AAMAS 2013 best-paper: “Mechanisms for Multi-Unit Combinatorial Auctions with a Few Distinct Goods” Piotr Krysta Orestis Telelis Carmine Ventre University of Liverpool, UK AUEB, Greece Teesside University, UK Multi-unit Combinatorial Auctions m goods Good j available in supply sj n bidders Each bidder has valuation functions for (multi) set of goods expressing his/her complex preferences, e.g., v( blue set ) = 290$ v( green set ) = 305$ Objective: find an allocation of goods to bidders that maximizes the social welfare (sum of the bidders’ valuations) (Multi-unit) CAs: applications CAs: paradigmatic problem in Algorithmic Mechanism Design “CAs is algorithms hard to approximate within “We can always return the (deterministic) Polynomial-time and √m and we have a polynomial-time optimum social welfare truthfully truthfulness? algorithm that guarantees that” (ie, when bidders lie) using VCG” VCG is in general not good to obtain approximate solutions [Nisan&Ronen, JAIR 2007] Few distinct goods Polynomial-time (deterministic) algorithms and truthfulness for m=O(1) and sj in N? VCG-based mechanisms do the job in this case! Our results at a glance Valuation Previous best apx NEW APX (m=O(1)) Single-minded 2-apx (m=1) [Mualem, Nisan’02] FPTAS (m=1) [BKV’05] Apx lower bound Weakly NP-complete Weakly NP-complete ( m , 1+ε, …,1+ε) hard (arbitrary m) [NEW] (1+ε,1+ε,…,1+ε)-FPTAS (m=O(1)) [GKLV’10] Multi-minded PTAS (m=1) (1+ε,1,…,1)-PTAS [Dobzinski, Nisan’07] Strongly NP-hard (m≥2) [ChK’00] No FPTAS (m=1) [DN’07] (1,1+ε,…,1+ε)-FPTAS Weakly NP-complete Submodular 1-apx (m=1) [Vickrey’61] (1+ε,1,…,1)-PTAS Exponential-time General 2-apx (m=1) [DN’07] (2, 1,…,1)-apx First deterministic poly-time mechanism even for m=1. Greatest improvement over previous result! ? 2-MiR-hard (m=1) [DN’07] VCG-based mechanisms: Maximum-inRange (MIR) algorithms [NR, JAIR 07] Algorithm A : b ® X is MIR, if it fully optimizes the Social Welfare over a subset R Í X of allocations. Truthful (Poly-Time) α-approximate VCG-based mechanism: 1. 2. 3. 4. Commit to a range, R, prior to the bidders’ declarations. Elicit declarations, b. Compute solution in R with best social welfare according to b. Use VCG payments. opt Tricky: R should be “big” enough to contain good approximations of opt for all b and “small” enough to guarantee step 3 to be quick. a · · X R Multi-minded bidders Bidders demand a collection of multi-sets of goods æm ö Di = {d1,..., dri } Í ç ´ {0,1,..., s j }÷ è j=1 ø d1 d2 1 2 ì max{v (d) : d £ x} if {d : d £ x} ¹ Æ ï dÎD i i vi (x) = í ïî 0 otherwise 1 2 3 1 2 d2 d1 x2 3 v(x1 ) = max{v(d1 ), v(d2 )} = 7 1 2 3 1 2 v(x2 ) = 5 3 v(d2 ) = 5 v(d1 ) = 7 Valuation Function x1 1 2 3 3 Allocation algorithm in input e > 0 1. Demands rounding m æê n × d(l) úö "i"d Î Di , define d ¢ = ççê ú÷÷ èë e × sl ûøl=1 1. Supply adjustment énù Define sl¢ = ê ú, for l =1, 2,..., m. êe ú 1. Optimize rounded instance by dynamic programming Optimality (1, 1+ε, …, 1+ε)-FPTAS: Feasible solutions to the original instance are feasible for the “rounded” instance Feasibility (1, 1+ε, …, 1+ε)-FPTAS: ê n × xi (l) ú n × xi (l) å e × s £ åêë e × s úû + {i : xi Î Di } l l i i énù n £ ê ú + n £ +1+ n êe ú e Truthfulness of the mechanism THEOREM: The allocation algorithm A is MiR. Proof: The set {x in X : there exists b s.t. A(b)= x} is the range of the algorithm. THEOREM: There is an economically efficient truthful FPTAS for multi-minded CAs, violating the supplies by (1 + ε), for any ε > 0. (Important: Bidders declare (and can lie about) both demand sets and values.) Violating the supply? • Theoretically needed to obtain an FPTAS – Strongly NP-hardness for m ≥ 2 • Common practice in multi-objective optimization literature • Sellers do that already! Conclusions • Studied Multi-Unit CAs with constant number of goods and arbitrary supply – most practically relevant CAs setting – dramatically changes the problem to be algorithmically tractable! • Designed best possible deterministic poly-time truthful mechanisms for broad classes of bidders: multiminded, submodular, general. – Mechanism for submodular valuations is the first deterministic poly-time. • Our assumptions (m = O(1), relaxing supplies) are provably necessary!