s strategy is the choice of lottery Risk aversion => insurance companies

26. Agent i chooses a strategy that maximizes expected utility maxstrategySoutcome p(outcome | strategy) ui(outcome) If ui’() = a ui() + b for a > 0 then the agent will choose the same strategy under utility function ui’ as it would under ui (ui has to be finite for each possible outcome; otherwise expected utility could be infinite for several strategies, so the strategies could not be compared.) Utility functions are scale-invariant

27. Full vs bounded rationality Bounded rationality Full rationality Descriptive vs. prescriptive theories of bounded rationality

28. Criteria for evaluating multiagent systems • Computational efficiency • Distribution of computation • Communication efficiency • Social welfare: maxoutcome ∑i ui(outcome) • Requires cardinal utility comparison • … but we just said that utility functions are arbitrary in terms of scale! • Surplus: social welfare of outcome – social welfare of status quo • Constant sum games have 0 surplus. Markets are not constant sum • Pareto efficiency: An outcome o is Pareto efficient if there exists no other outcome o’ s.t. some agent has higher utility in o’ than in o and no agent has lower • Social welfare maximization => Pareto efficiency • Individual rationality: Participating in the negotiation (or individual deal) is no worse than not participating • Stability: No agents can increase their utility by changing their strategies • Symmetry: No agent should be inherently preferred, e.g. dictator • …