Pushing the Envelope: new research topics at the interface of cs and econ/gt

1
Pushing the Envelopenew research topics at the
interface of cs and econ/gt

• Yoav Shoham
• Stanford University
• (many debts are due)

2
Primary areas of interaction so far

• Computing solution concepts, primarily NE
• Multi-agent learning
• Compact games (graphical games, MAIDs, game
  networks, local-effect games, social networks, )
• Mechanism design, in particular auctions

3
Talk Outline

• Computing solution concepts, primarily NE
• The role of NE unclear
• Multi-agent learning
• Ditto
• Compact games (graphical games, MAIDs, game
  networks, local-effect games, social networks, )
• Other forms of compactness, and what about
  coalitional games?
• Mechanism design, in particular auctions
• Behavioral Mechanism design
• Beyond GT Algorithmic Institutional Design

4
A game with a trivial, unique NE
Rock Paper Scissors
Rock 0,0 -1,1 1,-1
Paper 1,-1 0,0 -1,1
Scissors -1,1 1,-1 0,0
Heads Tails
Heads 1,-1 -1,1
Tails -1,1 1,-1
Matching Pennies Rochambeau (Rock-Paper-Scissors)
5
A game with a trivial, unique NE
Rock Paper Scissors
Rock 0,0 -1,1 1,-1
Paper 1,-1 0,0 -1,1
Scissors -1,1 1,-1 0,0
Heads Tails
Heads 1,-1 -1,1
Tails -1,1 1,-1
Matching Pennies Rochambeau (Rock-Paper-Scissors)
(www.worldrps.com)
6
A game with a trivial, unique NE
Rock Paper Scissors
Rock 0,0 -1,1 1,-1
Paper 1,-1 0,0 -1,1
Scissors -1,1 1,-1 0,0
Heads Tails
Heads 1,-1 -1,1
Tails -1,1 1,-1
Matching Pennies Rochambeau (Rock-Paper-Scissors)
(www.worldrps.com) Lesson Nash equilibrium not
necessarily instructive
7
Some Intuition about Learning
Left Right
Up 1,0 3,2
Down 2,1 4,0
Stackelberg Game
8
Some Intuition about Learning
Left Right
Up 1,0 3,2
Down 2,1 4,0
Stackelberg Game
Lesson cant separate learning from teaching
9
The typical GT work on MAL

• Define a certain learning procedure (or dynamics)
• fictitious play
• rational learning
• no-regret learning
• Prove conditions under which it converges in the
  limit
• to NE, Correlated NE, etc
• either in actual strategy or in empirical
  frequency
• And almost always in self play

10
Five Distinct Research Agendas in MAL

• Computation Quick-and-dirty method for (e.g.) NE
• Social science How people (institutions,
  animals) learn.
• Game theory puritanism Equilibria of learning
  strategies.
• Distributed control Learning in common-payoff
  games.
• Targeted learning Learning when you have some
  sense of how your opponents might behave.

11
Lesson Need to take NE with a grain of salt

• Beautiful, clever
• Makes it hard to back off from assumptions of
  perfect rationality can we have an alternative,
  constructive game theory?
• In any event, best response computation merits
  as much attention as eqm

12
Talk Outline

• Computing solution concepts, primarily NE
• The role of NE unclear
• Multi-agent learning
• Ditto
• Compact games (graphical games, MAIDs, game
  networks, local-effect games, social networks, )
• Other forms of compactness, and what about
  coalitional games?
• Mechanism design, in particular auctions
• Behavioral Mechanism design
• Beyond GT Algorithmic Institutional Design

13
On compact representations

• Compact representations are fine need more
• Programming constructs in strategy descriptions
  (programmatic rationality)
• Partial games (e.g., logic-based game
  description)
• What about coalitional games?

14
Marginal Contribution Nets

• Games represented by sets of rules
• pattern ? value
• a b c ? 5
• Value of a group S equals the sum of the values
  of the rules S satisfies
• v(S) ?r S satisfies r v(r)
• Focus on conjunction negation in pattern

15
Conciseness of MC-Nets

• Theorem MC-Nets generalize the multi-issue
  representation of CS04
• Theorem MC-Nets generalize the graphical
  representation of DP94

16
Computational Leverage

• Shapley value can be efficiently computed in
  MC-nets
• Exploiting Additivity and Symmetry
• Determining membership in core is hard, but one
  can determine membership in time exponential in
  treewidth
• Determining emptiness, or finding an arbitrary
  member of a non-empty core, are no harder

17
Talk Outline

• Computing solution concepts, primarily NE
• The role of NE unclear
• Multi-agent learning
• Ditto
• Compact games (graphical games, MAIDs, game
  networks, local-effect games, social networks, )
• Other forms of compactness, and what about
  coalitional games?
• Mechanism design, in particular auctions
• Behavioral Mechanism design
• Beyond GT Algorithmic Institutional Design

18
Recall some results from auction theory

• Informal observations
• Dutch First-price, sealed bid
• English ? Second-price, sealed bid (cf. proxy
  bidding)
• Japanese ? English
• Second-price and Japanese have dominant
  strategies
• For precise analyses, need to distinguish between
• Common values and independent values (winners
  curse)
• Risk averse, risk-neutral and risk-seeking
  bidders
• Formal results speak to
• Whether an auction is incentive compatible
• Whether the auction is efficient
• Whether the auction is revenue maximizing

19
Example of BMD Online marketing

• The X5 story
• What are we optimizing for?
• Behavioral requirements (BMD) (ack Moshe
  Tennenholtz)
• sign-ups
• return visits (magic number 5)
• Message injection
• Product education
• Truthful consumer surveys
• Yields a new perspective on existing mechanisms
• Suggests new mechanisms

20
Some new truths about auctions, from the
perspective of marketing

• First-price sealed-bid auction ? Dutch auction
• Second-price sealed-bid auction ? English auction
• Dominant-strategy mechanisms can be suboptimal
• Barter- and multiple-currency markets might trump
  markets with universal currency

21
Some new, marketing-oriented mechanisms

• Tournament auction
• Infinitely many equilibria
• Average-price auction
• Giving the little guy a chance
• Team bidding
• Cooperation
• Community auction
• Coopetition
• Online collectibles
• The marketing advantages of barter systems
• Preference auction
• Win-win for the auctioneer and buyers

22
Tournament auction
A series of sealed-bid auctions X make it to
the next day person with highest remaining
points wins.
23
Tournament auction
Other activities added to basic tournament auction
24
Inserting a population game into an auction
Capturing information about consumers and their
views of others the latter is particularly
truthful.
25
Average Price Game
The consumer who bids closest to the average of
all bids wins the prize.
26
Team Bidding
Bidders form teams and pool their bids.
27
Cariocas Community Auction
Community Auction
A global bid triggers the close of multiple
auctions.
28
Online collectibles
Online collection of digital objects, initially
assembled by various online activities.
29
Online collectibles
and then exchanged via online barter
30
Main takeaways

• Marketing considerations completely change the
  rules of the game. Some lessons of BMD
• new design criteria
• new perspectives on existing mechanisms
• new mechanisms
• Many applications beyond marketing. Example
  Captchas, ESP
• A lot more work is needed before this becomes a
  science

31
Talk Outline

• Computing solution concepts, primarily NE
• The role of NE unclear
• Multi-agent learning
• Ditto
• Compact games (graphical games, MAIDs, game
  networks, local-effect games, social networks, )
• Other forms of compactness, and what about
  coalitional games?
• Mechanism design, in particular auctions
• Behavioral Mechanism design
• Beyond GT Algorithmic Institutional Design

32
Algorithmic Institutional Design (ack Mike Munie)

• What is better The EE or CS qual structure at
  Stanford?
• Similar for job interviews, admissions, consumer
  surveys, etc
• Reminiscent of, but distinct from, the secretary
  problem
• The answer Depends on what youre optimizing
  for. And even given that, depends.

33
Formal Model, continued
34
Results

• Multiple versions
• Single prof?
• Single student?
• Parallel or sequential?
• Sample results
• Even in simplest case, selecting an optimal set
  of questions is NP-Hard, and is not submodular,
  so there is a not an obvious approximation
  algorithm
• Sequentiality can be maximally helpful
• In the multiagent setting, even deciding between
  committee structures is NP-Hard
• Seems like there are well behaved special cases

35
Talk Outline

• Computing solution concepts, primarily NE
• The role of NE unclear
• Multi-agent learning
• Ditto
• Compact games (graphical games, MAIDs, game
  networks, local-effect games, social networks, )
• Other forms of compactness, and what about
  coalitional games?
• Mechanism design, in particular auctions
• Behavioral Mechanism design
• Beyond GT Algorithmic Institutional Design

36
thank you!