Combinatorial Auctions By: Shai Roitman e-mail: shairoi@cs.huji.ac.il Auctions • • • • One to many mechanism Efficient Allocation of the items. Seller Auctions Buyer Auctions - Reversed Auctions Known Auction Types • Open Cry Auctions – English – Dutch • Sealed Bid Auctions – First Price – Second Price The equivalence of auctions • True Valuations – English – Sealed Bid Second Price • Winners Curse – Dutch – Sealed Bid First Price Sealed Bid Auctions advantages • Communication efficient • The value of the bid can be kept private. Items Value • Private Value - An Item has a value to the bidder regardless of the value to the other bidders – Example: Consumer goods • Public Value - The item has value in the context of other bidder estimations – Example: Stocks Strategies for the Auctions under private value assumptions • English Auction – Small increments until maximum price(true value) reached. • Second price Sealed Auction – Submit the evaluated value as the bid Strategies for Auctions continued • First price Sealed Auction & Dutch Auction – Need to evaluate others evaluation (may use some distribution on the values of the other bidders) and use this evaluation for setting the bid. – Winners Curse • Complex analysis Multi Item Auctions - Multi Stage Auction • Scenario – A set of items has to be sold • Naive Solution – Hold auctions for each item or set of items one at a time Multi Item Auctions - Problems • How to choose the order of the items to be sold? • How to bundle several dependant items? • If the items have dependencies multi stage auctions can lead to inefficient allocation Combinatorial auction • Items may be grouped as bundles. • => Takes into considerations the dependencies between the items. • => Greater economic efficiency The Utility function • • • • • Private - Public Value Super Additive - Supplemental items Sub Additive - Complementary items Monotonic - The more the better Convex - Diversity Uses for combinatorial auctions • • • • FCC Radio spectrum Logistics Scheduling Any purchase of dependant multiple items. Logistic explicit use case of combinatorial auctions • Logistics.com - OptiBid(TM) – Trucking companies bid on bundles of lanes – Logistics.com - More than $5 billion in transportation contracts been bid to date (January 2000) (Ford, Wal-Mart, K-Mart). Incentive Issues - An example • 3 bidders {1,2,3} • 2 items {x,y} • Bidder 1 values – {x,y}=100 {x}={y}=0 • Bidder 2 values – {x,y}=0 {x}={y}=75 • Bidder 3 values – {x,y}=0 {x}={y}=40 Incentive Issues - An example continued • If bid truthfully - x->2 , y->3 (Revenue 115) • If Bidder 2 and Bidder 3 belief that the others truthfully bid their values – Bidder 2 can shade his value of {x} and {y} to 65 and still get the same x->2 y->3 (Revenue 105) – Bidder 3 can shade his value of {x} and {y} to 30 and still get the same x->2 y->3 (Revenue 105) Incentive Issues - An example continued • If Bidder 2 & Bidder 3 shade their value (65 & 30) then they will lose as {x,y}->1 • => Lost of economic efficiency Threshold Problem • a collections of bidders whose combined valuation for distinct portions of a subset of items exceed the bid submitted on that subset by some other bidder. • Difficulty in coordination of their bids to outbid the single large bidder on that subset Auction Scheme assumptions • Independent private values for bidders • values draw from a commonly known distribution • risk neutral Auction Design- An optimal mechanism • Truth Revelation - revelation principle • No Bidder is made worse off by participating • Seller Maximum Expected Revenue Efficiency • If the allocation of objects to bidders chosen by the seller solves the following equations than the auction is efficient Let : N Set of Bidders M Set of distinct objects S .S M Bundles of objects 1 y(S , j) 0 Efficient Auction max v j Bundle S is allocated to j Otherwise (S ) y(S , j) jN S M s.t. y ( S , j ) 1 i M y(S , j) 1 j N iS jN S M y ( S , j ) 0,1S M , j N General CAP Formalization Let : N Set of Bidders M Set of distinct objects S .S M Bundles of objects b i ( S ) The bid that agent j N has announced for bundle S . b( S ) max b i ( S ) jN Bundle S is allocated to j 1 y(S , j) 0 Otherwise CAP (Combinatorial Auction problem ) max b j (S ) y(S , j) jN S M s.t. y ( S , j ) 1 i M y(S , j) 1 j N iS jN S M y ( S , j ) 0,1S M , j N Vickrey Clarke Groves (VCG) part 1 1. Agent j reports v j 2.The seller chooses the allocation that solves V max v j (S ) y(S , j) jN S M s.t. y ( S , j ) 1 i M y(S , j) 1 j N iS jN S M y ( S , j ) 0,1S M , j N Call this optimal allocation y * Vickrey Clarke Groves (VCG) part 2 3.To compute the payment that each bidder must make : for each k N V k max v j (S ) y(S , j) jN \ k S M s.t. y ( S , j ) 1 i M y(S , j) 1 j N \ k iS jN \ k S M y ( S , j ) 0,1S M , j N \ k Call this solution y k 4.The payment that bidder k makes is equal to p (k ) V k V v k ( S ) y * ( S , k ) S M p(k ) 0 Vickrey Clarke Groves (VCG) part 3 Seller Re venue V V k V kN • If no agent has a significant effect on the average V is close to V^(-k) thus the revenue is close to the maximum revenue defined in the General CAP. Problems in the VCG mechanism • Solving the CAP problem is hard (NP-Hard) • Using Approximate solutions => Not incentive compatible • Payments in VCG are sensitive to the choice of the solution General CAP Formalization Let : N Set of Bidders M Set of distinct objects S .S M Bundles of objects b i ( S ) The bid that agent j N has announced for bundle S . b( S ) max b i ( S ) jN Bundle S is allocated to j 1 y(S , j) 0 Otherwise CAP (Combinatorial Auction problem ) max b j (S ) y(S , j) jN S M s.t. y ( S , j ) 1 i M y(S , j) 1 j N iS jN S M y ( S , j ) 0,1S M , j N Multiple object in the CAP Formulation Let : N Set of Bidders M Set of distinct objects S .S M Bundles of objects b i ( S ) The bid that agent j N has announced for bundle S . b( S ) max b i ( S ) jN Bundle S is allocated to j 1 y(S , j) Otherwise 0 CAP (Combinatorial Auction problem ) max b j (S ) y(S , j) jN S M s.t. y(S , j) l i M l number of times object i is in the items y(S , j) 1 j N iS jN S M y ( S , j ) 0,1S M , j N The CAP (Combinatorial Auction Problem) • Bidders must submit bid for every subset • Transmitting the bid sets in a succinct manner Restriction of conditions => solvable solution - an example • Restriction – All bidders complement each other – all bidders are symmetric • Solution – Auction all the items as one item in an optimal single item auction Cybernomics experiments • Performed tests for additive values and valuations with synergies of small , medium or high intensity • Results – Combinatorial multi round auctions always superior in efficiency but lower in revenue – Slower convergence (finishing the auctions) The CAP - continued • partial solutions – Restriction on the way the bids are transmitted • OR / OR* Trees • Single mind restriction – Sending an Oracle • Problem of deciding the collection of bids to accept The SPP Problem • Given a set of M elements • collection V of subsets with weights • Find the largest weight collection of subsets that are pairwise disjoint. The SPP Formalization 1 if the j th set in V with weight c j is selected xj 0 otherwise 1 if the j th set in V contains element i M aij 0 otherwise The SPP Pr oblem max c j x j jV s.t. a jV ij x j 1 i M x j 0,1 j V SPP Related Problem - Set Partitioning Problem (SPA) 1 if the j th set in V with weight c j is selected xj 0 otherwise 1 f the j th set in V contains element i M aij 0 otherwise The SPA Pr oblem max c j x j jV s.t. a jV ij x j 1 i M x j 0,1 j V SPP Related Problem - Set Covering Problem 1 if the j th set in V with weight c j is selected xj 0 otherwise 1 f the j th set in V contains element i M aij 0 otherwise The SCP Pr oblem min c jV j xj s.t. a jV ij x j 1 i M x j 0,1 j V What is the complexity of SPP? • SPP Is a NP-Hard / Complete problem • SPP Problem is exponential in |V| (V the number of subsets of M)! • No Hope?? Effective solution to the CAP Problems • Requirements – Number of distinct bids is not large – Underlying SPP problem can be solved reasonable quick. SPP Approximation • There is no Polynomial algorithm that can deliver a worst case ration larger than n^(E1) for any E>0 • There is a worst case ratio of O(n/(log n)^2) algorithm (Polynomial algorithm) Other Approaches • Decentralized Methods – Setting up a fictitious market determining an allocation and prices – Choosing an allocation and bidders are required to send improvements Conclusions • Combinatorial Auctions can lead to higher economic efficiency • Practical Combinatorial Auctions are hard to implement with compliance to the truth revelation principle