fo for sel r selli ling ng in info form rmation ation
play

fo for Sel r Selli ling ng In Info form rmation ation Moshe - PowerPoint PPT Presentation

Op Opti tima mal Mec l Mecha hani nism sms fo for Sel r Selli ling ng In Info form rmation ation Moshe Babaioff Robert Kleinberg Renato Paes Leme (MSR-SVC) (Cornell) (Cornell) The Secret Agent The Secret Agent Moscow London


  1. Op Opti tima mal Mec l Mecha hani nism sms fo for Sel r Selli ling ng In Info form rmation ation Moshe Babaioff Robert Kleinberg Renato Paes Leme (MSR-SVC) (Cornell) (Cornell)

  2. The Secret Agent

  3. The Secret Agent Moscow London Valencia

  4. The Informant The Secret Agent Moscow London Valencia

  5. The Information The Informant The Secret Agent Moscow London Valencia

  6. The Information How to sell information ? The Informant The Secret Agent Moscow London Valencia

  7. Th This s is no s not (only) ly) a a tal alk ab about ut esp spiona onage. ge.

  8. cookies, u ser data… The Information How to sell information ? The Data Provider The Advertiser Potential ads to show

  9. The Information How to sell information ? The Informant The Secret Agent Moscow London Valencia

  10. The Information The Informant The Secret Agent Moscow London Valencia

  11. The Information The Informant The Secret Agent Moscow London Valencia

  12. The Information The Informant The Secret Agent Moscow London Valencia

  13. The Information Common Bayesian Prior The Informant The Secret Agent Moscow London Valencia

  14. The Information Common Bayesian Prior The Seller er The Buyer Moscow London Valencia

  15. More formally … • Seller knows .. Buyer knows . • Pair comes from a joint distribution that is common knowledge • Buyer needs to pick an action getting reward Co Context: text:

  16. Buyer (Secret Agent) Utility • If he doesn’t know (i.e. only knows ) • If he also knows

  17. Buyer (Secret Agent) Utility • If he doesn’t know (i.e. only knows ) • If he also knows • Expected surplus for full information

  18. Ho How mu w much h of th f this s su surp rplu lus can an the e se sell ller er (info formant) rmant) extract given that he doesn’t know w ? ?

  19. Why not post a price ?

  20. Why not post a price ? • with ¼ probability each • with ½ probability (danger level)

  21. Why not post a price ? • with ¼ probability each • with ½ probability (danger level) • if and 0 o.w. • if and 0 o.w.

  22. Why not post a price ? • with ¼ probability each • with ½ probability (danger level) • if and 0 o.w. • if and 0 o.w.

  23. Why not post a price ? •

  24. Why not post a price ? • • Post price 50 for the whole information will generate revenue ½ ∙ 50 = 25

  25. Why not post a price ? • • Post price 50 for the whole information will generate revenue ½ ∙ 50 = 25 • Post price 50 for and 0.5 for will generate revenue ½ ∙ 50 + ½ ∙ 0.5 = 25.25

  26. Why not post a price ? • • Post price 50 for the whole information will generate revenue ½ ∙ 50 = 25 • Post price 50 for and 0.5 for will generate revenue ½ ∙ 50 + ½ ∙ 0.5 = 25.25 Infor ormat mation ion is a lot t mo more e flex exibl ible e than traditional goods.

  27. What is a feasible mechanism ? $$$ message message $$$ message message message Informant proposes a mechanism based on . and commits to faithfully follow it. The agent is strategic. Informant wants to maximize revenue.

  28. How to design optimal mechanisms ? 1. Start with any possible interactive mechanism

  29. How to design optimal mechanisms ? 1. Start with any possible interactive mechanism 2. Show that they need to assume a particular format (revelation principle still holds but is not enough)

  30. How to design optimal mechanisms ? 1. Start with any possible interactive mechanism 2. Show that they need to assume a particular format (revelation principle still holds but is not enough) 3. Optimize over the set of reduced mechanisms (usually an infinite dimensional optimization problem)

  31. How to design optimal mechanisms ? 1. Start with any possible interactive mechanism 2. Show that they need to assume a particular format (revelation principle still holds but is not enough) 3. Optimize over the set of reduced mechanisms (usually an infinite dimensional optimization problem) 4. Give a structural characterization that brings it down to a manageable (polynomial) size.

  32. How to design optimal mechanisms ? 1. Start with any possible interactive mechanism 2. Show that they need to assume a particular format (revelation principle still holds but is not enough) 3. Optimize over the set of reduced mechanisms (usually an infinite dimensional optimization problem) 4. Give a structural characterization that brings it down to a manageable (polynomial) size.

  33. Independent and . Theorem: If and are independent, there exists an optimal mechanism that offers to the buyer a list where is a random variable correlated with and is its price.

  34. Independent and . Theorem: If and are independent, there exists an optimal mechanism that offers to the buyer a list where is a random variable correlated with and is its price. Examples: Y i is a “noisy” version of : • a subset of the bits • the XOR of two bits • with prob ½ and random with prob ½

  35. Independent and . What does this theorem mean? $$$ message message $$$ $$ $$ message message message

  36. Correlated and . Theorem: If and are correlated, there exists an optimal mechanism that offers to the buyer a list where is a random variable correlated with and is its price depending on the outcome of .

  37. Correlated and . Theorem: If and are correlated, there exists an optimal mechanism that offers to the buyer a list where is a random variable correlated with and is its price depending on the outcome of . We can find the optimal mechanism in polynomial time using convex programming.

  38. $$ $$ $$ $$ Independent case Correlated case

  39. What bad can happen ? $$ $$ Correlated case

  40. What bad can happen ?

  41. What bad can happen ?

  42. What bad can happen ?

  43. What bad can happen ?

  44. What bad can happen ?

  45. What bad can happen ?

  46. What bad can happen ? ???

  47. What bad can happen ? ??? This mechanism doesn’t work if the buyer is allowed to defect at any point.

  48. One possible fix : large deposit upfront $$ $$

  49. One possible fix : large deposit upfront $$ $$

  50. One possible fix : large deposit upfront $$ $$ Increases participation cost, creates incentives for the informant to defect, …

  51. Qu Questi estion on: Wh What at is t s the he re revenue enue op optimal mal me mechanism anism wh wher ere (1) bu buye yer is all r is allow owed ed to o de defe fect (2) no o po posi siti tive ve tra ransfer sfers s ar are all e allow owed ed

  52. Th The an e answ swer er is s pu puzzli ling. ng.

  53. Mechanisms for uncommitted buyers with no positive transfers Theorem: Interactive mechanism are necessary in order to get optimal revenue.

  54. Mechanisms for uncommitted buyers with no positive transfers Theorem: Interactive mechanism are necessary in order to get optimal revenue. $$ $$

  55. Mechanisms for uncommitted buyers with no positive transfers How long can the protocol be ? How to optimize over interactive mechanisms ? How to do mechanisms design beyond the one-round revelation principle ?

  56. Mechanisms for uncommitted buyers with no positive transfers How long can the protocol be ? ?? ?? How to optimize over interactive mechanisms ? How to do mechanisms design beyond the one-round revelation principle ?

  57. Open Problems How to design optimal interactive mechanisms ?

  58. Open Problems How to design optimal interactive mechanisms ? Multiple buyers and sellers : a market for information

  59. Open Problems How to design optimal interactive mechanisms ? Multiple buyers and sellers : a market for information Coupling goods and information

  60. Open Problems How to design optimal interactive mechanisms ? Multiple buyers and sellers : a market for information Coupling goods and information Crypto primitives and computationally bounded agents

  61. Open Problems How to design optimal interactive mechanisms ? Multiple buyers and sellers : a market for information Coupling goods and information Crypto primitives and computationally bounded agents Continuous type spaces

Recommend


More recommend