Presentation is loading. Please wait.

Presentation is loading. Please wait.

Competitive Auctions and Digital Goods Andrew Goldberg, Jason Hartline, and Andrew Wright presenting: Keren Horowitz, Ziv Yirmeyahu.

Similar presentations


Presentation on theme: "Competitive Auctions and Digital Goods Andrew Goldberg, Jason Hartline, and Andrew Wright presenting: Keren Horowitz, Ziv Yirmeyahu."— Presentation transcript:

1 Competitive Auctions and Digital Goods Andrew Goldberg, Jason Hartline, and Andrew Wright presenting: Keren Horowitz, Ziv Yirmeyahu

2 Introduction Selling a large number of items every customer wants a single item unlimited supply examples: –downloadable software –pay-per-view movies

3 Motivation Mechanism Design purpose: –to maximize seller’s revenues! Using: –Fixed price auctions –Multiple price auctions –Deterministic auctions –Randomized auctions –All single round, sealed bids

4 The economic view Quantity Price

5 The economic view Revenues = Price X Quantity Quantity Price

6 Notations Set of Bids, B=, sorted:, so = utility of bidder I T = F = Optimal fixed price revenues R = current game revenues assumption: h is small compare to F

7 Notations 2 Truthful auction: encourage bidders to bid their utility value Competitive auction: yields revenues within a constant factor of optimal fixed price. We use computer science analysis similar to online algorithms analysis where we check if

8 For example: The k-item vickery auction: –truthful, single price Worst case scenario: –k bidders bid at h, n-k bidders bid at 1. Vickery revenues (R) = k optimal fixed price revenues (F)

9 How T relates to F Theorem: T/(2log h)  F First, see that F uses the price opt(B):

10 How T relates to F (2) Proof: 1. Divide the bids into log h bins. 2. the sum of the bids = T => there exists a bin, say X, that sums to at least T/(log h). 3. the lowest bid in X is at least 1/2 the highest. 4. choose price(B) to be the lowest bid in that bin. 5. Each bid in that bin will contribute at least half its value. 6. X

11 Random Sampling Auctions Select a subset B’ of B (the set of bids), Compute f(B’) to be the optimal fixed price for B’ Use f(B’) as a threshold to B\B’ Bidder I wins the auction at price f(B’) if

12 Performance analysis Theorem: assume, then with a probability Proof is basing in the Chernoff bound. By this Theorem:

13 Weighted Pairing Auction A bid-independent, truthful, multi price auction: Bids distribution

14 Weighted pairing performance if, then for the weighted pairing auction E(R)=  (T/(log h)). But still, weighted pairing isn’t competitive, in the worst case, it has a tight bound of:

15 An upper bound We can see that no truthful auction, even a multi price one, can do better then F in the average case: Theorem: for any truthful auction:

16 An upper bound - proof Let’s define: – the probability a bid I is satisfy – expected cost to winning bidder I – expected profit (gain) for bidder I (1)

17 An upper bound - proof(2) Lemma: in a truthful auction: Proof: if, then together: since, we have.

18 An upper bound - proof(3) Theorem: for any truthful auction: Proof: if bidder I-1 had the utility value of his gain will not exceed: (2) so:

19 An upper bound - proof(4) We can recursively extend in the same way until: define the total expected revenue from bidder I as: rewrite (1) as: using equation (3) we get:

20 An upper bound - proof(5)

21 An upper bound - proof(6) define, this is also the revenues from a fixed price auction using as the price. Note that. Rearrange the sum over and define :

22 An upper bound - proof(7) but and by the lemma in the beginning we showed that so: this sum telescopes to but so:

23 Deterministic Optimal Threshold Auction A bid-independent auction, meaning bidder i‘s bid value should only determine whether bidder i wins or losses. The bid value should not determine bidder i’s price. Uses optimal threshold function opt (Bi),to be used as a threshold for bidder i, which on set of bids B\b i returns the fixed price at which items should be sold to achieve revenue F. This action has a worst case input that causes it not to be competitive.

24 Deterministic Optimal Threshold Auction (cont.) An example for worst case input : –r bids with value h. –(h-1)r-1 bids with value 1. –Optimal single price auction take r bid with price h s.t F= rh => threshold is h. (the second best threshold is 1 taking all bid for revenue of hr-1). – Deterministic optimal threshold auction takes r high bids with price 1 s.t R=r. (the high are taken in price 1 because removing one h bids causes the opt. threshold to switch to 1) –So R/F=1/h.

25 Upper bound for deterministic Bid-independent action Theorem: For any truthful deterministic bid-independent auction and any constant α there exist an input for which R/F =O(1/h) and αh  F.

26 Truthful deterministic Auction are Bid-independent Lemma: Any truthful deterministic auction is bid independent Proof: –B i = B\{i} –B ix = b i replaced with value x. –A(B) = the result of running auction A on set of bids B. –A i (B) – the result for bidder i(  if i is rejected). –We will define function g as: g(x) = Ai(B ix ).

27 Proof(cont.) It is left to show that g is  except for some region (v,  ), [v,  ) or {} where it has the value v. This would imply A is bid-independent. If g(x)=  for all x than we have empty interval {}. Otherwise, let b = inf{x:g(x)   } and v=min x g(x). It is left to show b=v and for all b’>b, g(b’)=v

28 Proof(cont.) 1.v=b: v  b since g(x) is defined as  (loss) or g(x)  x (a bid will win at price necessarily at most x). v  b since else there might exist b’ s.t v<b’<b. The bidder with utility value b’ would be better off biding greater than b winning the auction and pay v which is less that b’. This is a contradiction of A been truthful. Thus, v=b. 2.For all b’>b, g(b’)=v: since that if g(b’) will be bigger than v, a bidder with utility b’ will be better off bidding b. This is a contradiction of A been truthful.

29 Upper bound for deterministic Bid-independent action Theorem: For any truthful deterministic bid-independent auction and any constant α there exist an input for which R/F =O(1/h) and αh  F.

30 Why Simulation is needed? In terms of asymptotic worst-case performance, deterministic auctions are significantly worse that randomized auction => this is not to say deterministic auctions are bad to use for all input families. Constant factors (not necessarily tight) obtained in the analysis do not enable determination which auction does better. Theoretical analysis of the presented auctions for specific distributions is non-trivial.

31 Some simulation results On uniform and normal distribution (“average case” families) for large n the dual-price sampling optimal threshold and the deterministic optimal threshold have the best results. The above families have the property that any uniformly chosen random subset of the bids has the same distribution as the original. Because of this random sampling auction perform very good on such families. Weighted pairing is the worst auction for average case families.

32 Some simulation results (cont.) Even on contrived worst-case families the auction revenue is a large constant fraction of F. Sampling about a square root of the number of bids for the sampling auction seems to balance out the loss due to rejecting all of the sample and the loss due to having a non representative sample.

33 Extension for Bounded Supply The bounded supply case is a generalization of the unlimited supply case as items are available in unlimited supply when the number of available items is the same as the number of bidders. Denote the number of items available by k. Denote by F k the revenue in the bounded supply case. Let opt k be the function that given a set of bids return the optimal threshold that sells k item or less.

34 Extension for Bounded Supply(cont.) The single price sampling optimal threshold will be modified to use threshold function opt mk/(n-m) on sample of size m. If more than k bids are satisfied, we arbitrary reject bids until there are only k left. The dual-price auction with sample size of m=n/2 use opt k/2 so that about k/2 bids are selected from each of the sample and the non-sample. Both auction can be shown to be competitive.

35 Conclusion There exist truthful auction for unlimited supply market. Randomized auction are competitive in that they yield revenue that is within a constant factor of optimal fixed pricing. No deterministic auction is competitive in the worst case. The presented auctions compare favorably to fixed pricing with market analysis (simulation result)

36 Appendix: Competitive Auction for Multiple Digital Goods.

37 Problem Definition The problem of selling several items, with each item available in unlimited supply. Each bidder wants only a single item. Example: concurrent broadcast of several movies. The input for an auction: –Number of bidders n. –Number of items m. –Set of bids {a ij }.

38 Problem Definition (cont.) Auction outcome: an assignment of a subset of wining bidders to items. Each bidder i in the subset is assigned a single item j and a sales price of at most a ij. Stable auction: bidding U ij is a dominant strategy for bidder i. It maximize the bidder profit at the bidder’s utility value. Note: the set of bidders strategies where bidder i bids u ij for item j is a Nash equilibrium.

39 “Fixed Price” Auction Bidders supply bids a ij, seller supplies sale prices r j 1  j  m. Definition c ij =a ij -r j. The auction assigns each bidder i to the item j with the maximum c ij. The sale price is r j. Lemma: Suppose the sale prices are set independently of input bids. Then fixed price auction is stable.

40 Optimal Sale Prices The problem of finding for a given set of bids and a set of prices an assignment such that the fixed price auction brings the highest revenue can be view as a matching problem. Translating this to an integer program we get the following linear program: max  j  i X ij (a ij -r j ) s.t.  j X ij  1 1  i  n Xij  0 1  i  n, 1  j  m X ij is one exactly when bidder i get item j.

41 Optimal Sale Prices (cont.) The dual problem is expressed it terms of p i, the profit of the corresponding bidders: min  i p i s.t. p i  a ij –r j 1  i  n, 1  j  m p i  0 1  i  n If we combine both of these optimization problems we get:

42 Optimal Sale Prices(cont.) Find X ij and P i s.t  j X ij  1 1  i  n Xij  0 1  i  n, 1  j  m p i + r j  a ij 1  i  n, 1  j  m  i p i =  j  i X ij (a ij - r j )

43 Optimal Sale Prices(cont.) Treating r j as variable the optimal sale price problem as the following mathematical programming problem: max  j  i X ij r j s.t. r m =0  j X ij  1 1  i  n Xij  0 1  i  n, m  j  m p i + r j  a ij 1  i  n, m  j  m  i p i =  j  i X ij (a ij - r j )

44 Extension of the random sampling auction Pick a random sample S of the set of bidders. Let N be the set of bidders not in the sample. Compute the optimal sale prices for S. The result of the random sampling auction is then just the result of running the fixed price auction on N using the sale prices computed in the previous step.

45 Extension of the Deterministic Auction Delete i from B. Compute optimal prices for the remaining bidders Choose the most profitable item for i under these prices. This is done independently for each bidder.

46 Conclusion The random sample auction is competitive. In order to studies this auction using simulation there is a need for a fast algorithm for calculation the optimal fixed pricing problem.


Download ppt "Competitive Auctions and Digital Goods Andrew Goldberg, Jason Hartline, and Andrew Wright presenting: Keren Horowitz, Ziv Yirmeyahu."

Similar presentations


Ads by Google