The Sample Complexity of Revenue Maximization in the Hierarchy of Deterministic Combinatorial Auctions

Transcription

1 The Sample Complexity of Revenue Maximization in the Hierarchy of Deterministic Combinatorial Auctions Ellen Vitercik Joint work with Nina Balcan and Tuomas Sandholm Theory Lunch 27 April 2016

2 Combinatorial (multi-item) auctions : $5 : $5 : $6 Combinatorial auctions allow bidders to express preferences for bundles of goods

3 Real-world examples US Government wireless spectrum auctions [FCC] Sourcing auctions [Sandholm 2013] Airport time slot allocation [Rassenti 1982] Building development, e.g. office space in GHC (no money) Property sales

4 Mechanism design Mechanism designer must determine: Allocation function: Who gets what? Payment function: What does the auctioneer charge? Goal: design strategy-proof mechanisms Easy for the bidders to compute the optimal strategy Easy for designer to analyze possible outcomes

5 Warm-up: single-item auctions : $5 Second-Price Auction NINA, -$3 Allocation (N:$5, T:$3) = give carrot to Nina : $3 ø, -$0 Payment (N:$5, T:$3) = charge Nina $3 TUOMAS Second-price auction: the classic strategy-proof, single-item auction.

6 Revenue-maximizing combinatorial auctions Standard assumptions: bidders valuations drawn from distribution D, mechanism designer knows D Allocation and payment rules often depend on D

7 Revenue-maximizing combinatorial auctions Design Challenges Support of D might be doublyexponential NP-hard to determine the revenue-maximizing deterministic auction with respect to D [Conitzer and Sandholm 2002] Feasible Solutions Draw samples from D instead Fix a rich class of auctions. Can we learn the revenuemaximizing combinatorial auction in that class with respect to D given samples drawn from D? Central problem in Automated Mechanism Design [Conitzer and Sandholm 2002, 2003, 2004, Likhodedov and Sandholm 2004, 2005, 2015, Sandholm 2003] No theory that relates the performance of the designed mechanism on the samples to that mechanism s expected performance on D, until now.

8 Outline Introduction Hierarchy of deterministic combinatorial auction classes Our contribution: how many samples are needed to learn over the hierarchy of auctions? Affine maximizer auctions and Rademacher complexity Mixed-bundling auctions and pseudo-dimension Summary and future directions

9 Combinatorial auctions NINA TUOMAS : $1 : 50 : $0 : $1 : 50 : possible outcomes o = (o 1, o 2 ) For example, o = (, )

10 A natural generalization of second price NINA TUOMAS : $1 : 50 : $0 : $1 : 50 : 50 Social Welfare o = SW o = i Bidders v i o o maximizes SW o SW -i o = j Bidders i v j o o i maximizes SW -i o Allocation: o Payment: Nina pays SW Nina o Nina SW Nina o The Vickrey-Clarke-Groves mechanism (VCG).

11 VCG in action NINA TUOMAS : $1 : 50 : $0 : $1 : 50 : 50 o =, o Nina = (,, ) Nina pays v Tuomas (, ) v Tuomas ( ) = 0 How do we get the bidders to pay more?

12 Outcome boosting NINA TUOMAS : $1 : 50 : $0 : $1 : 50 : 50 value,, = v Nina + v Tuomas (, ) = 50

13 Outcome boosting NINA TUOMAS : $1 : 50 : $0 : $1 : 50 : 50 value,, = v Nina + v Tuomas (, ) = o =, o Nina = (,, )

14 Outcome boosting NINA TUOMAS : $1 : 50 : $0 : $1 : 50 : 50 value,, = v Nina + v Tuomas (, ) = o =, o Nina = (,, ) Nina pays v Tuomas (, ) + 99 v Tuomas ( ) = 99

15 Affine maximizer auctions (AMAs) Boost outcomes: λ(o) Take bids v Compute outcome: n j Bidders λ o = argmax o SW o + λ o Compute Bidder i s payment: SW i o i + λ o i SW i o + λ o

16 Affine maximizer auctions (AMAs) Boost outcomes: λ(o) Take bids v Compute outcome: o = argmax o v j o + λ o Compute Bidder i s payment: n j Bidders v j o i + λ o i v j o + λ o j Bidders {i} j Bidders {i}

17 Affine maximizer auctions (AMAs) Boost outcomes: λ(o); Weight bidders: w i Take bids v Compute outcome: o = argmax o v j o + λ o Compute Bidder i s payment: n j Bidders v j o i + λ o i v j o + λ o j Bidders {i} j Bidders {i}

18 Affine maximizer auctions (AMAs) Boost outcomes: λ(o); Weight bidders: w i Take bids v Compute outcome: o = argmax o w j v j o + λ o Compute Bidder i s payment: n j Bidders v j o i + λ o i v j o + λ o j Bidders {i} j Bidders {i}

19 Affine maximizer auctions (AMAs) Boost outcomes: λ(o); Weight bidders: w i Take bids v Compute outcome: o = argmax o w j v j o + λ o n j Bidders Compute Bidder i s payment: 1 w i w j v j o i j Bidders {i} + λ o i w j v j o j Bidders {i} + λ o

20 Hierarchy of parameterized auction classes Affine maximizer auctions [R79] w i, λ o R Virtual valuation combinatorial auctions [SL03] λ o = λ i o i Bidders λ-auctions [J07] w i = 1 λ o R Mixed bundling auctions with reserve prices [TS12] Mixed bundling auctions [J07] w i = 1 λ o = 0 except any outcome where a bidder gets all items item reserve prices w i = 1 λ o = 0 except outcome where a bidder gets all items

21 Outline Introduction Hierarchy of deterministic combinatorial auction classes Our contribution: how many samples are needed to learn over the hierarchy of auctions? Affine maximizer auctions and Rademacher complexity Mixed-bundling auctions and pseudo-dimension Summary and future directions

22 Our contribution Optimize λ o and w given a sample S~D N (Automated Mechanism Design) We want: The auction with best revenue over the sample has almost optimal expected revenue Any approximately revenue-maximizing auction over the sample will have approximately optimal expected revenue For any auction we output, we want S large enough such that: empirical revenue expected revenue < ε In other words, how many samples S = N do we need to ensure that empirical revenue expected revenue = 1 N v S rev A v E v~d rev A v for all auctions A in the class? (We can only do this with high probability.) < ε

23 How many samples do we need? Affine maximizer auctions [R79] N = O Un m m U + n m/2 /ε 2 Virtual valuation combinatorial auctions [SL03] N = O Un m m U + n m/2 /ε 2 λ-auctions [J07] N = O Un m m U + n m/2 /ε 2 Variables N: sample size n: # bidders m: # items U: maximum revenue achievable over the support of the bidders valuation distributions Mixed bundling auctions with reserve prices [TS12] N = O U/ε 2 m 3 Mixed bundling auctions [J07] N = O U/ε 2

24 How many samples do we need? Affine maximizer auctions [R79] N = O Un m m U + n m/2 /ε 2 Virtual valuation combinatorial auctions [SL03] N = O Un m m U + n m/2 /ε 2 λ-auctions [J07] N = O Un m m U + n m/2 /ε 2 Variables N: sample size n: # bidders m: # items U: maximum revenue achievable over the support of the bidders valuation distributions Mixed bundling auctions with reserve prices [TS12] N = O U/ε 2 m 3 Mixed bundling auctions [J07] N = O U/ε 2 Nearly-matching exponential lower bounds.

25 How many samples do we need? Affine maximizer auctions [R79] N = O Un m m U + n m/2 /ε 2 Virtual valuation combinatorial auctions [SL03] N = O Un m m U + n m/2 /ε 2 λ-auctions [J07] N = O Un m m U + n m/2 /ε 2 Variables N: sample size n: # bidders m: # items U: maximum revenue achievable over the support of the bidders valuation distributions Mixed bundling auctions with reserve prices [TS12] N = O U/ε 2 m 3 Mixed bundling auctions [J07] N = O U/ε 2 Learning theory tool: Rademacher complexity

26 How many samples do we need? Affine maximizer auctions [R79] N = O Un m m U + n m/2 /ε 2 Virtual valuation combinatorial auctions [SL03] N = O Un m m U + n m/2 /ε 2 λ-auctions [J07] N = O Un m m U + n m/2 /ε 2 Variables N: sample size n: # bidders m: # items U: maximum revenue achievable over the support of the bidders valuation distributions Mixed bundling auctions with reserve prices [TS12] N = O U/ε 2 m 3 Mixed bundling auctions [J07] N = O U/ε 2 Learning theory tool: Pseudo-dimension

27 Outline Introduction Hierarchy of deterministic combinatorial auction classes Our contribution: how many samples are needed to learn over the hierarchy of auctions? Affine maximizer auctions and Rademacher complexity Mixed-bundling auctions and pseudo-dimension Summary and future directions

28 How many samples do we need? Affine maximizer auctions [R79] N = O Un m m U + n m/2 /ε 2 Virtual valuation combinatorial auctions [SL03] N = O Un m m U + n m/2 /ε 2 λ-auctions [J07] N = O Un m m U + n m/2 /ε 2 Variables N: sample size n: # bidders m: # items U: maximum revenue achievable over the support of the bidders valuation distributions Mixed bundling auctions with reserve prices [TS12] N = O U/ε 2 m 3 Mixed bundling auctions [J07] N = O U/ε 2

29 Key challenge Our problem... Boost outcomes: λ(o); Weight bidders: w i Take bids v Compute outcome: o = argmax o w j v j o + λ o n j Bidders Compute Bidder i s payment: 1 w i w j v j o i j Bidders {i} + λ o i w j v j o j Bidders {i} + λ o Whereas typically in machine learning

30 Rademacher complexity More expressive function classes need more samples to learn How to measure expressivity? How well do functions from the class fit random noise? Empirical Rademacher complexity: x 1,, x N ~ 1, 1 N, S = v 1,, v N 1 R S A = E x sup A A N x i rev A v i, where Rademacher complexity: R N A = E S~D N R S A With probability at least 1 δ, for all A A, empirical revenue expected revenue 2R N A + U 2 ln 2/δ N *U is the maximum revenue achievable over the support of the bidders valuation distributions

31 Rademacher complexity More expressive function classes need more samples to learn How to measure expressivity? How well do functions from the class fit random noise? Empirical Rademacher complexity: x 1,, x N ~ 1, 1 N, S = v 1,, v N 1 R S A = E x sup A A N x i rev A v i, where Rademacher complexity: R N A = E S~D N R S A A = all binary valued functions R N A = 1 2 A = one binary valued function R N A = 0

32 Rademacher complexity of AMAs Theorem Let A be the class of n-bidder, m-item AMA revenue functions. If N = O Un m m U + n m/2 /ε 2, then with high probability over a sample S~D N, empirical revenue expected revenue < ε for all rev A A. Key idea: split revenue function into its simpler components Weighted social welfare without any one bidder s participation (n components) Amount of revenue subtracted out to maintain strategyproof property Then use compositional properties of Rademacher complexity and other tricks, for example: If F = f f = g + h, g G, h H, then R N F R N G + R N H

33 Outline Introduction Hierarchy of deterministic combinatorial auction classes Our contribution: how many samples are needed to learn over the hierarchy of auctions? Affine maximizer auctions and Rademacher complexity Mixed-bundling auctions and pseudo-dimension Summary and future directions

34 How many samples do we need? Affine maximizer auctions [R79] N = O Un m m U + n m/2 /ε 2 Virtual valuation combinatorial auctions [SL03] N = O Un m m U + n m/2 /ε 2 λ-auctions [J07] N = O Un m m U + n m/2 /ε 2 Mixed bundling auctions with reserve prices [TS12] N = O U/ε 2 m 3 Variables N: sample size n: # bidders m: # items U: maximum revenue achievable over the support of the bidders valuation distributions Mixed bundling auctions [J07] N = O U/ε 2

35 Mixed bundling auctions (MBAs) NINA : $1 : $0 : $1 TUOMAS : 50 : 50 : 50 Class of auctions parameterized by a scalar c Boost the allocations where one bidder gets all goods by c value,, = v Nina + v Tuomas (, ) = value,, = v Nina, + v Tuomas ( ) = How large must the sample S be in order to ensure that for all MBAs, empirical revenue expected revenue < ε?

36 revenue Structural properties of MBA revenue functions Lemma Fix v S. Then rev v c is piecewise linear with at most n + 1 discontinuities. c

37 VC-dimension Complexity measure for binary-valued functions only Example: F = {single interval on the real line} No set of size 3 can be labeled in all 2 3 ways by F Class of functions F shatters set S = x 1,, x N if for all b 0, 1 N, there exists f F such that f x i = b i VC-dimension of F is the cardinality of the largest set S that can be shattered by F How can we extend VC-dim to real-valued functions?

38 Pseudo-dimension x 1 f x 1 r (1) 0 x 2 f x 2 r (2) 0 x 3 f x 3 > r (3) 1 x 4 f x 4 r (4) 0 x 5 f x 5 > r (5) 1 x 6 f x 6 r (6) 0 x 7 f x 7 > r (7) 1 x 8 f x 8 > r (8) 1 x 9 f x 9 r (9) 0 Sample S = x 1,, x N Class of functions F into U, U r = r (1),, r (N) R N witnesses the shattering of S by F if for all T S, there exists f T F such that f T x i r (i) iff x i T Pseudo-dimension of F is the cardinality of the largest sample S that can be shattered by F P-dim(F) = VC-dim( (x, r) 1 f x r>0 f F )

39 How many samples do we need? Set of auction revenue functions A with range in 0, U, distribution D over valuations v. For every ε > 0, δ 0, 1, if N = O U ε 2 P dim(a) ln U ε + ln 1 δ, then with probability at least 1 δ over a sample S~D N, empirical revenue expected revenue < ε for every rev A A. Pseudo-dimension allows us to derive strong sample complexity bounds.

40 How many samples do we need? Affine maximizer auctions [R79] N = O Un m m U + n m/2 /ε 2 Virtual valuation combinatorial auctions [SL03] N = O Un m m U + n m/2 /ε 2 λ-auctions [J07] N = O Un m m U + n m/2 /ε 2 Mixed bundling auctions with reserve prices [TS12] N = O U/ε 2 m 3 Variables N: sample size n: # bidders m: # items U: maximum revenue achievable over the support of the bidders valuation distributions Mixed bundling auctions [J07] N = O U/ε 2

41 2-bidder MBA pseudo-dimension Theorem Let A = rev c c 0 be the class of n-bidder, m-item mixed bundling auction revenue functions. Then P-dim(A) = O log n.

42 2-bidder MBA pseudo-dimension Theorem Let A = rev c c 0 be the class of 2-bidder, m-item mixed bundling auction revenue functions. Then P-dim(A) = 2. Proof sketch. Fact: there exists a set of 2 samples that is shattered by A. We need to show that no set of 3 samples can be shattered by A. Suppose, for a contradiction, that S = v 1, v 2, v 3 is shatterable. Recall v 1 = v 1 1 1, v 2 This means: There exists r = r 1, r 2, r 3 R 3 and 2 S = 8 MBA parameters C = c 1,, c 8 such that rev c1,, rev c8 induce all 8 binary labelings on S with respect to r.

43 revenue 2-bidder MBA pseudo-dimension Lemma Fix v i S. Then rev v i c is piecewise linear with one discontinuity, with a slope of 2 followed by a constant function with value i i min v 1 m, v 2 m. min v 1 i m, v 2 i m c i c

44 revenue Case 1: r 3 < min v 1 3 m, v 2 3 m min v 2 3 m, v 2 3 m r 3 c i c

45 Case 1: r 3 < min v 1 3 m, v 2 3 m rev v 3(c) increasing rev v 3 c = min v 1 3 m, v 2 3 m c 3 c

46 Case 1: r 3 < min v 1 3 m, v 2 3 m rev v 3(c) increasing rev v 3 c = min v 3 1 m, v 3 2 m rev v 2(c) increasing rev v 2 c = min v 2 1 m, v 2 2 m c 3 c 2 c

47 Case 1: r 3 < min v 1 3 m, v 2 3 m rev v 3(c) increasing rev v 3 c = min v 1 3 m, v 2 3 m rev v 2(c) increasing rev v 2 c = min v 1 2 m, v 2 2 m rev v 1(c) increasing rev v 1 c = min v 1 1 m, v 2 1 m c 3 c 2 c 1 c

48 Case 1: r 3 < min v 1 3 m, v 2 3 m rev v 3(c) increasing rev v 3 c = min v 1 3 m, v 2 3 m rev v 2(c) increasing rev v 2 c = min v 1 2 m, v 2 2 m rev v 1(c) increasing rev v 1 c = min v 1 1 m, v 2 1 m c 3 c 2 c 1 c We need: rev v 3(c) increasing rev v 2(c) increasing rev v 1(c) increasing This is impossible, so we reach a contradiction. Therefore, no set of size 3 can be shattered by the class of 2-bidder MBA revenue functions, so the pseudo-dimension is at most 2.

49 Summary Analyzed the sample complexity of learning over a hierarchy of deterministic combinatorial auctions Uncovered structural properties of these auctions revenue functions along the way Of independent interest beyond sample complexity results