An exact and two heuristic strategies for truthful bidding incombinatorial transport auctions

Tobias Buer1

Computational Logistics Junior Research Group, University of Bremen, Germany

Abstract

To support a freight carrier in a combinatorial transport auction, we proposes an exact and two heuristic strategies forbidding on subsets of requests. The exact bidding strategy is based on the concept of elementary request combinations.We show that it is sufficient and necessary for a carrier to bid on each elementary request combination in order toguarantee the same result as bidding on each element of the powerset of the set of tendered requests. Both heuristicbidding strategies identify promising request combinations. For this, pairwise synergies based on saving values aswell as the capacitated p-median problem are used. The bidding strategies are evaluated by a computational study thatsimulates an auction. It is based on 174 benchmark instances and therefore easily extendable by other researchers. Onaverage, the two heuristic strategies achieve 91 percent and 81 percent of the available sales potential while generating36 and only 4 percent of the bundle bids of the exact strategy. Therefore, the proposed bidding strategies help a carrierto increase her chance to win and at the same time reduce the computational burden to participate in a combinatorialtransport auction.

Keywords: combinatorial auction, contract procurement, bid generation problem, bundle bidding

1. Introduction

1.1. Relevance of combinatorial transport auctionsIn order to buy and sell services for multiple heterogenous transport requests combinatorial transport auctions are

used. On the one hand, combinatorial auctions are used by large shippers to buy their required transport servicesfrom many carriers (Caplice and Sheffi, 2003, 2006; Sheffi, 2004). The shipper acts as the auctioneer and the carriersare the bidders in the auction. In such a scenario, usually long-term commitments in the sense of a frameworkagreement are negotiated. Real-world transport procurement auctions are voluminous, i. e., in a single auction somehundred requests with an annual value of several million US-Dollar are allocated. Examples of companies that usecombinatorial transport auctions for this purpose are Home Depot, Wal Mart, Procter & Gamble, and Ford Motors(Caplice and Sheffi, 2006). On the other hand, combinatorial auctions can also be used by carriers to implementcollaborative planning for a coalition of carriers (Kopfer and Pankratz, 1999; Krajewska and Kopfer, 2006; Bergerand Bierwirth, 2010). Here, the planning horizion is rather short-term. Transport requests are exchanged betweenmembers of the coalition. The exchange can be organized by combinatorial transport auctions and the goal of theexchange is to find an efficient allocation of transport requests to carriers.

In contrast to conventional auctions, combinatorial auctions allow bidders to submit bundle bids. A bundle bid(also referred to as package bid) is a bid on any subset of the tendered transport requests. The bid price is the amountof money the bidding carrier charges to fulfill all transport requests of the subset. Furthermore, bundle bids are all-or-nothing bids, i. e., a carrier either wins all requests composed in a bundle bid or none. For this reason, carriers

Email address: tobias.buer@uni-bremen.de (Tobias Buer)URL: http://www.cl.uni-bremen.de (Tobias Buer)

1Computational Logistics – Cooperative Junior Research Group of University of Bremen and ISL - Institute of Shipping Economics andLogistics, Wilhelm-Herbst-Str. 5, 28359 Bremen, Germany

Preprint submitted to Elsevier Received: date / Accepted: date

arX

iv:1

406.

1928

v1 [

cs.G

T]

7 J

un 2

014

are safe to express their valuations for combinations of requests which usually differ from the sum of the valuationsof the respective individual requests. This is due to economies of scope and the effects of the compatibility betweentransport requests. That is, the cost to fulfill a request significantly depend on the amount and characteristics ofadditional requests a carrier has to fulfill as well, e. g. the requests can be executed at lower costs, if they can becombined in a well balanced tour. Carriers benefit from more efficient transport plans and an increased planningreliability which is enabled by bundle bids. What is more, the shipper has a significant potential to lower his totalprocurement costs due to bundle bidding.

Certainly, the advantage to fully express valuations for request combinations poses a challenge for the bidders. Abidder that does not submit bundle bids is no longer competitive. Determining the bidder’s preferences becomes morecomplicated, because the bidder has to evaluate a huge number of request combinations which grows exponentiallywith the number of tendered transport requests. In line with this, Plummer (2003) and Elmaghraby and Keskinocak(2004) report that real world carriers are often far from unlocking the full potential of bundle bidding. Furthermore,the results from classical auction theory with single-item based auctions are difficult to apply to combinatorial auctionswith multiple items (i.e., transport requests). Therefore, we are interested in providing decision support for bundlebidding in a combinatorial transport auction. We want to know: Which bidding strategies reduce the effort to calculatepreferences for all possible request combinations without lowering the chance to win in the auction? Furthermore, howcan we evaluate those bidding strategies by means of computational experiments and make them easily comparablefor other researchers?

If bidding carriers would know more about bidding strategies, they could gain several benefits. They could increasethe chance of their bids to be accepted during winner determination. At the same time, the effort to generated the bidscould be reduced. This could lower the entrance barriers for smaller carriers to participate in transport auctions.Likewise, if the transport auction is used for collaborative transport planning within a groupage system, more andmore diversified carriers are able to cooperate and it may become easier to lift synergies. On the other hand, theshipper benefits from more bids and in particular more competitive bids. Of course, from an economic point of view,more competitive bids finally lead to a more efficient usage of the available transport resources.

To get more insights into bidding in combinatorial transport auctions, we formulate a bid generation problem forless-than-truckload (LTL) requests and we study the interdependencies between valuations of request combinations.Based on this, we propose three different bidding strategies. These are applicable without knowledge about bidsof rivaling carriers, i.e., we assume truthful bidding. The performance of the strategies is evaluated by means of acomputational benchmark study. The study is setup such that it is easily repeatable by other researchers who mightwant to propose their own bidding strategies and compare them with the strategies proposed in this paper.

1.2. Review of the literature and contribution

Two main problems in combinatorial auctions are the bid generation problem (BGP) and the winner determinationproblem (WDP). In the BGP, a carrier (acting as a bidder) has to decide on which sets of requests to bid. This includesa valuation of possible bids and a selection of bids that are actually submitted. After bidding is complete the shipper(acting as an auctioneer) solves a WDP, i.e., the shipper chooses a set of winning bids from the set of all submittedbids. The WDP is usually modeled based on the set partitioning problem or the set covering problem. Designing abidding strategy depends on the used auction rules and the formulation of the WDP. Therefore, we start with a briefreview of the literature on WDP in transport auctions.

The formulation and solution of variants of the WDP in transport auctions still receives the most attention in therelevant literature. In this class of NP-hard combinatorial optimization problems, the task of an auctioneer is to decide,which of the submitted bundle bids should be accepted as winning bids such that the required transport services areprocured with minimal total cost. Integrating extensive real-world business constraints (e. g., Sheffi, 2004; Caplice andSheffi, 2006), considering transport quality by means of Pareto optimization (e. g., Buer and Pankratz, 2010b,a; Buerand Kopfer, 2014) or taking into account stochastic effects (Remli and Rekik, 2013) are discussed. For a literaturereview on winner determination problems see e.g. Abrache et al. (2007). Typically, WDP models assume that the setof bundle bids submitted by all bidders is a given input parameter.

Compared to the number of papers on the WDP only a few papers deal with the BGP. The following approachesdeal with innovative bidding strategies on individual, non-bundled requests. Figliozzi et al. (2007) study a dynamicvehicle routing scenario with bidding in a competitive environment. The sequential auction mechanism of van Duin

2

et al. (2007) also deals with short-term excess transport demand by means of an online planning approach. Requestsare periodically announced by the shipper and carriers may bid on single requests. Garrido (2007) propose a doubleauction format where shippers and carriers act as bidders at the same time. More double auction mechanisms fortransport requests are proposed by Huang and Xu (2013). Essentially, a multi-unit auction is performed for eachrequest. As the mechanisms are incentive compatible, the best strategy for each bidder is to bid truthfully. Xu andHuang (2013) advance their double auction mechanisms in order to cope with a dynamic environment. However, allthese approaches are concerned with bidding on individual requests and do not support bundle bidding.

Bundle bidding strategies may be distinguished between strategies that are independent of the rivaling bidders’valuations and strategies that do require information about the rivaling bidders’ valuations. For the latter, the amountof required information may vary significantly. While some strategies require complete information, i.e., the pref-erence function to calculate all bids, other estimate the prices of individual requests or a few request combinationsonly.

Complete information approaches require that carriers are willing to share sensitive data (e.g. utilization of vehiclefleet, cost rates, or operational constraints) with the auctioneer. Alternatively, the auctioneer may estimate suchinformation. In the end, the bid price is calculated by the auctioneer and not by the carrier (Chen et al., 2009;Beil et al., 2007). Possible scenarios are the ex ante existence of framework agreements that already specify whichresources a carrier has to provide; here, the auction is used for operational fine-tuning. Also, a setting were carriershave very high incentives to collaborate applies. For example, because the carriers’ do not represent independentcompanies but profit centers of a single company which are more open for sharing sensitive data, e.g. as consideredin Krajewska et al. (2008).

The following bidding strategies require some information (deterministic or stochastic) on the valuations of rival-ing carriers. Lee et al. (2007) propose a quadratically constrained formulation of a BGP. The model assumes a multipleround combinatorial auction. For each request, an ask prices announced by the auctioneer (see Kwon et al., 2005)is given. Although the assumptions about bundle valuations are not made by the bidders, the bidders’ require the askprices from auctioneer in order to decide about their own bid. The efficiency and effectiveness of solution techniquesthat compute a bundle bid are evaluated. However, the study does not go beyond that point and does not focus on theperformance of the computed bids in an auction setting. Bichler et al. (2009) analyse iterative combinatorial auctionswith respect to the allocative efficiency and utility distribution between the auctioneer and the bidders. Nevertheless,six different bidding strategies are presented and compared. These are based on a generic value model and the askprices announced by the auctioneer. Furthermore, the used WDP assumes, at most one bid per bidder may win. Forthese reasons, the focus is rather on the selection of bids as well as on the evaluation of the auction mechanism.

Triki et al. (2014) present a stochastic optimization model that requires assumptions about the cheapest pricefor each subset of requests offered by some rivaling bidders. These prices are considered stochastic. Stochasticoptimization is used to calculate a single bundle bid together with a bid price. This bid is expected to maximize thebidder’s profit. The model may be solved repeatedly, if multiple bundle bids are required. An et al. (2005) proposetwo domain independent bundle bidding strategies. The bidder’s valuation for each item as well as her synergyvalue for each pair of items are given. With it, the valuations for the most promising set of items are calculated.While one strategy requires information on the single-item valuation of rivaling bidders, the other strategy requiresno information on the valuations of rivaling bidders. Both strategies are compared, among others, with a strategy thatenumerates the complete set of bundle bids. At this, auctions with up to 10 items are considered.

Finally, there are bidding strategies that work independent of the valuations of rivaling carriers. The strategiesare truthful, because the carrier bids according to her true valuations. However, this is not always enforced by anincentive compatible auction mechanism for which truthful bidding is the dominant strategy. Rather, it is due to thecomplexity of the BGP in transport auctions which also makes systematic cheating a nontrivial task. Schönberger andKopfer (2004) study an auction-based request reallocation mechanism for a coalition of carriers. Tendered are thoserequests which are not profitable for an individual carrier. A bid is generated by solving a pickup and delivery selectionproblem which maximizes the carrier’s profit. All new requests in the solution are combined into a bundle bid and thecharged price equals the marginal increase of the carrier’s profit. The evaluation focuses on the increase in the totalprofit of a coalition of carriers but not on the performance of the bidding strategy. Song and Regan (2005) considerthe procurement of full truckload pick up and delivery requests. To generate bids, a truckload vehicle routing problemwhich minimizes the empty movement costs is solved. Each tour of this solution is interpreted as a candidate bid; thebid price corresponds to the empty movement costs. To select bids from the set of candidate bids that are actually

3

submitted to the auctioneer, a set covering problem is solved which minimizes the sum of the prices of the selectedbids and ensures that on each tendered request at least on bid is submitted. The latter constraint might be unnecessaryin other transport auction scenarios, as it is often not required the each carrier provides a bids on each request. Toevaluate the strategies, two agents compete against each other in a computational simulation of an auction with fourto ten requests. Wang and Xia (2005) explain first-order and second-order synergies between set of requests and theypresent the notion of a carrier’s set of optimal bundle bids. However, the synergy notions are neither implemented inthe two bidding strategies and therefore not evaluated. Chang (2009) considers the BGP in a full-truckload scenarioand models the problem as a variant of the minimum cost flow problem with pairwise synergies between arcs. Theevaluation focuses on the performance of the proposed method to solve the synergistic flow problem. However, it isnot studied how competitive the generated bids are in an auction setting, which is the focus of our paper.

The contribution of this paper is to propose three strategies for bidding in a combinatorial transport auction thatrequire no information about the valuation of rivaling bidders. The competitiveness of the bundle bids generated bythe strategies are evaluated by means of a computational benchmark study. In detail:

1. An exact bundle bidding strategy is proposed. With this strategy, a bidder always achieves the same outcome asbidding on each element of the powerset of the set of requests.

2. A first heuristic bidding strategy is presented that restricts the search in the bid space by using pairwise synergiesbetween requests based on modified Clarke and Wright savings values.

3. A second heuristic strategy uses the capacitated p-median problem as a clustering approach to identify promis-ing request combinations and to restrict the search in the bid space even stronger.

4. None of the strategies requires information about valuations of rivaling carriers. This is advantageous, because itis very difficult for a bidder to estimate what prices rivaling carriers may charge for which request combinations.

5. The computational evaluation by means of test instances (see electronic appendix) is set up such that otherresearcher can easily compare their own strategies with the results of the strategies presented in this paper.

1.3. OrganizationThe remaining paper is organized as follows. In Section 2, we formulate a basic bid generation problem from the

auctioneer and from the carrier point of view. In Section 3 we present an exact bidding strategy for a carrier. Theheuristic bidding strategies are introduced in Section 4. The performance of the three bidding strategies is evaluatedby means of a computational study in Section 5. Section 6 concludes the paper and gives an outlook on future researchpossibilities.

2. A bundle bid generation problem

A bundle bid generation problem (BGP) has to be solved by each carrier who participates in a combinatorialtransport auction. A carrier c has to decide on which subsets of transport requests to bid and which price to charge.The goal of carrier c is to maximize her revenue won in the auction, i.e., the sum of prices of her bundle bids thatare accepted by the auctioneer. What makes the BGP difficult is the distributed nature of the problem. On the onehand, there are rivaling carriers that submit competing bundle bids whose composition and prices are unknown tocarrier c, i.e., there is asymmetric information. On the other hand, the authority to decide which bids to accept aswinning bids lies not with carrier c, but with the auctioneer, i.e., there are multiple decision makers with a distributeddecision-making authority.

2.1. Assumptions on bidding and winner determination in a combinatorial transport auctionThe process of a transport auction is shown by Figure 1. Rectangles represent optimization problems, rounded

rectangles show problem data, and edges represent input-output relations. A directed edge from a data box to aproblem box indicates that the data is required as input parameter to solve the problem. Vice versa, an edge from aproblem box to a data box indicates that the solution of the optimization problem provides the data. So, the auctioneerannounces the set R of tendered requests which is used as input data by the bidders. Each of the n bidders solves

4

auctioneertenders

requests R

BGP ofcarrier 1

BGP ofcarrier . . .

BGP ofcarrier n

bundle bidsB1

bundle bidsB...

bundle bidsBn

auctioneer’swinner

determinationproblem

winning bids WW ⊂ B1 ∪ . . . ∪ Bn

Figure 1: Selected problems and decision makers in combinatorial transport auctions.

her BGP, the outcome for each bidder i is a set Bi of bundle bids. To decide which bundle bid is a winning bid,the auctioneer solves a WDP; the solution of the WDP is a set W of winning bids, which is a subset of the setB := B1 ∪ . . . ∪ Bn of all bids in the auction.

We take the point of view of a single carrier and study her bidding strategy, i.e., the way she solves the BGP. Thehatched area in Figure 1 highlights the problem and the data which are controlled by this carrier. The main problem isof course, that the success of the bidding strategy depends both on the rivaling bids and on the winner determinationdecision of the auctioneer, which both lie not in the circle of influence of the carrier at hand.

With respect to the auction setting, we assume the following.

• Bundle bidding is allowed. A bundle bid b is defined as a triple (cb, pb,Rb). The carrier cb offers to performthe set Rb of requests for the price pb ∈ N. The set Rb is any non-empty subset of the set of tendered requests(Rb ⊆ R,Rb , ∅). Let P(R) denote the powerset of R without the empty set.

• A carrier may bid on each element of P(R). The cardinality of P(R) is 2|R| − 1.

• Bidders bid truthfully. The price of a bundle bid reflects the true valuations of the bidder and there is no strategicbidding. From this follows, that each carrier looks only on her own and she makes no assumptions on the bidsof rivaling carriers.

• Bidders submit OR-bids. Nisan (2000) and Day and Raghavan (2009) discuss several bidding languages forcombinatorial auctions. Here, a bidding language is characterized by logical connections between the bundlebids like OR or XOR, which allow the bidders to express different preference for combinations of bundle bids.We use OR-bids, i.e., the auctioneer is allowed to select any subset of the submitted bids.

• The carriers use cost plus pricing. The price of a bundle bid is proportional to the costs of performing therequests. Without loss of generality, we assume a markup of zero percent, i.e., the cost for performing a set ofrequests are equal to the price charged by a carrier to execute these requests.

• The WDP of the auctioneer is formulated as the set covering problem. It is given by formula (1).

minW⊆B

f a =∑

b∈Wpb such that

⋃

b∈WRb = R (1)

The auctioneer’s task is to select a minimum cost set W of winning bids (W ⊆ B) that covers each tenderedrequest r ∈ R at least once. A set W of bids covers a request r ∈ R if it contains at least one bid b ∈ W withr ∈ Rb.

For the transport auction scenario at hand we prefer a WDP-formulation based on the set covering model to otherformulations based on set packing or set partitioning models. WDP formulations based on the set packing model

5

Figure 1: Selected problems and decision makers in combinatorial transport auctions.

her BGP, the outcome for each bidder i is a set Bi of bundle bids. To decide which bundle bid is a winning bid,the auctioneer solves a WDP; the solution of the WDP is a set W of winning bids, which is a subset of the setB := B1 ∪ . . . ∪ Bn of all bids in the auction.

We take the point of view of a single carrier and study her bidding strategy, i.e., the way she solves the BGP. Thehatched area in Figure 1 highlights the problem and the data which are controlled by this carrier. The main problem isof course, that the success of the bidding strategy depends both on the rivaling bids and on the winner determinationdecision of the auctioneer, which both lie not in the circle of influence of the carrier at hand.

With respect to the auction setting, we assume the following.

• Bundle bidding is allowed. A bundle bid b is defined as a triple (cb, pb,Rb). The carrier cb offers to performthe set Rb of requests for the price pb ∈ N. The set Rb is any non-empty subset of the set of tendered requests(Rb ⊆ R,Rb , ∅). Let P(R) denote the powerset of R without the empty set.

• A carrier may bid on each element of P(R). The cardinality of P(R) is 2|R| − 1.

• Bidders bid truthfully. The price of a bundle bid reflects the true valuations of the bidder and there is no strategicbidding. From this follows, that each carrier looks only on her own and she makes no assumptions on the bidsof rivaling carriers.

• Bidders submit OR-bids. Nisan (2000) and Day and Raghavan (2009) discuss several bidding languages forcombinatorial auctions. Here, a bidding language is characterized by logical connections between the bundlebids like OR or XOR, which allow the bidders to express different preference for combinations of bundle bids.We use OR-bids, i.e., the auctioneer is allowed to select any subset of the submitted bids.

• The carriers use cost plus pricing. The price of a bundle bid is proportional to the costs of performing therequests. Without loss of generality, we assume a markup of zero percent, i.e., the cost for performing a set ofrequests are equal to the price charged by a carrier to execute these requests.

• The WDP of the auctioneer is formulated as the set covering problem. It is given by formula (1).

minW⊆B

f a =∑

b∈Wpb such that

⋃

b∈WRb = R (1)

The auctioneer’s task is to select a minimum cost set W of winning bids (W ⊆ B) that covers each tenderedrequest r ∈ R at least once. A set W of bids covers a request r ∈ R if it contains at least one bid b ∈ W withr ∈ Rb.

For the transport auction scenario at hand we prefer a WDP-formulation based on the set covering model to otherformulations based on set packing or set partitioning models. WDP formulations based on the set packing model

5

(see, e.g. de Vries and Vohra, 2003) are better suited for scenarios where the auctioneer wants to sell goods,rather than to buy goods like in the scenario at hand. WDP formulations based on the set partitioning modelrequire that each transport request is part of exactly one winning bundle bid. However, given a set of bundlebids B, the total cost of the optimal set W of winning bids under the set covering formulation is guaranteed tobe lower or at most equal than the total cost of the optimal set partitioning solution for B (de Vries and Vohra,2003; Song and Regan, 2005; Buer and Pankratz, 2010a). This is clearly preferred by the auctioneer.

2.2. Assumptions on the tendered transport requests and the bidding carrier

We consider a shipper whose customers have to be serviced from the shipper’s warehouse by LTL requests. Withrespect to the structure of the transport requests and the carrier at hand, our transport scenario is based in the well-known capacitated vehicle routing problem (CVRP) with the following modifications and additional assumptions.

• The shipper operates a single warehouse w from which a set V of customers has to be served.

• The relevant carrier operates a fleet of homogeneous vehicles. The capacity of each vehicle is cap.

• The shipper tenders a set R of requests. Each request r ∈ R requires a pickup of load at the shipper’s warehousew and its delivery to a customer location i ∈ V . We assume there is exactly one request for each customer, i.e.,|V | = |R|. Therefore, requests may be also identified in terms of the involved customer location.

• Like in the CVRP, we assume LTL requests, i.e., each request r ∈ R includes a load lr (0 ≤ lr ≤ cap) which hasto be fulfilled by a single vehicle. However, unlike the CVRP, a vehicle has to pickup the load at the shipper’swarehouse w and not at the carrier’s depot.

• Each tour (d,w, . . . , d) of a vehicle starts and ends at the home depot d of the carrier. Furthermore, a vehicledrives immediately to the warehouse w after leaving the depot to pickup the loads required to service thedemands of the customers in the tour. In each tour, the home depot is visited twice, the warehouse is visitedonce, and each customer is visited at most once.

• The carrier estimates costs by focusing on the tendered requests only, existing requests and possible futurerequests by other shippers which may be acquire during the planning period are not considered.

Combinatorial transport auctions are often used to procure framework contracts for a longer period of severalmonth. In such a case the tendered requests may be simply interpreted, for example, as recurring requests during theplanning period.

With respect to the rivaling carriers, we make the same assumptions. Therefore, the competitive differences ofthe carriers are only caused by their different depot locations in relation to the shipper’s warehouse and the shipper’scustomers locations. For our computational study, we consider these aspects during the generation of test instancesas described in Section 5.2. Due to the instance generation procedure, we can simplify the problem by setting thedistance of the shipper’s warehouse and the carrier’s depot at hand to zero. Without loss of generality, this allows theuse of sophisticated standard solution procedures which eases the computational experiments.

2.3. Bid generation problem of an omniscient carrier

Provided that the carrier at hand knew the bundle bids submitted by her rivaling carriers she could examine thechances of winning in the auction by means of the model given by the bid generation problem of an omniscientcarrier (BGPO) as defined in (2) – (4). Although it is not possible for the carrier to use this model directly – due toinformation asymmetry and lacking decision authority – we use to clarify the optimization problem and to evaluatethe proposed bidding strategies (cf. Section 5).

6

For the BGPO, the set of bundle bids B is partitioned into two subsets Bc and Br with Bc∪Br = B and Bc∩Br = ∅.Let Bc and Br denote the set of bids submitted by carrier c and by all rivals of carrier c, respectively.

lex minW⊆B

f a =∑

b∈Wpb, (2)

lex maxBc

f b =∑

b∈Bc∩W

pb, (3)

s. t.⋃

b∈WRb = R. (4)

In the BGPO there are two objective functions which are lexicographically ordered. If carrier c wants to simulatethe outcome of the auction and design her bundle bids appropriately, she first of all has to take into account, that theshipper will select bundle bids such that his total procurement costs are minimal (2). Under this precondition, thecarrier may then generate a set of bundle bids Bc which maximizes her total revenue won in the auction, representedby objective function f carrier c. Clearly, both decisions are interdependent. Of course, in practice the set of bundle bidsBr submitted by rivaling carriers is usually unknown to carrier c which makes it even more challenging to formulatean adequate optimization model for the BGP taking into account asymmetric information.

With respect to the BGPO the question is which set of bids Bc maximizes the winning revenue by carrier c? Underthe assumption of truthful bidding, the best chance to come out on top of the bids of rivaling carriers is simply tosubmit all possible bundle bids (brute-force strategy), i.e., a bid on each request combination in P(R). This strategy,however, is computational infeasible even for a small number of tendered request, because to calculate the bid price aNP-hard vehicle routing problem has to be solved for all 2|R| − 1 request combinations.

3. An optimal bidding strategy based on elementary request combinations

We propose a bidding strategy denoted as elementary bundle bid search (EBBS). The outcome of EBBS is denotedoptimal (or exact), because it provides the same result as a brute-force strategy that bids on each request combinationin P(R).

3.1. Bidding strategy and bidding spaceA bidding strategy φ in a combinatorial auction is a method that generates a set B of bundle bids. Considering

a truthful carrier c, a bidding strategy φc is a function on the set R of requests which outputs a set Bc of carrier’s cbundle bids b := (c,Rb, pb), b ∈ Bc. At this, Rb is a subset of R and the bid price pb is calculated according to a costfunction p(Rb) ∈ N. Thus, the bidding function of c may be formally defined as in (5). For future reference, we omitthe superscript c for ease of notation because we focus on the individual carrier c only.

φc : R→ {b ∈ Bc|b = (c,Rb, pb) with Rb ∈ P(R) and p : R→ N}. (5)

The bidding strategy φ generates atomic bids only (Nisan and Ronen, 2001). Because the winner determinationproblem assumes OR-bids, additional logical dependencies between bundle bids do not have to be modeled. That is,the auctioneer can accept any subset of the submitted bundle bids; on the other hand, a bidder cannot enforce logicalconstraints between atomic bids in the fashion of ’if you accept one of these bids then you cannot accept that bid.’

The bid space is P(R) ×N with a cardinality of |P(R)|. There will be at most one bid on each request combinationof P(R). In particular, it is not reasonable for a carrier to generate multiple bundle bids with different prices on thesame request combination A, A ∈ P(R). As each of these bids offers an identical set A of requests, the bid with thelowest price dominates all other bids on A. Dominated bids will never be part of an optimal solution of the WDP.

The carrier uses her cost function p(A) in order to calculate a bid price for A ∈ P(R). The price charged forexecuting the requests R is equal to the minimum cost solution of the VRP described in Section 2.2. Form this bidprice function, three plain bid strategies arise naturally. Given a set R of tendered requests, a truthful bidder could:

1. solve the VRP for R and submit a single bid on R only,

2. solve the VRP for R and submit a bid on each generated tour, or

7

3. solve the VRP for each request combination in P(R) and submit a bid on each element of P(R).

As introduced in Section 2.3, the auctioneer minimizes his total procurement costs by solving the WDP givenby formula (1). The first strategy generates a single bid which is as efficient as it can be for the carrier at hand.Nevertheless, it will be not very competitive in the WDP as it somewhat ignores the rivaling carriers which alsosubmit bids. In order to win any revenue in the auction, the bid of our carrier has to offer the lowest total costs forR compared to all possible combinations of bundle bids which also cover R. This seems highly unlikely for largertransport auctions.

The second strategy generates several bids which inhibit the same expressiveness as the single bid of the firststrategy. If all bids of the second strategy are accepted by the auctioneer, the assigned requests and the won revenueare identical. With this in mind, the third strategy will be at least as successful as the first and the second strategy. Onthe one hand, simply because the set of the generated bids are supersets of the bids generated by the fist and the secondstrategy. On the other hand, because it generates more bids which offer a higher chance to be a good match with bidsof rivaling carriers. Of course, the computational effort for solving 2|R| − 1 vehicle routing problems is prohibitivehigh considering the number of requests in many real world auctions. In the following, we develop an approach whichguarantees the same results as the brute force strategy but generates significantly less bids.

3.2. Relations between the valuations of sets of requests

Some terminology with respect to the relationships between costs for performing sets of requests are introduced.First, performing an additional request always increases (or at least does not decrease) the total costs of a carrier. LetA and B be two sets of requests (A, B ⊆ R). If A ⊆ B it follows that p(A) ≤ p(B). This relationship is denoted as freedisposal. The term originates from the point of view of an auctioneer that does not have to pay a fee for disposingpurchased items wherefore the auctioneer’s utility of receiving additional items given the same price (ceteris paribus)never decreases. Due to the pricing function p, free disposal is a guaranteed characteristic of the scenario at hand.

Second, Nisan (2000) introduced subadditive, superadditive, and additive valuations between disjoint sets of itemsin combinatorial auctions. We apply their terminology with respect to the valuation of request combinations from thepoint of view of a single carrier. Let A and B be two disjoint subsets of requests (A, B ⊆ R, A ∩ B = ∅).• If p(A ∪ B) < p(A) + p(B), the valuation of A and B is strictly cost subadditive.

• If p(A ∪ B) > p(A) + p(B), the valuation of A and B is strictly cost superadditive.

• If p(A ∪ B) = p(A) + p(B), the valuation of A and B is cost additive.

Under the conditions of the vehicle routing problem at hand, there are only additive and subadditive valuationsbetween sets of requests but no superadditive valuations. Consider Figure 2 with three requests a, b, c represented bythree customers nodes. Let A and B be two sets of requests with A = {a} and B = {b, c}. It is clear, that p({a, b, c}) iseither smaller than p({a}) + p({b, c}), e.g., when all three request can be performed in a single tour and a depot edgecan be removed. Or, if two tours are required to perform the requests due to capacity constraints, the total costs areunchanged. Therefore, p({a, b, c}) > p({a}) + p({b, c}) is not possible, i.e., the valuation is not superadditive.

3.3. Elementary request combinations and elementary bundle bids

We show that given free disposal and the absence of superadditive valuations, it is sufficient to bid on subadditiverequest combinations. We start by introducing the concept of elementary request combinations. A subset E of theset of requests R is denoted as elementary if this set of requests can be performed by at most one vehicle, i.e.,∑

r∈E dr ≤ cap with E ∈ P(R). The set of all elementary request combinations is:

E := {E ∈ P(R) |∑

i∈E

di ≤ cap}. (6)

Consequently, the set of non-elementary request combinations is defined as

E := P(R) \ E. (7)

8

depot

a

c

b

warehouse free disposal:p({b}) ≤ p({b, c})p({c}) ≤ p({b, c})

subadditive valuations only:p({b}) + p({c}) ≥ p({b, c})p({a}) + p({b, c}) ≥ p({a, b, c})

Figure 2: Example of relations between the valuations of request combinations

Table 1: Example of inferred information on bundle bids

Elementary bundle bids Inferred bundle bids

Rb used capacity p(Rb) R′ inferred prices for R′

{a} 0.8 pa {a,c} pa + pc

{b} 0.2 pb {a,d} pa + pd

{c} 0.5 pc {c,d} pc + pd

{d} 0.7 pd {a,b,c} min{pA + pb + pcC; pab + pc}{a,b} 1.0 pab {a,b,d} min{pab + pd; pa + pb + pd; pa + pbd}{b,c} 0.7 pbc {a,c,d} min{pab + pd; pa + pb + pD; pA + pbd}{b,d} 0.9 pbd {b,c,d} min{pab + pd; pa + pb + pd; pa + pbd}

{a,b,c,d} min{pa + pb + pc + pd; . . . ; pa + pc + pbd}

A bundle bid b = (pb, cb,Rb) on an elementary request combination (Rb ∈ E) is an elementary bid. A bid on anon-elementary request combination in E is a non-elementary bid. Note, an elementary request combination for onecarrier needs not to be elementary from another carrier’s point of view.

All elementary request combinations exhibit subadditive valuations. On the other hand, non-elementary requestcombinations exhibit additive valuations. Superadditive valuations are not present in the scenario at hand. Therefore,an auctioneer is able to infer the prices of all non-elementary bids of a carrier given that the carrier submits allelementary bundle bids. To infer the price for performing a set R′ of requests (R′ ⊆ R) the auctioneer has to select acost minimal set W of bids whose request combinations is a partition of R′, i.e., the auction has to solve the followingset partitioning problem:

min∑

b∈Wpb such that

⋃

b∈WRb = R′ and

⋂

b∈WRb = ∅ (8)

Table 1 shows an example. The auctioneer tenders four requests a, b, c, and d; seven subadditive bundle bids havenbeen submitted by a carrier, the required relative capacity (with cap = 1) is given in the second column and the bidprice is given in the column p(Rb). Now, the auctioneer can infer from these seven subadditive bids the prices of allremaining bids by solving set partitioning problems.

For these reasons, it is sufficient if the bidding strategy φ of a carrier takes into account all elementary requestcombinations E ⊂ P(R) only, because the auctioneer is able to infer the bid prices of all non-elementary bids which isalways implicitly done during winner determination. That is, a bid strategy can focus on those request combinationswhich can be executed by a single vehicle. This brings up another question. Do we have to bid on each elementaryrequest combination or can we exclude additional request combinations a priori, e.g., is it reasonable to focus onrequest combinations which are performed in a highly efficient tour?

9

Figure 2: Example of relations between the valuations of request combinations

Table 1: Example of inferred information on bundle bids

Elementary bundle bids Inferred bundle bids

Rb used capacity p(Rb) R′ inferred prices for R′

{a} 0.8 pa {a,c} pa + pc

{b} 0.2 pb {a,d} pa + pd

{c} 0.5 pc {c,d} pc + pd

{d} 0.7 pd {a,b,c} min{pA + pb + pcC; pab + pc}{a,b} 1.0 pab {a,b,d} min{pab + pd; pa + pb + pd; pa + pbd}{b,c} 0.7 pbc {a,c,d} min{pab + pd; pa + pb + pD; pA + pbd}{b,d} 0.9 pbd {b,c,d} min{pab + pd; pa + pb + pd; pa + pbd}

{a,b,c,d} min{pa + pb + pc + pd; . . . ; pa + pc + pbd}

A bundle bid b = (pb, cb,Rb) on an elementary request combination (Rb ∈ E) is an elementary bid. A bid on anon-elementary request combination in E is a non-elementary bid. Note, an elementary request combination for onecarrier needs not to be elementary from another carrier’s point of view.

All elementary request combinations exhibit subadditive valuations. On the other hand, non-elementary requestcombinations exhibit additive valuations. Superadditive valuations are not present in the scenario at hand. Therefore,an auctioneer is able to infer the prices of all non-elementary bids of a carrier given that the carrier submits allelementary bundle bids. To infer the price for performing a set R′ of requests (R′ ⊆ R) the auctioneer has to select acost minimal set W of bids whose request combinations is a partition of R′, i.e., the auction has to solve the followingset partitioning problem:

min∑

b∈Wpb such that

⋃

b∈WRb = R′ and

⋂

b∈WRb = ∅ (8)

Table 1 shows an example. The auctioneer tenders four requests a, b, c, and d; seven subadditive bundle bids havenbeen submitted by a carrier, the required relative capacity (with cap = 1) is given in the second column and the bidprice is given in the column p(Rb). Now, the auctioneer can infer from these seven subadditive bids the prices of allremaining bids by solving set partitioning problems.

For these reasons, it is sufficient if the bidding strategy φ of a carrier takes into account all elementary requestcombinations E ⊂ P(R) only, because the auctioneer is able to infer the bid prices of all non-elementary bids which isalways implicitly done during winner determination. That is, a bid strategy can focus on those request combinationswhich can be executed by a single vehicle. This brings up another question. Do we have to bid on each elementaryrequest combination or can we exclude additional request combinations a priori, e.g., is it reasonable to focus onrequest combinations which are performed in a highly efficient tour?

9

D1

qA = 0.2 W

qB = 0.2

qC = 0.2

D2

qD = 0.2

Figure 3: Inefficient tours might still have a good chance of being selected as winning bid.

3.4. Focusing solely on efficient tours is not reasonable

The question is whether it is necessary to compute all elementary bids in order to achieve an optimal solution ofthe BGP or does it suffice to submit a subset of the elementary bids to achieve the same outcome. The question is ofpractical importance due to the combinatorial nature of the BGP. For a transport auction with fifty requests auctionedand an average size five requests per elementary bid there are

(505

)or more than 2,000,000 elementary bids; with an

average size of ten requests per elementary bid there are already more than 10,000,000,000 elementary bids.A natural thought to reduce the number of bids is to focus on those (elementary) request combinations that can

be combined in highly efficient tours. The idea is that an efficient tour leads to lower cost per request and therefore abid on an efficient tour is probably more competitive, i.e., it has a higher chance to be selected as a winning bid. Onthe other hand, bids on (elementary) request combinations with a low utilization of the vehicle capacity or bids onrequests which are geographically distributed in an unfavorable manner are considered as inefficient. Although theseconsiderations are reasonable, one has to keep in mind that ultimately the auctioneer decides, by solving the WDP,which subset of the submitted bids will be the set of winning bids. Therefore, efficiency of bids is predominantlyjudged by the auctioneer taking into account

• the objective function and the constraints of the used winner determination problem and

• all bundle bids submitted by all rivaling carriers.

For the auctioneer, a bundle bid is efficient when it contributes to minimize the WDP’s total procurement costs. This,however, does not necessarily imply that low utilized pendular tour might not have better chances to be selected as awinning bid.

Consider Figure 3 with two carriers c1, c2 and five bids b1, . . . , b5. Carrier c2 bids on the set of requests {A, B,C,D}and {B,C,D}. On the other hand, carrier c1 bids on {A} and {A, B,C,D}. Bid b3 = (p3, c1, {A}) can be considered asinefficient. However, it complements very well bid b2 of carrier 1. Therefore, it has a good chance to being selectedas winning bid. Assume, carrier c2 with depot D2 submits a bid on request A which leads to a pendular tour

b1 = (p1, c2, {A, B,C,D})b2 = (p2, c2, {B,C,D})b3 = (p3, c1, {A})b4 = (p4, c1, {A, B,C})b5 = (p5, c1, {A, B,C,D})Assume, carrier c1 with depot D1 submits a bid on request A which leads to a pendular tour with a low utilized

vehicle capacity of only twenty percent. Such a pendular tour is usually considered inefficient. From the point ofview of carrier, the construction of efficient bids does not necessarily increase her chance to win in the auction. As thecarriers is not aware of the actions of her rivals, the construction of efficient bids For a carrier, it is not a good strategyto focus on the construction of feasible bids only.

From these considerations, it follows that focusing on the subset of elementary bundle bids on efficient (elemen-tary) tours from the carrier point of view cannot be the optimal bidding strategy. Of course, this does not exclude anadoption of these ideas within heuristic bidding strategies. Finally, like in most auction scenarios and in accordance

10

Figure 3: Inefficient tours might still have a good chance of being selected as winning bid.

3.4. Focusing solely on efficient tours is not reasonable

The question is whether it is necessary to compute all elementary bids in order to achieve an optimal solution ofthe BGP or does it suffice to submit a subset of the elementary bids to achieve the same outcome. The question is ofpractical importance due to the combinatorial nature of the BGP. For a transport auction with fifty requests auctionedand an average size five requests per elementary bid there are

(505

)or more than 2,000,000 elementary bids; with an

average size of ten requests per elementary bid there are already more than 10,000,000,000 elementary bids.A natural thought to reduce the number of bids is to focus on those (elementary) request combinations that can

be combined in highly efficient tours. The idea is that an efficient tour leads to lower cost per request and therefore abid on an efficient tour is probably more competitive, i.e., it has a higher chance to be selected as a winning bid. Onthe other hand, bids on (elementary) request combinations with a low utilization of the vehicle capacity or bids onrequests which are geographically distributed in an unfavorable manner are considered as inefficient. Although theseconsiderations are reasonable, one has to keep in mind that ultimately the auctioneer decides, by solving the WDP,which subset of the submitted bids will be the set of winning bids. Therefore, efficiency of bids is predominantlyjudged by the auctioneer taking into account

• the objective function and the constraints of the used winner determination problem and

• all bundle bids submitted by all rivaling carriers.

For the auctioneer, a bundle bid is efficient when it contributes to minimize the WDP’s total procurement costs. This,however, does not necessarily imply that low utilized pendular tour might not have better chances to be selected as awinning bid.

Consider Figure 3 with two carriers c1, c2 and five bids b1, . . . , b5. Carrier c2 bids on the set of requests {A, B,C,D}and {B,C,D}. On the other hand, carrier c1 bids on {A} and {A, B,C,D}. Bid b3 = (p3, c1, {A}) can be considered asinefficient. However, it complements very well bid b2 of carrier 1. Therefore, it has a good chance to being selectedas winning bid. Assume, carrier c2 with depot D2 submits a bid on request A which leads to a pendular tour

b1 = (p1, c2, {A, B,C,D})b2 = (p2, c2, {B,C,D})b3 = (p3, c1, {A})b4 = (p4, c1, {A, B,C})b5 = (p5, c1, {A, B,C,D})Assume, carrier c1 with depot D1 submits a bid on request A which leads to a pendular tour with a low utilized

vehicle capacity of only twenty percent. Such a pendular tour is usually considered inefficient. From the point ofview of carrier, the construction of efficient bids does not necessarily increase her chance to win in the auction. As thecarriers is not aware of the actions of her rivals, the construction of efficient bids For a carrier, it is not a good strategyto focus on the construction of feasible bids only.

From these considerations, it follows that focusing on the subset of elementary bundle bids on efficient (elemen-tary) tours from the carrier point of view cannot be the optimal bidding strategy. Of course, this does not exclude anadoption of these ideas within heuristic bidding strategies. Finally, like in most auction scenarios and in accordance

10

request A B A ∪ B Z1 Z2 Z3

1 x x x x2 x x x x3 x x x4 x

p(.) 2 1 3 3 2.1 1.1

Table 2: A ∪ B is redundant and less competitive than A and B.

to many real world auctions, we assume the bidding carrier does not have knowledge about the bids of the rivalingcarriers. Therefore we cannot use this information in order to identify a subset of the set of elementary bids whichleads to an optimal solution for the carrier. For the optimal bidding strategy under the assumptions of in the transportauction at hand a carrier is required to generate all elementary bundle bids.

3.5. A tree search approach for generating elementary bundle bids (EBBS)

We propose a bidding strategy which is denoted as elementary bundle bidding strategy (EBBS). EBBS consistsof two phases. An overview is given by Algorithm 1. In the first phase (cf. Section 3.5.1), a set of all elementaryrequest combinations is generated. In the second phase (cf. Section 3.5.2), the price of each request combination iscalculated.

Algorithm 1: Elementary bundle bidding strategy (EBBS) – OverviewInput: depot node d of carrier c, customer nodes V , demand di, ∀i ∈ V

set of bundle bids: B← {}R ← generateAllElementaryRequestCombinations(R, d) // with R ⊂ P(R)foreach S ∈ R do

p← calculateBidPriceForAnElementaryRequestCombination(S , d)b← (c, S , p)B← B ∪ {b}

endreturn B

3.5.1. Generation of all elementary request combinationsTo generate all elementary request combinations E,E ⊂ P(R) binary tree search is used (cf. Figure 4). Assume

all requests are indexed in non-ascending order of their capacity demand, i.e., the sequence of requests (r1, r2, . . . , r|R|)implies that dri ≤ dri+1 for all ri ∈ R.

Each node of the binary tree represents a subset S of the set of requests (S ⊆ R). The root node of the binary treerepresents an empty set of requests. The number of edges k on a path from the root to a given node i is denoted as thelevel k of node i, the root node is level 0. The maximum level or the height of the binary tree is |R|. Let i be a parentnode on level k that represents S . The left child of node i represents the set of requests S ∪ {rk+1} where rk+1 is therequest indexed with k + 1. The right child of node i represents S in turn. In other words, on level k of the tree it isdecided whether request rk+1 is added to the set of requests represented by the parent node or not.

Because the requests are indexed in non-ascending order of their capacity demand, we can prune as soon as thecapacity demand of the set S of requests represented by a left node exceeds the vehicle capacity, i.e., we prune if∑

ri∈S qi ≥ cap. If the required capacity of the request represented by a node l exceeds the vehicle capacity capthe node l and all its descendant nodes do not have to be visited, because they do not represent elementary requestcombinations. The leaf nodes of the generated tree represent all elementary request combinations.

11

request i b c d ademand di 0.2 0.5 0.7 0.8

{a, b}0.9

{b}0.2

{a}0.8

{}

db + dc + dd

= 1.4 > 1.0{b, d}

0.9{b} {d} {}

{b, c}0.7

{b}0.2

{c}0.5

{}0.0

{b}0.2

{}0.0

b b

c c c c

d d d dd d d d

a aa a

a aa a

Figure 4: Binary tree search approach to generate all elementary request combinations.

3.5.2. Calculation of bid pricesAccording to the assumptions stated in Section 3.1 (in particular, depot d equals warehouse w), the price of bundle

bid equals the cost of performing all the requests of the bid. By definition, elementary request combinations can beperformed in a single tour, therefore we are not required to solve a vehicle routing problem but a computationally lesschallenging traveling salesman problem. The problem of calculating a bid price p for an elementary set of requestsS ∈ E is to find the traveling salesman tour for S of minimum cost.

4. Heuristic bidding strategies based on promising sets of elementary transport requests

Although intelligent transportation decision support systems disseminate, the optimal bidding strategy remainscomputationally challenging with an increasing number of tendered requests. Therefore, we propose two additionalheuristic bidding strategies. They are denoted as pairwise synergies based request clustering (PSC) and capacitatedp-median based request clustering (CPMC). Both seek to identify a subset of elementary bids which are promising.Section 4.1 describes the common elements of both bidding strategies, while Section 4.2 presents PSC and Section 4.3describes CPMC. Two random strategies are introduced in Section 4.4 for evaluation purposes.

4.1. General procedureThe basic idea of the heuristic bidding strategies is to generate only those elementary request combinations that are

supersets of promising request combinations (see Algorithm 2 for an overview). The bid price calculation is identicalto EBBS, however, the generation of bundle bids differs for the heuristic strategies. A set P of promising elementaryrequest combinations (P ⊂ E) is identified by first calling the procedure cluster. The following steps slightly differcompared to the procedure of EBBS in Section 3.5:

• The procedure generateAllElementaryRequestCombinations is called for each S ∈ P.

• S is used as the root node of procedure generateAllElementaryRequestCombinations.

• The outcome of generateAllElementaryRequestCombinations is a set T of requests that is a superset ofS (T ⊃ S ,T ∈ E).

• The set union of all generated request combinations is denoted as R (R ⊂ E).

12

Figure 4: Binary tree search approach to generate all elementary request combinations.

3.5.2. Calculation of bid pricesAccording to the assumptions stated in Section 3.1 (in particular, depot d equals warehouse w), the price of bundle

bid equals the cost of performing all the requests of the bid. By definition, elementary request combinations can beperformed in a single tour, therefore we are not required to solve a vehicle routing problem but a computationally lesschallenging traveling salesman problem. The problem of calculating a bid price p for an elementary set of requestsS ∈ E is to find the traveling salesman tour for S of minimum cost.

4. Heuristic bidding strategies based on promising sets of elementary transport requests

Although intelligent transportation decision support systems disseminate, the optimal bidding strategy remainscomputationally challenging with an increasing number of tendered requests. Therefore, we propose two additionalheuristic bidding strategies. They are denoted as pairwise synergies based request clustering (PSC) and capacitatedp-median based request clustering (CPMC). Both seek to identify a subset of elementary bids which are promising.Section 4.1 describes the common elements of both bidding strategies, while Section 4.2 presents PSC and Section 4.3describes CPMC. Two random strategies are introduced in Section 4.4 for evaluation purposes.

4.1. General procedureThe basic idea of the heuristic bidding strategies is to generate only those elementary request combinations that are

supersets of promising request combinations (see Algorithm 2 for an overview). The bid price calculation is identicalto EBBS, however, the generation of bundle bids differs for the heuristic strategies. A set P of promising elementaryrequest combinations (P ⊂ E) is identified by first calling the procedure cluster. The following steps slightly differcompared to the procedure of EBBS in Section 3.5:

• The procedure generateAllElementaryRequestCombinations is called for each S ∈ P.

• S is used as the root node of procedure generateAllElementaryRequestCombinations.

• The outcome of generateAllElementaryRequestCombinations is a set T of requests that is a superset ofS (T ⊃ S ,T ∈ E).

• The set union of all generated request combinations is denoted as R (R ⊂ E).

12

Algorithm 2: General heuristic bidding strategyInput: depot d of carrier c, warehouse w, set of requests R, vehicle capacity cap, demand di, ∀i ∈ R

set of elementary request combinations R ← {}identify promising request combinations: P ← cluster(d,w,cap, R)foreach S ∈ P doR′ ← generateAllElementaryRequestCombinations(d,w, S )R ← R ∪ R′

endbundle bids B← {}foreach S ∈ E do

p← calculateBidPriceForAnElementaryRequestCombination(d, S )B← B ∪ {b} with b← (p, c, S )

endreturn set B of bundle bids submitted by carrier c

The set of bundle bids generated by Algorithm 2 is always a subset of the set of bundle bids generated by EBBS. Inthe following Sections 4.2, 4.3, and 4.4, respectively, we discuss different implementations of the procedure clusterwhich identifies a promising set P of requests.

4.2. Pairwise synergies based request clustering (PSC)

The core of the strategy PSC is to identify pairs of requests that offer high synergies compared to other pairs ofrequests. The idea is: if two requests i and j (i, j ∈ R) do not provide high synergies, then it is no-good to considersupersets of {i, j}. Vice versa, if i and j already offer high synergies, then all supersets of {i, j} are worth a closer look.Strategy PSC follows Algorithm 2, the cluster-procedure is implemented like Algorithm 3.

Algorithm 3: clusterPairwiseSynergies (PSC)Input: depot d of carrier c, warehouse w, set R of requests, demand dr, ∀r ∈ R, threshold α

pairs of requests: R2 ← {S ∈ 2R : |S | = 2 ∧∑r∈S dr ≤ cap}

foreach S ∈ R2 doσ(S )← (distd,i + distd, j − disti, j)︸ ︷︷ ︸

distance saving

· (cap − di − d j)︸ ︷︷ ︸idle vehicle capacity

with i, j ∈ S , i , j

endreturn {S ∈ R2 |σ(S ) ∈ α-fractile of the synergy distribution}

Algorithm 3 looks at all(|R|

2

)pairs of requests. For each pair S of requests a synergy measure σ(S ) is computed.

Requests offer synergies, if they complement each other nicely, i.e., if they are cost subadditive. We say the synergybetween two requests i and j ({i, j} = S , i , j) is the higher, the higher the saved travel distance due to the combinedfulfillment of i and j is. Saved travel distance is measured by use of the well-known Clarke and Wright (1964)savings heuristic. Furthermore, we say the synergy of pair {i, j} is the higher, the lower the required vehicle capacity(cap − di + d j) is. Given a lower utilized vehicle, ceteris paribus, the chance is higher to include more requests whichcomplement the pair {i,j}. Finally, Algorithm 3 returns those pairs of requests that are among the α-percent pairs withthe highest synergy.

4.3. Capacitated p-median based request clustering (CPMC)

The second way to identify promising request combinations is borrowed from facility location. Facility locationproblems address decisions about the location of facilities in a network and the allocation of demand points to thesefacilities. Finding such an allocation corresponds to determining a set of clusters in a network which is why facility

13

location problems are frequently used as clustering approaches. A well-known problem of this class is the capacitatedp-median problem (PMP). For an overview of model formulations and solution approaches see Reese (2006). ThePMP considers a set of n candidate points, each point has a demand. The goal is to find a subset of p points (p ≤ n)which are denoted as medians. Each point has to be assigned to a median, such that the capacity of the median is notexceeded by the sum of the demands of the assigned points, and the total sum of distances between the points andtheir assigned medians is minimal. The solution of the PMP is a partition of the n points into p clusters. Here, we usethe PMP as an approach to determine elementary request combinations by interpreting the PMP as follows:

min∑

i∈V

∑

i∈Vci jxi j (9)

subject to∑

j∈Vxi j = 1, for all i ∈ V (10)

∑

j∈Vx j j = p, (11)

x j j ≥ xi j, for all i, j ∈ V (12)∑

j∈Vd jxi j ≤ cap, for all i ∈ V (13)

xi j ∈ {0, 1}, for all i, j ∈ V. (14)

The decision variable xi j defined in formula (14) represents whether point i is assigned to the same cluster as j(xi j = 1) or not (xi j = 0). Point i and j stand for locations of customers (i, j ∈ V). In other words, if xi j = 1 then therequests associated to customers i and j are part of the same elementary request combination. The objective function(9) minimizes the total sum of the distances between medians and associated nodes. The costs ci j of assigning node ito node j are represented by the Euclidean distance between customers i and j. Equation (10) ensures that each nodeis assigned to exactly one median. Equation (11) guarantees that there are p medians and that each median is assignedto itself. So, if x j j = 1, then point j is a median. Points may only be assigned to medians (12). The capacity constraint(13) ensures that the total demand assigned to a median does not exceed the vehicle capacity cap.

PMP is NP-hard if the number p of medians is a decision variable; if p is given, PMP is solvable in polynomialtime (Garey and Johnson, 1979). Nevertheless, it is still difficult to solve even when p is given. We preset p to theminimum number of medians which are required to solve PMP:

p =

⌈∑ni=1 di

cap

⌉. (15)

We can interpret a solution of the PMP as a set of elementary request combinations because all requests of a clustermay be performed by a single vehicle. The heuristic bidding strategy that clusters requests by means of solving thePMP is denoted as PMP-based clustering bid strategy (CPMC). Comparing the clustering approaches of the biddingstrategies PSC and CPMC, we expect the following differences with respect to the structure of the computed requestcombinations:

• CPMC generates a significantly smaller number of request clusters than PSC. The number of clusters generatedby PMP is p, see equation (15). PSC, however, generates all clusters which lie in the α-fractile according to thedetermined synergy distribution. For most values of α the latter will be significantly larger, since there are

(|R|2

)

possible request clusters.

• PSC generates clusters which contain two requests. The number of requests in a cluster generated by CPMCvaries. Although a CPMC cluster may contain a single request only, most CPMC clusters will contain morethan two requests which together use the available capacity cap to large extend. This is induced due to thedefinition of p and the solution of PMP.

• The sum of the demand of the requests in a cluster is higher with CPMC compared to PSC. Therefore, CPMCclusters offer a lower potential to generate additional supersets of elementary request clusters.

14

• PSC considers the carrier’s depot d and the shipper’s warehouse while generating promising request clusters.CPMC does not exploit this information.

• CPMC clusters cover the tendered requests smoothly while PSC favors requests which are geographically closerthe to carrier’s depot and the shipper’s warehouse. The PMP-clusters are a partition of the requests, i.e., theCPMC places at least one bundle bid on each request.

As this discussion of the structural differences between the clustering approaches shows, both heuristic biddingstrategies appear sufficiently different. It is hard to anticipate which of the strategies will be more successful in atransport auction setting why we refer to the computational study in Section 5.

4.4. Random request clustering for evaluation purposes (R1 and R2)For evaluation purposes we propose two random bidding strategies denoted as R1 and R2. R1 generates |R|

random bids. A random bundle bid (pb, cb,Rb) is a bid where Rb ⊆ R is generated randomly but the charged pricepb is calculated deterministic as described in Section 3.5.2. Starting with Rb = ∅ a request r ∈ R is selected withprobability 1

|R| . Add r to Rb, if the vehicle capacity cap suffices to perform Rb ∪ {r}, update R := R \ {r} and updateRb := Rb ∪ {r}. Continue extending Rb until the vehicle capacity is exhausted, i.e., dr +

∑s∈Rb

ds > cap. Compute theprice pb for Rb according to Section 3.5.2. All in all, generate |R| random bundle bids.

The random strategy R2 follows Algorithm 2 like PSC and CPMC, however, it simply uses R1 as an implementa-tion for the cluster-procedure.

5. Computational benchmark study

We evaluate the effectiveness and the efficiency of the proposed bidding strategies by a computational benchmarkstudy. For this, we take the bidder’s as well as the auctioneer’s point of view. The test setup and the used performancecriteria are presented in Section 5.1, test instance generation is described in Section 5.2, and the computational resultsare discussed in Section 5.3.

5.1. Evaluation framework and test setupThe evaluation of heuristics usually involves solving a set of test instances and measuring the performance in the

light of the objective function values of the computed solutions. A solution of a bidding strategy is a set of bundlebids. However, an objective function which can measure the quality of the generated bids directly is not available.Therefore, we choose an indirect approach that takes into account the independent decision authorities of bidders andauctioneer as well as the asymmetric distribution of information.

Consider Figure 1 again, the evaluation is aligned according to this auction process. There are n carriers, weconsider the point of view of carrier c. The test instances (see next section) include the set R of tendered requests andthe set Br of bundle bids submitted by the rivals of carrier c. In Figure 1 they are represented by the white boxes,carrier c corresponds to carrier 1 and Br would be defined as B2 ∪ B3 ∪ . . . ∪ Bn. We apply a bid strategy to computethe set Bc of bundle bids of carrier c. After that, the set B := Bc ∪ Br of bundle bids is used as input for the winnerdetermination problem (cf. Section 2.1) which is solved to optimality by the commercial solver CPLEX. The resultingset W of winning bundle bids (W ⊂ B) and the set Wc of winning bundle bids submitted by carrier c (Wc := Bc ∩W)are used to compute performance criteria.

Let Bφc be the set of bids generated by carrier c using bidding strategy φ ∈ Φ, Φ := {EBBS,PSC,CPMC,R1,R2}.Let Wφ represent the set of all winning bids if carrier c uses strategy φ ∈ Φ. Then, the set of carrier’s c winning bidsis Wφ

c := Wφ ∩ Bφc .Four performance criteria κφ1 , κ

φ2 , κ

φ3 , and κφ4 are introduced to evaluate a bidding strategy φ. From the auctioneer

point of view, κφ1 measures how much using φ increases the auctioneer’s total procurement costs f a(W) compared tothe exact strategy EBBS. Lower values of κφ1 are better. κφ1 ≥ 0.0 is expected, because a heuristic bidding strategygenerates a subset of the bids generated by EBBS, which may only increase the auctioneer’s total procurement cost.

κφ1 :=

f a(Wφ) − f a(WEBBS)f a(WEBBS)

. (16)

15

The remaining criteria κφ2 , κφ3 , and κφ4 consider the point of view of the bidding carrier c. κφ2 measures the used sales

potential. The actual sales volume awarded to c when using a heuristic strategy φ ∈ Φ \ {EBBS} is set in relation tothe sales potential. The carrier’s sales potential is given by using the exact bidding strategy EBBS and summing upthe prices of c’s winning bids:

κφ2 := (1 +

f b(Wφ) − f b(WEBBS)f b(WEBBS)

) · 100. (17)

The success rate of a strategy φ, that is the number |Wφc | of carrier c’s winning bids versus the number |Bφc | of carrier

c’s bids, is measured by κφ3 :

κφ3 :=

|Wφc ||Bφc |

· 100. (18)

The ratio of the number of generated bids by strategy φ to the number of generated bids by strategy EBBSis expressedby κφ4 :

κφ4 :=

|Bφc ||BEBBS

c | · 100. (19)

The computational efficiency of a strategy φ is measured by κφ3 and κφ4 , while the effectiveness of φ is measured by κφ1and κφ2 from the auctioneer’s and the bidder’s point of view, respectively.

Finally, some remarks on the implementation of the bidding strategies and the used hardware. The bidding strate-gies are implemented in Java 1.7. The Java code does not benefit from multicore computer architectures as paral-lelization of code segments was not implemented. The prices of a bundle bid on an elementary set of requests arecomputed by the TSP solver Concorde by Applegate et al. (2007). The winner determination problem is solved bythe commercial solver IBM CPLEX Optimizer 12.5. All computations are performed on a standard desktop computerwith Intel Core i7-3770 CPU with 3.4 GHz and 16 GB of working memory using Windows 7 as operating system.

5.2. Generation of benchmark instances

The bidding strategies are compared by means of benchmark instances. An instance consists of a set R of requeststendered by the auctioneer and a set Br of bundle bids of the rivals of carrier c. Furthermore, the vehicle capacity capis given. All in all, 240 instances have been generated in which between 15 and 40 requests are tendered and up to5000 bids of rivaling carriers are present.

The set of requests is implicitly defined by a depot node d and a set V of customer nodes, each with a required loadli (i ∈ V). The warehouse node w and the depot node d are set equal, as explained below. The distance between twonodes is calculated as the Euclidean distance (L2-norm) rounded to the nearest integer value. The customer nodes anddemands are taken from the popular instances2 of the capacitated vehicle routing problem introduced by Christofideset al. (1979). The first node of a CVRP instance is used as the depot node d. The subsequent m = |R| nodes of a CVRPinstance together with their resource demands are considered as the customer nodes of a BGP instance.

Each bundle bid b ∈ Br, b = (pb, c,Rb) of a rivaling carrier is randomly constructed as follows. First, the initialempty set Rb := ∅ of elementary requests is extended:

Step 1. The predefined vehicle capacity cap is randomly reduced to cap′, cap′ ≤ cap, i.e., cap is multiplied by arandom, uniformly distributed number between zero and one.

Step 2. Draw a request r randomly. If r < R and the capacity constraint∑

r∈Rblr ≤ cap′ is satisfied, add r to Rb.

Step 3. Repeat Step 2 until a drawn request violates the capacity constraint for the first time.

2available at http://people.brunel.ac.uk/ mastjjb/jeb/orlib/vrpinfo.html, files vrpnc1.txt to vrpnc5.txt, vrpnc11.txt, and vrpnc12.txt

16

Because the actual vehicle capacity is temporary reduced from cap to cap′ in Step 1, the method generates someelementary request combinations with a low degree of capacity utilization with respect to the vehicle capacity cap. Indoing so, we do not have to make assumptions about the rivaling carriers vehicle fleet, their bidding strategies or theiralready existing requests.

The price pb of a bundle bid b is calculated as the minimum length of the traveling salesman tour that containsthe warehouse w and all customer nodes associated with requests in Rb. The price pb is multiplied by a randomnumber that is chosen randomly from the interval [0.7, 1.3]. By this means, we cope with different depot locations ofdifferent carriers without precisely modeling the rivaling carriers and their depot locations. As the test results show,the generated instances are in many cases very challenging.

These rivaling bundle bids are considered as given. For the test it does not matter whether they reflect the truevaluations of the rivaling carriers or not. Because of the random variations of the vehicle capacities and the bidprices of the rivaling carriers, it is highly likely that a bid generated by one of the bidding strategies at hand has adifferent bid price than a given bid by a rivaling carrier, even if both bids are on the same set of requests. With this,the assumption that all carriers are equal except for different depot locations is adequately integrated into the testinstances. Furthermore, this is the reason why we may simplify the bid price calculation of carrier c by setting thephysical location of the depot of carrier c equal to the physical location of the warehouse w.

5.3. Results and discussion

The detailed results for κ1, κ2, and κ3 as well as run times per test instance as well as averages, median, andquartiles are reported in Table 5 in the electronic appendix.

5.3.1. Elementary bundle bidding strategy (EBBS)For each of the 240 test instances we allowed a maximal computation time of 24 hours. Within the time limit,

the exact bidding strategy EBBS solved 209 out of 240 instances. Among these 209 instances, the bidder did notwin any business at all in 35 cases, i.e., f b = 0. That means, any truthful bidding strategy following the assumptionsof Section 3.1 will not be able to win any of the tendered requests in these 35 instances, because the competition ofthe rivaling bids is too strong. Furthermore, none of the proposed heuristic strategies will generate winning bids forthese instances, because the set of bids generated by the heuristic strategies is always a subset of the bids generatedby EBBS. Therefore, we compare the bidding strategies on those 174 instances.

As expected, the computing time of EBBS reported in Table 5 in the Appendix increases with an increasingnumber of possible elementary bundle bids per instance. The number of potential elementary bundle bids increaseswith an increasing number of tendered requests as well as an increasing vehicle capacity cap. Higher values of capallow more requests to be serviced by a single vehicle. Therefore, the time limit of EBBS was exceeded for some ofthe largest instances with 35 or 40 tendered requests and cap ≥ 60. Nevertheless, even some of the largest instanceswith 40 requests and cap = 80 could be solved within a few minutes, probably because the distribution of the customerloads allowed EBBS to prune early compared to the instances in which EBBS exceeded the time limit.

5.3.2. Comparing effectiveness of the bidding strategiesThe effectiveness of a bidding strategy φ is measured by the used sales potential κφ2 . Table 3 shows the aggregated

results for each of the four heuristic strategies for κφ2 . Two effects become clear. In contrast to both PSC and CPMC,the random bidding strategy R1 is not able to utilize the potential sales volume to a noteworthy degree for the vastmajority of the instances. Therefore, the performance of PSC and CPMC is more than simply bidding randomly.Taking R2 into account, it also becomes visible, that generating random clusters of promising requests which areused as a nucleus to construct elementary bundle bids (cf. Algorithm 2) is less effective. A systematic approach toidentify promising requests clusters as used by PSC and CPMC is meaningful and a crucial element for the design ofsuccessful bidding strategies.

With respect to comparing the effectiveness of PSC and CPMC, PSC seems to outperform CPMC at least on thecoarse categories given in Table 3. To provide more insights, we visualize the individual κφ2 values for each of the 174instances in Figure 5. The chart shows, for each bidding strategy φ, the obtained κφ2 values in descending order. BothPSC and CPMC achieve a sales potential of for more than 100% for several instances (κPSC

2 , κCPMC2 > 100). How is

this possible and is it a good signal?

17

Table 3: Distribution of the used sales potential for 174 tested instances

PSC CPMC R1 R2

# of instances κφ2 ≥ 100% 87 72 3 15# of instances κφ2 ≥ 75% 136 103 3 21# of instances κφ2 ≥ 50% 160 122 5 35# of instances κφ2 < 50% 14 52 169 139# of instances κφ2 = 0% 6 24 148 76

0

50

100

150

200

250

300

0 25 50 75 100 125 150 175

CPMC

PSC

R1R2

Figure 5: Comparing the used sales potential κφ2 in percent for φ = PSC,CPMC,R1,R2

Recall, EBBS is denoted as an exact bidding strategy because it guarantees the same outcome as bidding on eachrequest combination in P(R) and not because it guarantees the maximum value for f b. A heuristic bidding strategyφ that generates a subset of all elementary bundle bids may achieve higher values for f b than the exact biddingstrategy EBBS. EBBS generates the most competitive bids with respect to the request combinations and the chargedprice. Depending on the set of rivaling bids, however, some of the EBBS’s bids could win even if higher prices arecharged. Therefore, the heuristic strategies might generate a higher sales volume compared to bidding on each requestcombination which implies κφ2 > 100%.

With this in mind, however, a bidding strategy that generates values of κφ2 ≥ 100 percent is double-edged. Onthe one hand, we prefer strategies that generate higher sales volumes f b and therefore higher values of κφ2 are alwaysbetter. On the other hand, there is no free lunch. Shifting from the static view on the issue to a dynamic view, theportion of sales which exceeds the sales potential, i.e., κφ2 − 100 > 0, signals a competitive gap. This gap is not earnedsystematically but based on fortunate circumstances with respect to the actual composition of the rivaling bundle bidswhich cannot be influenced by the bidder. It also indicates the auctioneer overpays.

Looking at Figure 5 the strategy PSC is more robust than CPMC. PSC uses the sales volume more frequently, i.e.,it achieves values of κφ2 near hundred percent with a higher frequency. Both methods solve a roughly equal numberof instances with a potential of hundred or more percent. Nevertheless, the CPMC-graph declines faster than thePSC-graph. All in all, for about half of the tested instances, both PSC and CPMC achieve exactly the same or an evensuperior outcome—with the discussed limitations—as EBBS.

5.3.3. Comparing efficiency of the bidding strategiesThe strategies EBBS, PSC, CPMC significantly vary with respect to the computational times. Over all 209 solved

instances (including the 35 instances with f b = 0), the median runtime for EBBS, PSC, CPMC is 152 seconds, 9seconds, and 3 seconds, respectively. However, the averaged runtime for EBBS, PSC, CPMC is 153 minutes, 20minutes and 0.75 minutes, respectively. Strategy CPMC is an order of magnitude faster than PSC, and PSC is anorder of magnitude faster than EBBS. This effect may be attributed almost exclusively to differences in the number of

18

0

10

20

30

40

50

60

70

0 20 40 60 80 100 120

#bid

sPS

C/#

bids

EB

BS,

κ4

in%

used sales potential, κ2 in %

PSC

0

10

20

30

40

50

60

70

0 20 40 60 80 100 120

#bid

sC

PMC

/#bi

dsE

BB

S,κ4

in%

used sales potential, κ2 in %

CPMC

Figure 6: Efficiency of PSC and CPMC.

generated bundle bids. The average success rates are κEBBS3 = 0.035 percent, κPSC

3 = 0.081 percent, and κCPMC3 = 5.72

percent. Therefore, the heuristic strategies are a serious alternative to EBBS if the trade-off between used sales volumeand runtime are considered.

To gain more insights on the efficiency and effectiveness of the bidding strategies look at Figure 6. On bothdiagrams, the abscissa shows the effectiveness measured in the used sales potential κφ2 . Higher values are better. Note,a few data points with κφ2 ≥ 120 are not shown. The ordinate shows the efficiency measure κφ4 . Lower values are better.Efficiency of a heuristic strategy is measured in terms of the number of computed bundle bids relative to the numberof bundle bids computed by the exact strategy EBBS. Each circle in the diagrams shows the results for one of the174 test instances solved by the strategy PSC (left diagram) and by the strategy CPMC (right diagram). Circles inlower-right area of a diagram indicate better results.

Strategy CPMC is more efficient than PSC. Although both strategies solve many instances with κ2 ∼ 100%, CPMCrequires less bundle bids to achieve these results. Actually, for most of the instances the number of computed bundlebids is significantly less than five percent compared to EBBS. In contrast, PSC requires for all instances significantlymore than five percent of the bundle bids of EBBS. Averaged over all 174 instances, the strategy PSC requires only36.1 percent of the bundle bids of EBBS in order to achieve 90.8 percent of the sales volume won by strategy EBBS.The bids of strategy CPMC are even more efficient. On average, the strategy CPMC requires only 4.3 percent of thebundle bids of EBBS in order to achieve 80.9 percent of the sales volume won by strategy EBBS.

5.3.4. Effects on auctioneer’s total procurement costsThe auctioneer’s total procurement costs are also affected by the choice of a bidding strategy. Table 4 shows

that the maximum increase of the procurement costs is about 11 percent if the carrier uses strategy PSC. The otherstrategies CPMC, R1, and R2 provide an inferior worst-case performance, because the cost increase for each strategyis about 25 percent. For the auctioneer, strategy PSC is more competitive than strategy CPMC which follows fromthe mean and median values of Table 4. As the table indicates, the auctioneer will probably be able to significantlyreduce his procurement costs, if he provides advanced data and decision support tools in order to support the biddingprocess of the carriers.

6. Conclusion and outlook

In order to support a freight carrier in a combinatorial transport auction we introduced and evaluated three strate-gies for truthful bundle bidding. The bidding strategies help a bidder (i.e., a carrier) to decide on which subsets of theset of tendered requests to bid. The bidder’s goal is to maximize the sum of the prices of those bundle bids accepted by

19

κPSC1 κCPMC

1 κR11 κR2

1 κPSC1 κCPMC

1 κR11 κR2

1

min 0.0 0.0 0.0 0.0 25%-quantile 0.0 1.0 2.9 2.2max 11.1 23.1 28.6 22.4 median 0.8 2.8 6.3 5.0mean 1.4 3.5 7.8 6.7 75%-quantile 2.4 4.9 11.9 9.7

Table 4: Increase κφ1 of the auctioneer’s total procurement costs f a over 174 test instances.

the auctioneer (i.e., the shipper). All three strategies do not require information about the bundle bids of the rivalingcarriers which is advantageous because such information is often very difficult to estimate.

The bidding strategy EBBS is called exact, because it guarantees the same results as bidding on each requestcombination in the powerset of the set of tendered requests. In contrast to such a naïve brute-force strategy, EBBScomputes a significantly smaller number of bundle bids by using the notion of elementary bundle bids. The exactbidding strategy EBBS was motivated by showing that it is—given the transport and auction scenarios at hand—necessary and sufficient to compute all elementary bundle bids in order to achieve the same outcome as a brute-forcestrategy. The bidding strategies denoted as PSC and CPMC are called heuristic. The idea of these heuristic strategiesis to generate only those elementary bundle bids which are supersets of promising request combinations. Bid strategyPSC identifies promising request combinations by measuring pairwise synergies through modified Clarke-and-Wrightsavings values. In contrast, bid strategy CPMC identifies promising request combinations by solving the capacitatedp-median problem. The computational evaluation showed that EBBS may be used in combinatorial transport auctionswith up to 40 tendered requests. This is about twice the size of most previous truthful bidding strategies. On average,bid strategy PSC requires 36 percent of the bundle bids generated by EBBS in order to achieve 91 percent of thepossible revenue. Bid strategy CPMC is more efficient, it computes only 4.3 percent of the bundle bids generated byEBBS, but achieves 80.9 percent of the possible revenue. However, the results of PSC are more robust compared toCPMC.

By showing that heuristic bundle bidding strategies are viable and proposing ways and means to evaluate theperformance of (heuristic) strategies for bundle bidding a basis for future research has been created. If we learnmore about promising request combinations and heuristic bidding strategies, carriers as well as shippers may benefit.Carriers increase their chance to win additional business. The entrance barriers to successfully bid in transport auctionsare reduced, because the computational effort to generate bids decreases with more powerful strategies. On the onehand, future research should study features of winning bundle bids and non-winning bundle bids in order to gainmore insights into characteristics of promising request combinations. The bidding strategies should be adapted toother transport auction scenarios with, e.g., existing transport commitments, general pickup-and-delivery requests, orrequests with time windows. On the other hand, shippers also benefit from effective bidding strategies. Therefore,shippers should provide tools that comply with the carrier’s private information requirements in order to support thecarriers to generate competitive bundle bids.

Acknowledgements

The cooperative junior research group on Computational Logistics is funded by the University of Bremen in linewith the Excellence Initiative of German federal and state governments.

References

Abrache, J., Crainic, T., Gendreau, M., Rekik, M., 2007. Combinatorial auctions. Annals of Operations Research 153, 131–164.An, N., Elmaghraby, W., Keskinocak, P., 2005. Bidding strategies and their impact on revenues in combinatorial auctions. Journal of Revenue and

Pricing Management 3, 337–357.Applegate, D.L., Bixby, R.E., Chvátal, V., Cook, W.J., 2007. The Traveling Salesman Problem: A Computational Study. Princeton University

Press.Beil, D.R., Cohn, A., Sinha, A., 2007. Simplified Bidding and Solution Methodology for Truckload Procurement and Other Vcg Combinatorial

Auctions. Ross School of Business Working Paper Series Working Paper No. 1034. Stephen M. Ross School of Business, University ofMichigan.

20

Berger, S., Bierwirth, C., 2010. Solutions to the request reassignment problem in collaborative carrier networks. Transportation Research Part E:Logistics and Transportation Review 46, 627–638.

Bichler, M., Shabalin, P., Pikovsky, A., 2009. A computational analysis of linear price iterative combinatorial auction formats. Information SystemsResearch 20, 33–59.

Buer, T., Kopfer, H., 2014. A Pareto-metaheuristic for a bi-objective winner determination problem in a combinatorial reverse auction. Computers& Operations Research 41, 208–220.

Buer, T., Pankratz, G., 2010a. Grasp with hybrid path relinking for bi-objective winner determination in combinatorial transportation auctions.Business Research 3, 192–213.

Buer, T., Pankratz, G., 2010b. Solving a bi-objective winner determination problem in a transportation procurement auction. Logistics Research 2,65–78.

Caplice, C., Sheffi, Y., 2003. Optimization-based procurement for transportation services. Journal of Business Logistics 24, 109–128.Caplice, C., Sheffi, Y., 2006. Combinatorial auctions for truckload transportation, in: Cramton, P., Shoaham, Y., Steinberg, R. (Eds.), Combinatorial

Auctions. MIT Press, Cambridge, MA, pp. 539–571.Chang, T.S., 2009. Decision support for truckload carriers in one-shot combinatorial auctions. Transportation Research Part B: Methodological

43, 522 – 541.Chen, R.L.Y., AhmadBeygi, S., Cohn, A., Beil, D.R., Sinha, A., 2009. Solving truckload procurement auctions over an exponential number of

bundles. Transportation Science 43, 493–510.Christofides, N., Mingozzi, A., Toth, P., 1979. The vehicle routing problem, in: Christofides, A., Mingozzi, Toth, P., Sandi, C. (Eds.), Combinatorial

Optimization. Wiley, Chichester, UK, pp. 315–338.Clarke, G., Wright, J., 1964. Scheduling of vehicles from a central depot to a number of delivery points. Operations Research 12, 568–581.Day, R.W., Raghavan, S., 2009. Matrix bidding in combinatorial auctions. Operations Research 57, 916–933.van Duin, J., Tavasszy, L., Taniguchi, E., 2007. Real time simulation of auctioning and re-scheduling processes in hybrid freight markets. Trans-

portation Research Part B: Methodological 41, 1050 – 1066. <ce:title>Behavioural insights into the Modelling of Freight Transportation andDistribution Systems</ce:title>.

Elmaghraby, W., Keskinocak, P., 2004. Combinatorial auctions in procurement, in: Harrison, T., Lee, H., Neale, J. (Eds.), The Practice of SupplyChain Management: Where Theory and Application Converge. Springer-Verlag, New York, pp. 245–258.

Figliozzi, M.A., Mahmassani, H.S., Jaillet, P., 2007. Pricing in dynamic vehicle routing problems. Transportation Science 41, 302–318. http://transci.journal.informs.org/cgi/reprint/41/3/302.pdf.

Garey, M.R., Johnson, D.S., 1979. Computers and Intractability - A Guide to the Theory of NP-Completeness. W. H. Freeman and Company.Garrido, R.A., 2007. Procurement of transportation services in spot markets under a double-auction scheme with elastic demand. Transportation

Research Part B: Methodological 41, 1067 – 1078. Behavioural insights into the Modelling of Freight Transportation and Distribution Systems.Huang, G.Q., Xu, S.X., 2013. Truthful multi-unit transportation procurement auctions for logistics e-marketplaces. Transportation Research Part

B: Methodological 47, 127 – 148.Kopfer, H., Pankratz, G., 1999. Das Groupage-Problem kooperierender Verkehrsträger, in: Kall, P., Lüthi, H.J. (Eds.), Operations Research

Proceedings 1998, Springer, Berlin. pp. 453–462.Krajewska, M., Kopfer, H., 2006. Collaborating freight forwarding enterprises. OR Spectrum 28, 301–317.Krajewska, M.A., Kopfer, H., Laporte, G., Ropke, S., Zaccour, G., 2008. Horizontal cooperation among freight carriers: request allocation and

profit sharing. Journal of the Operational Research Society 59, 1483–1491.Kwon, R.H., Anandalingam, G., Ungar, L.H., 2005. Iterative combinatorial auctions with bidder-determined combinations. Management Science

51, 407–418.Lee, C.G., Kwon, R.H., Ma, Z., 2007. A carrier’s optimal bid generation problem in combinatorial auctions for transportation procurement.

Transportation Research Part E: Logistics and Transportation Review 43, 173 – 191.Nisan, N., 2000. Bidding and allocation in combinatorial auctions, in: Proceedings of the 2nd ACM Conference on Electronic Commerce, pp.

1–12.Nisan, N., Ronen, A., 2001. Algorithmic mechanism design. Games and Economic Behavior 35, 166 – 196.Plummer, C.L., 2003. Bidder response to combinatorial auctions in truckload procurement. Master’s thesis. Massachusetts Institute of Technology.

Engineering Systems Division.Reese, J., 2006. Solution methods for the p-median problem: An annotated bibliography. Networks 48, 125–142.Remli, N., Rekik, M., 2013. A robust winner determination problem for combinatorial transportation auctions under uncertain shipment volumes.

Transportation Research Part C: Emerging Technologies 35, 204–217.Schönberger, J., Kopfer, H., 2004. A collaborative approach for solving the portfolio re-composition problem in transport cooperations, in:

Spengler, T., Voß, S., Kopfer, H. (Eds.), Logistik Management. Physica-Verlag HD, pp. 37–49.Sheffi, Y., 2004. Combinatorial auctions in the procurement of transportation services. Interfaces 34, 245–252.Song, J., Regan, A., 2005. Approximation algorithms for the bid construction problem in combinatorial auctions for the procurement of freight

transportation contracts. Transportation Research Part B: Methodological 39, 914 – 933.Triki, C., Oprea, S., Beraldi, P., Crainic, T.G., 2014. The stochastic bid generation problem in combinatorial transportation auctions. European

Journal of Operational Research 236, 991–999.de Vries, S., Vohra, R.V., 2003. Combinatorial auctions: A survey. INFORMS Journal on Computing 15, 284–309.Wang, X., Xia, M., 2005. Combinatorial bid generation problem for transportation service procurement. Transportation Research Record: Journal

of the Transportation Research Board 1923, 189–198.Xu, S.X., Huang, G.Q., 2013. Transportation service procurement in periodic sealed double auctions with stochastic demand and supply. Trans-

portation Research Part B: Methodological 56, 136–160.

21

Supplementary data / electronic appendix

22

Tabl

e5:

Det

aile

dco

mpa

riso

nof

EB

BS,

PSC

,CPM

C,R

1,an

dR

2

Inst

ance

Auc

tione

erB

iddi

ngca

rrie

r,us

edpo

tent

ial

Bid

ding

carr

ier,

succ

ess

rate

(%)

Tim

e(s

)

Id|R|

cap

fa (WE

BB

S)

κPSC

1κC

PMC

1κR

11

κR2

1fb (W

EB

BS)

κPSC

2κC

PMC

2κR

12

κR2

2κE

BB

S3

κPSC

3κC

PMC

3κR

13

κR2

3E

BB

SPS

CC

PMC

115

5029

5.96

––

––

0–

––

––

––

––

00

02

1550

335.

17–

––

–0

––

––

––

––

–0

00

315

5029

7.74

3.0

0.0

3.0

3.0

680.

010

0.0

0.0

0.0

0.1

0.0

0.2

0.0

0.0

00

04

1550

355.

982.

54.

94.

94.

920

998

.10.

00.

00.

00.

10.

20.

00.

00.

00

00

515

5025

7.60

0.0

0.0

1.2

0.0

7910

0.0

100.

00.

010

0.0

0.1

0.3

1.2

0.0

2.1

00

06

1560

285.

02–

––

–0

––

––

––

––

–0

00

715

6029

0.62

0.0

0.7

0.7

0.7

7210

0.0

0.0

0.0

0.0

0.3

0.5

0.0

0.0

0.0

00

08

1560

291.

640.

02.

45.

35.

379

100.

023

6.7

0.0

0.0

0.0

0.2

3.2

0.0

0.0

00

09

1560

315.

82–

––

–0

––

––

––

––

–0

00

1015

6023

9.47

2.5

2.5

6.3

3.0

5025

0.0

176.

018

0.0

92.0

0.1

0.4

2.2

13.3

0.8

00

011

1570

273.

01–

––

–0

––

––

––

––

–0

00

1215

7029

1.33

0.0

0.0

1.0

1.0

145

100.

010

0.0

0.0

0.0

0.3

0.7

0.9

0.0

0.0

00

013

1570

264.

56–

––

–0

––

––

––

––

–1

00

1415

7027

1.08

0.8

12.1

12.7

6.8

114

93.9

187.

70.

099

.10.

00.

022

.20.

02.

41

10

1515

7021

4.04

0.0

3.4

7.0

6.6

118

100.

013

0.5

0.0

77.1

0.1

0.2

6.9

0.0

1.9

00

016

1580

250.

420.

00.

01.

11.

198

100.

010

0.0

0.0

0.0

0.0

0.1

0.4

0.0

0.0

00

017

1580

286.

390.

02.

92.

92.

955

278.

20.

00.

00.

00.

10.

50.

00.

00.

00

00

1815

8027

5.40

0.0

0.0

3.2

3.2

9810

0.0

100.

00.

00.

00.

00.

11.

60.

00.

03

00

1915

8026

1.44

0.0

0.8

1.9

1.9

9510

0.0

104.

20.

00.

00.

00.

00.

50.

00.

04

30

2015

8018

4.50

0.5

1.4

5.2

2.0

4112

4.4

107.

311

9.5

112.

20.

00.

10.

26.

70.

51

00

2120

5039

2.36

2.5

0.3

3.3

0.0

136

120.

648

.50.

010

0.0

0.1

0.2

0.9

0.0

0.5

00

022

2050

453.

24–

––

–0

––

––

––

––

–0

00

2320

5038

5.32

1.8

0.5

6.5

6.2

311

86.2

100.

60.

022

.80.

10.

10.

20.

00.

21

00

2420

5059

0.37

0.8

1.1

2.9

0.0

9616

0.4

109.

40.

010

0.0

0.0

0.0

1.4

0.0

0.7

10

125

2050

424.

781.

92.

72.

72.

720

989

.00.

00.

00.

00.

20.

40.

00.

00.

00

00

2620

6039

0.34

0.3

3.6

3.6

3.4

226

89.8

0.0

0.0

35.4

0.1

0.1

0.0

0.0

0.1

00

027

2060

419.

080.

00.

00.

00.

414

810

0.0

100.

044

.60.

00.

20.

41.

65.

00.

00

00

2820

6037

1.19

3.8

5.7

10.9

10.9

323

101.

210

6.5

0.0

0.0

0.0

0.1

6.0

0.0

0.0

100

029

2060

510.

380.

00.

07.

77.

726

910

0.0

100.

00.

00.

00.

00.

04.

40.

00.

06

50

3020

6035

4.12

––

––

0–

––

––

––

––

10

131

2070

361.

281.

10.

51.

71.

715

743

.952

.90.

00.

00.

00.

00.

80.

00.

03

11

3220

7038

3.11

0.0

2.8

2.8

2.8

119

100.

00.

00.

00.

00.

10.

20.

00.

00.

00

00

3320

7035

3.62

1.5

2.1

7.0

7.0

255

78.0

141.

60.

00.

00.

00.

036

.40.

00.

015

05

234

2070

440.

500.

74.

911

.511

.527

598

.560

.70.

00.

00.

00.

00.

60.

00.

012

133

135

2070

317.

000.

01.

86.

84.

614

210

0.0

58.5

0.0

35.2

0.0

0.1

0.4

0.0

1.0

50

036

2080

350.

960.

80.

03.

03.

021

170

.610

0.0

0.0

0.0

0.0

0.0

1.1

0.0

0.0

243

137

2080

343.

742.

82.

92.

83.

113

352

.648

.152

.60.

00.

00.

13.

15.

00.

01

00

23

Tabl

e5:

Det

aile

dco

mpa

riso

nof

EB

BS,

PSC

,CPM

C,R

1,an

dR

2

Inst

ance

Auc

tione

erB

iddi

ngca

rrie

r,us

edpo

tent

ial

Bid

ding

carr

ier,

succ

ess

rate

(%)

Tim

e(s

)

Id|R|

cap

fa (WE

BB

S)

κPSC

1κC

PMC

1κR

11

κR2

1fb (W

EB

BS)

κPSC

2κC

PMC

2κR

12

κR2

2κE

BB

S3

κPSC

3κC

PMC

3κR

13

κR2

3E

BB

SPS

CC

PMC

3820

8032

9.25

––

––

0–

––

––

––

––

712

302

3920

8042

1.99

––

––

0–

––

––

––

––

655

711

4020

8029

5.96

0.7

3.4

8.1

8.1

140

97.9

125.

00.

00.

00.

00.

11.

10.

00.

048

21

4125

5051

2.32

0.0

5.3

8.2

6.7

294

100.

049

.015

.021

.10.

10.

217

.64.

00.

12

01

4225

5061

7.06

0.0

0.8

1.4

1.4

255

100.

023

.50.

00.

00.

30.

710

.00.

00.

00

02

4325

5050

6.16

1.6

2.2

9.5

9.9

450

68.9

87.3

13.6

0.0

0.0

0.1

19.2

4.0

0.0

789

044

2550

763.

190.

03.

95.

16.

916

316

0.1

291.

410

2.5

100.

00.

00.

03.

24.

00.

332

14

4525

5048

1.48

0.1

1.3

3.5

0.8

274

96.4

78.1

0.0

96.4

0.1

0.1

3.1

0.0

1.9

10

346

2560

472.

990.

32.

28.

68.

232

113

2.7

121.

20.

048

.00.

00.

11.

80.

00.

130

20

4725

6055

0.27

0.7

3.4

6.7

4.0

310

121.

612

2.6

0.0

24.8

0.1

0.4

10.9

0.0

0.3

00

348

2560

472.

722.

04.

18.

88.

834

977

.114

1.0

0.0

0.0

0.0

0.0

9.2

0.0

0.0

760

52

4925

6066

2.70

0.0

4.4

8.9

8.3

270

100.

077

.80.

041

.50.

00.

03.

00.

00.

447

22

150

2560

425.

413.

43.

510

.65.

819

284

.979

.215

.689

.60.

00.

010

.04.

00.

512

22

5125

7044

0.18

1.6

5.2

12.4

9.4

382

85.3

96.3

0.0

19.1

0.0

0.0

2.0

0.0

0.1

224

253

5225

7051

4.24

0.0

2.1

5.1

3.7

299

100.

087

.60.

046

.20.

00.

23.

20.

00.

75

01

5325

7044

8.90

2.9

9.3

19.3

19.3

374

103.

581

.60.

052

.40.

00.

00.

00.

00.

44,

954

7052

354

2570

579.

544.

94.

915

.315

.327

759

.659

.60.

00.

00.

00.

01.

20.

00.

03,

595

52

5525

7037

1.72

4.1

2.4

14.8

10.4

353

36.3

9.6

0.0

24.1

0.0

0.0

1.1

0.0

0.5

4320

256

2580

402.

824.

29.

717

.115

.833

810

5.0

130.

80.

021

.60.

00.

05.

40.

00.

11,

224

231

157

2580

477.

053.

01.

69.

78.

640

187

.510

1.5

0.0

39.7

0.0

0.1

5.1

0.0

0.8

261

158

2580

419.

364.

55.

421

.417

.939

286

.771

.40.

041

.60.

00.

00.

30.

00.

127

,705

537

2159

2580

516.

329.

04.

217

.117

.139

20.

013

7.2

0.0

0.0

0.0

0.0

100.

00.

00.

023

,597

922

6025

8032

5.34

4.5

5.7

15.1

14.9

214

44.9

136.

40.

043

.00.

00.

00.

10.

00.

278

916

04

6130

5059

4.02

1.2

9.0

13.3

7.5

419

114.

613

1.3

0.0

54.2

0.0

0.1

53.3

0.0

0.1

424

662

3050

686.

540.

01.

81.

81.

315

810

0.0

0.0

0.0

30.4

0.2

0.3

0.0

0.0

0.5

10

363

3050

577.

940.

72.

515

.711

.543

496

.110

8.8

0.0

37.6

0.0

0.0

6.5

0.0

0.2

332

321

6430

5095

1.51

4.3

7.8

12.0

9.8

542

60.5

160.

730

.170

.70.

00.

03.

43.

30.

773

516

465

3050

547.

290.

23.

511

.16.

735

787

.111

2.6

26.9

47.3

0.0

0.1

14.3

3.3

0.7

232

566

3060

543.

580.

60.

79.

86.

549

666

.510

0.8

0.0

13.3

0.0

0.0

1.3

0.0

0.0

558

825

6730

6063

0.46

1.8

6.4

7.9

6.5

415

104.

132

.30.

051

.60.

10.

320

.00.

01.

01

03

6830

6052

6.40

2.1

2.7

22.3

21.3

504

74.8

97.8

0.0

19.2

0.0

0.0

0.0

0.0

0.1

3,07

692

884

6930

6078

2.77

11.1

11.2

16.4

14.3

488

34.4

78.3

0.0

35.2

0.0

0.0

1.8

0.0

0.3

6,56

533

53

7030

6048

2.31

1.5

6.4

14.3

11.2

244

99.6

198.

014

.810

8.2

0.0

0.0

21.6

3.3

0.2

386

643

7130

7049

6.60

2.4

11.9

17.8

16.9

448

87.5

95.8

0.0

25.2

0.0

0.0

0.4

0.0

0.0

3,21

421

04

7230

7056

9.62

0.6

2.0

11.2

11.4

418

107.

411

2.0

20.8

35.4

0.0

0.1

43.8

3.3

0.4

291

373

3070

492.

402.

69.

318

.118

.138

988

.913

8.3

0.0

0.0

0.0

0.0

60.0

0.0

0.0

23,3

1172

23

7430

7071

5.25

5.9

5.0

13.0

11.4

289

67.1

191.

30.

067

.80.

00.

00.

60.

00.

165

,971

7,82

311

24

Tabl

e5:

Det

aile

dco

mpa

riso

nof

EB

BS,

PSC

,CPM

C,R

1,an

dR

2

Inst

ance

Auc

tione

erB

iddi

ngca

rrie

r,us

edpo

tent

ial

Bid

ding

carr

ier,

succ

ess

rate

(%)

Tim

e(s

)

Id|R|

cap

fa (WE

BB

S)

κPSC

1κC

PMC

1κR

11

κR2

1fb (W

EB

BS)

κPSC

2κC

PMC

2κR

12

κR2

2κE

BB

S3

κPSC

3κC

PMC

3κR

13

κR2

3E

BB

SPS

CC

PMC

7530

7041

4.17

1.7

4.4

19.1

14.0

257

106.

612

6.8

0.0

105.

40.

00.

02.

50.

00.

22,

267

1,04

43

7630

8045

9.84

4.0

8.1

20.3

20.3

448

101.

610

4.5

0.0

0.0

0.0

0.0

0.0

0.0

0.0

21,7

741,

220

329

7730

8052

6.80

1.7

6.3

16.8

18.8

450

87.3

104.

718

.20.

00.

00.

00.

13.

30.

029

511

980

3080

380.

935.

24.

618

.916

.428

710

9.8

119.

530

.748

.10.

00.

00.

83.

30.

312

,712

4,36

01

8135

5071

9.04

0.0

5.0

13.6

8.9

584

100.

585

.10.

050

.30.

00.

10.

10.

00.

155

53

8235

5082

1.34

0.0

1.1

6.8

4.3

430

100.

012

1.6

14.0

47.0

0.2

0.5

40.0

2.9

0.6

10

385

3550

719.

870.

63.

011

.39.

051

188

.393

.017

.443

.10.

00.

14.

52.

90.

263

125

8735

6071

9.88

0.0

2.2

7.5

7.0

462

100.

073

.40.

014

.90.

10.

20.

50.

00.

15

03

9135

7058

2.76

3.5

6.3

21.1

18.8

527

109.

786

.50.

013

.70.

00.

01.

70.

00.

114

,213

683

392

3570

663.

641.

611

.011

.911

.662

680

.413

.112

.010

.20.

00.

110

.02.

90.

378

33

9335

7059

3.25

2.5

5.8

17.4

17.2

571

78.3

89.3

0.0

15.9

0.0

0.0

0.4

0.0

0.1

88,8

092,

838

3095

3570

549.

642.

12.

414

.412

.543

011

0.0

99.5

0.0

19.3

0.0

0.0

2.3

0.0

0.1

6,20

62,

342

396

3580

559.

281.

67.

123

.216

.250

699

.411

8.4

0.0

25.9

0.0

0.0

3.0

0.0

0.0

72,8

655,

623

597

3580

612.

581.

18.

215

.313

.541

092

.091

.50.

025

.10.

00.

038

.50.

00.

21,

140

433

100

3580

497.

492.

24.

916

.116

.128

313

2.9

134.

60.

00.

00.

00.

027

.30.

00.

035

,111

13,4

403

101

4050

822.

580.

44.

912

.17.

169

583

.282

.60.

033

.70.

00.

029

.20.

00.

076

110

63

102

4050

929.

920.

05.

610

.12.

874

310

0.0

78.9

6.7

46.2

0.3

0.6

3.2

2.5

0.4

10

410

340

5083

3.67

0.0

2.3

18.0

16.6

595

102.

213

6.0

25.5

18.5

0.0

0.0

5.9

5.0

0.0

5,30

71,

141

310

440

501,

282.

775.

72.

417

.113

.483

959

.215

3.6

13.3

49.2

0.0

0.0

0.0

2.5

0.1

64,4

9032

,635

2,05

010

540

5086

1.04

0.8

5.6

12.8

9.2

434

97.9

149.

118

.468

.20.

00.

021

.62.

50.

117

454

510

640

6074

0.78

1.6

10.4

21.9

17.2

644

98.0

81.4

0.0

28.1

0.0

0.0

5.6

0.0

0.1

11,2

241,

250

310

740

6083

7.59

1.0

4.0

13.1

11.5

686

87.3

97.4

24.9

46.5

0.1

0.2

0.7

5.0

0.6

281

310

840

6072

8.93

2.7

4.7

23.1

18.2

476

98.7

133.

422

.949

.80.

00.

03.

32.

50.

066

,825

4,62

26

110

4060

762.

910.

85.

418

.113

.356

682

.211

1.5

0.0

71.0

0.0

0.0

3.3

0.0

0.1

3,01

287

03

111

4070

667.

063.

523

.128

.622

.459

411

3.6

46.3

0.0

37.5

0.0

0.0

33.3

0.0

0.1

116,

071

11,1

613

112

4070

739.

052.

66.

217

.617

.558

210

2.7

101.

90.

013

.60.

00.

11.

50.

00.

251

031

111

540

7066

8.00

2.1

4.4

21.2

20.5

647

96.6

92.1

0.0

27.2

0.0

0.0

3.9

0.0

0.1

24,3

657,

120

411

740

8069

2.38

1.7

18.5

19.1

14.3

532

101.

39.

40.

062

.60.

00.

010

.00.

00.

54,

228

329

3

121

1550

274.

14–

––

–0

––

––

––

––

–0

00

122

1550

309.

930.

00.

10.

10.

067

100.

00.

00.

010

0.0

0.5

0.8

0.0

0.0

1.5

00

112

315

5028

3.73

––

––

0–

––

––

––

––

00

012

415

5036

0.05

––

––

0–

––

––

––

––

00

012

515

5025

6.74

0.0

0.2

2.7

0.8

3910

0.0

94.9

0.0

112.

80.

10.

31.

20.

01.

60

02

126

1560

261.

21–

––

–0

––

––

––

––

–0

02

127

1560

277.

12–

––

–0

––

––

––

––

–0

01

128

1560

245.

90–

––

–0

––

––

––

––

–0

00

129

1560

303.

87–

––

–0

––

––

––

––

–0

02

25

Tabl

e5:

Det

aile

dco

mpa

riso

nof

EB

BS,

PSC

,CPM

C,R

1,an

dR

2

Inst

ance

Auc

tione

erB

iddi

ngca

rrie

r,us

edpo

tent

ial

Bid

ding

carr

ier,

succ

ess

rate

(%)

Tim

e(s

)

Id|R|

cap

fa (WE

BB

S)

κPSC

1κC

PMC

1κR

11

κR2

1fb (W

EB

BS)

κPSC

2κC

PMC

2κR

12

κR2

2κE

BB

S3

κPSC

3κC

PMC

3κR

13

κR2

3E

BB

SPS

CC

PMC

130

1560

221.

050.

51.

64.

90.

045

100.

010

0.0

0.0

100.

00.

10.

21.

10.

00.

70

00

131

1570

239.

67–

––

–0

––

––

––

––

–0

02

132

1570

264.

26–

––

–0

––

––

––

––

–0

01

133

1570

263.

42–

––

–0

––

––

––

––

–1

01

134

1570

262.

36–

––

–0

––

––

––

––

–2

10

135

1570

200.

472.

73.

53.

52.

045

113.

30.

00.

010

8.9

0.0

0.1

0.0

0.0

1.2

00

013

615

8024

1.28

––

––

0–

––

––

––

––

00

113

715

8024

1.20

––

––

0–

––

––

––

––

00

113

815

8025

1.18

––

––

0–

––

––

––

––

50

113

915

8023

2.55

––

––

0–

––

––

––

––

84

014

015

8017

8.57

––

––

0–

––

––

––

––

10

1

141

2050

373.

970.

00.

10.

10.

082

100.

00.

00.

010

0.0

0.1

0.1

0.0

0.0

0.2

00

114

220

5040

8.76

––

––

0–

––

––

––

––

00

114

320

5036

9.26

0.3

0.0

0.3

0.0

610.

010

0.0

0.0

100.

00.

00.

00.

10.

00.

31

00

144

2050

561.

340.

00.

30.

30.

393

100.

00.

00.

00.

00.

00.

00.

00.

00.

01

03

145

2050

407.

61–

––

–0

––

––

––

––

–0

01

146

2060

348.

35–

––

–0

––

––

––

––

–0

01

147

2060

390.

26–

––

–0

––

––

––

––

–0

01

148

2060

357.

601.

50.

01.

91.

999

75.8

100.

00.

00.

00.

00.

01.

50.

00.

026

10

149

2060

477.

120.

00.

01.

01.

015

710

0.0

100.

00.

00.

00.

00.

02.

20.

00.

019

72

150

2060

335.

960.

40.

40.

40.

445

0.0

0.0

0.0

0.0

0.0

0.0

0.0

0.0

0.0

10

215

120

7033

3.91

––

––

0–

––

––

––

––

41

315

220

7034

1.04

––

––

0–

––

––

––

––

00

015

320

7032

8.78

3.0

3.0

3.0

3.0

107

0.0

0.0

0.0

0.0

0.0

0.0

0.0

0.0

0.0

154

41

154

2070

406.

86–

––

–0

––

––

––

––

–13

632

115

520

7030

5.93

0.0

0.5

2.2

2.2

8510

0.0

52.9

0.0

0.0

0.0

0.1

0.4

0.0

0.0

70

115

620

8030

8.28

––

––

0–

––

––

––

––

315

215

720

8034

5.90

0.0

0.0

1.2

1.2

9610

0.0

100.

00.

00.

00.

00.

13.

10.

00.

01

00

158

2080

321.

801.

51.

41.

51.

593

89.2

117.

20.

00.

00.

00.

00.

30.

00.

068

934

215

920

8039

9.66

0.0

0.0

2.0

2.0

164

100.

010

0.0

0.0

0.0

0.0

0.0

0.4

0.0

0.0

619

124

116

020

8028

1.35

2.7

1.9

4.4

3.4

130

67.7

102.

368

.578

.50.

00.

01.

15.

00.

553

10

161

2550

479.

78–

––

–0

––

––

––

––

–1

00

162

2550

553.

120.

00.

50.

50.

077

100.

00.

00.

010

0.0

0.1

0.2

0.0

0.0

0.3

00

216

325

5048

6.77

0.0

0.0

5.4

5.4

154

100.

010

0.0

0.0

0.0

0.0

0.0

7.7

0.0

0.0

807

016

425

5074

1.27

0.0

0.7

1.9

1.9

202

130.

749

.50.

00.

00.

00.

01.

10.

00.

038

13

165

2550

464.

960.

50.

91.

61.

613

610

6.6

58.1

0.0

0.0

0.1

0.1

1.0

0.0

0.0

10

216

625

6044

0.15

0.0

0.0

0.3

0.3

8510

0.0

100.

00.

00.

00.

00.

00.

40.

00.

032

21

26

Tabl

e5:

Det

aile

dco

mpa

riso

nof

EB

BS,

PSC

,CPM

C,R

1,an

dR

2

Inst

ance

Auc

tione

erB

iddi

ngca

rrie

r,us

edpo

tent

ial

Bid

ding

carr

ier,

succ

ess

rate

(%)

Tim

e(s

)

Id|R|

cap

fa (WE

BB

S)

κPSC

1κC

PMC

1κR

11

κR2

1fb (W

EB

BS)

κPSC

2κC

PMC

2κR

12

κR2

2κE

BB

S3

κPSC

3κC

PMC

3κR

13

κR2

3E

BB

SPS

CC

PMC

167

2560

516.

600.

01.

42.

42.

425

710

0.0

29.6

0.0

0.0

0.1

0.2

2.2

0.0

0.0

00

216

825

6043

7.64

3.4

4.3

4.3

4.3

183

96.7

0.0

0.0

0.0

0.0

0.0

0.0

0.0

0.0

735

91

169

2560

626.

350.

01.

61.

61.

616

210

0.0

0.0

0.0

0.0

0.0

0.0

0.0

0.0

0.0

458

22

170

2560

396.

801.

00.

53.

21.

477

58.4

41.6

0.0

55.8

0.0

0.0

3.3

0.0

0.1

203

217

125

7042

1.64

0.0

0.8

1.0

1.0

168

100.

032

.70.

00.

00.

00.

00.

40.

00.

021

635

317

225

7046

6.35

0.0

0.1

0.1

0.1

8710

0.0

0.0

0.0

0.0

0.0

0.0

0.0

0.0

0.0

80

117

325

7043

4.06

0.0

4.5

6.2

6.2

258

100.

032

.20.

00.

00.

00.

00.

00.

00.

05,

149

9260

817

425

7055

7.85

1.8

2.5

5.4

5.4

270

58.9

100.

00.

00.

00.

00.

02.

30.

00.

02,

490

81

175

2570

348.

740.

02.

34.

04.

045

100.

019

5.6

0.0

0.0

0.0

0.0

2.2

0.0

0.0

142

341

176

2580

395.

480.

00.

30.

40.

415

910

0.0

49.7

0.0

0.0

0.0

0.0

0.9

0.0

0.0

1,13

523

31

177

2580

452.

250.

00.

80.

80.

861

100.

00.

00.

00.

00.

00.

00.

00.

00.

056

12

178

2580

417.

701.

32.

44.

64.

629

974

.264

.50.

00.

00.

00.

00.

20.

00.

024

,233

541

1817

925

8050

4.00

3.9

3.9

7.5

7.5

385

43.1

43.1

0.0

0.0

0.0

0.0

25.0

0.0

0.0

26,3

9290

218

025

8032

7.76

1.2

1.6

4.5

4.5

123

74.0

146.

30.

00.

00.

00.

00.

10.

00.

085

718

06

181

3050

588.

740.

41.

92.

31.

134

050

.933

.20.

027

.40.

00.

013

.30.

00.

150

44

182

3050

635.

93–

––

–0

––

––

––

––

–0

04

183

3050

546.

940.

00.

12.

70.

115

410

0.0

48.7

0.0

48.7

0.0

0.0

1.1

0.0

0.0

355

350

184

3050

900.

631.

43.

03.

03.

027

568

.00.

00.

00.

00.

00.

00.

00.

00.

052

826

518

530

5051

8.32

2.4

3.8

4.8

3.1

162

150.

026

.50.

023

.50.

00.

11.

80.

00.

039

65

186

3060

529.

490.

61.

92.

22.

232

920

.113

0.7

0.0

0.0

0.0

0.0

1.1

0.0

0.0

469

994

187

3060

585.

66-0

.01.

82.

82.

813

610

0.0

50.7

0.0

0.0

0.0

0.1

10.0

0.0

0.0

20

318

830

6051

2.68

1.9

1.3

7.2

6.8

349

80.8

121.

20.

024

.60.

00.

00.

00.

00.

24,

141

106

662

189

3060

773.

852.

72.

73.

93.

928

864

.634

.40.

00.

00.

00.

00.

60.

00.

07,

920

400

319

030

6045

4.22

0.7

2.9

4.2

4.2

123

36.6

100.

80.

00.

00.

00.

08.

10.

00.

044

713

65

191

3070

466.

062.

65.

25.

25.

223

385

.80.

00.

00.

00.

00.

00.

00.

00.

05,

216

233

319

230

7053

8.09

0.0

0.7

4.4

4.3

220

100.

062

.30.

027

.30.

00.

112

.50.

00.

246

23

193

3070

479.

642.

36.

26.

26.

227

588

.70.

00.

00.

00.

00.

00.

00.

00.

036

,019

773

319

430

7067

6.79

3.5

8.7

8.7

8.7

990.

017

3.7

0.0

0.0

0.0

0.0

0.2

0.0

0.0

66,2

343,

338

519

530

7039

3.02

2.7

4.3

4.3

4.3

133

66.9

0.0

0.0

0.0

0.0

0.0

0.0

0.0

0.0

3,52

891

55

196

3080

453.

820.

95.

07.

67.

235

697

.276

.10.

019

.10.

00.

00.

00.

00.

124

,388

1,65

232

719

730

8051

6.45

0.1

0.2

3.2

3.2

278

54.7

57.6

0.0

0.0

0.0

0.0

0.0

0.0

0.0

366

2622

198

3080

452.

932.

93.

411

.011

.027

797

.876

.50.

00.

00.

00.

00.

10.

00.

013

8,57

13,

949

230

200

3080

360.

284.

36.

98.

98.

920

670

.987

.40.

041

.70.

00.

00.

40.

00.

117

,276

4,34

91

201

3550

696.

960.

00.

63.

23.

130

910

0.0

66.7

0.0

29.8

0.0

0.0

0.0

0.0

0.0

8019

1520

235

5075

7.20

0.0

0.4

0.4

1.1

122

100.

049

.249

.20.

00.

10.

15.

02.

90.

00

04

203

3550

674.

830.

02.

66.

96.

939

710

0.0

81.9

0.0

0.0

0.0

0.0

30.8

0.0

0.0

1,20

115

53

204

3550

966.

871.

04.

97.

87.

728

966

.834

.60.

064

.40.

00.

00.

10.

00.

124

,941

10,9

534

27

Tabl

e5:

Det

aile

dco

mpa

riso

nof

EB

BS,

PSC

,CPM

C,R

1,an

dR

2

Inst

ance

Auc

tione

erB

iddi

ngca

rrie

r,us

edpo

tent

ial

Bid

ding

carr

ier,

succ

ess

rate

(%)

Tim

e(s

)

Id|R|

cap

fa (WE

BB

S)

κPSC

1κC

PMC

1κR

11

κR2

1fb (W

EB

BS)

κPSC

2κC

PMC

2κR

12

κR2

2κE

BB

S3

κPSC

3κC

PMC

3κR

13

κR2

3E

BB

SPS

CC

PMC

205

3550

672.

010.

62.

43.

61.

412

010

0.0

166.

70.

067

.50.

00.

02.

30.

00.

060

245

206

3560

626.

660.

34.

24.

44.

425

298

.834

.10.

00.

00.

00.

04.

30.

00.

01,

384

293

420

735

6068

8.44

0.0

1.0

2.7

2.7

319

100.

065

.20.

00.

00.

10.

20.

30.

00.

016

04

208

3560

615.

541.

23.

67.

37.

041

849

.899

.021

.317

.50.

00.

01.

92.

90.

011

,933

477

321

035

6060

4.65

1.1

0.1

6.0

3.4

355

59.7

101.

40.

023

.40.

00.

00.

00.

00.

192

630

770

211

3570

582.

312.

32.

98.

68.

347

371

.074

.00.

019

.20.

00.

01.

40.

00.

012

,182

914

221

235

7063

9.77

0.0

2.6

2.6

2.6

269

100.

00.

00.

00.

00.

00.

00.

00.

00.

013

38

321

335

7058

3.71

1.2

4.3

6.4

6.4

466

78.1

56.2

0.0

0.0

0.0

0.0

0.2

0.0

0.0

75,4

343,

830

2521

535

7053

5.41

0.8

1.4

7.0

5.0

258

114.

012

5.6

0.0

64.3

0.0

0.0

1.6

0.0

0.1

7,69

62,

288

321

635

8054

2.79

3.4

8.7

12.4

11.9

424

93.9

45.5

0.0

18.2

0.0

0.0

1.1

0.0

0.0

79,8

527,

956

421

735

8058

7.12

2.7

3.8

4.5

4.4

166

104.

831

.30.

032

.50.

00.

00.

20.

00.

21,

192

643

220

3580

478.

641.

84.

411

.59.

424

688

.689

.00.

028

.90.

00.

018

.20.

00.

135

,438

13,5

612

221

4050

780.

250.

36.

37.

13.

839

210

0.5

35.5

0.0

32.1

0.0

0.0

8.3

0.0

0.0

1,22

411

63

222

4050

891.

710.

00.

30.

90.

729

510

0.0

75.9

0.0

12.2

0.1

0.2

1.3

0.0

0.2

10

322

340

5078

8.17

0.2

1.9

4.6

4.5

240

80.0

79.2

0.0

26.7

0.0

0.0

1.1

0.0

0.0

5,26

31,

063

222

440

501,

203.

445.

26.

29.

87.

628

612

9.7

260.

10.

065

.70.

00.

00.

00.

00.

064

,483

26,7

931,

745

225

4050

853.

920.

11.

15.

32.

231

410

5.4

104.

10.

010

0.3

0.0

0.0

0.1

0.0

0.1

213

679

226

4060

729.

510.

74.

46.

96.

756

185

.652

.40.

018

.20.

00.

03.

20.

00.

012

,993

1,11

93

227

4060

767.

960.

02.

34.

13.

131

810

0.0

42.5

0.0

42.8

0.0

0.1

0.1

0.0

0.6

464

222

840

6072

7.64

2.2

7.8

9.7

9.7

673

81.4

54.5

0.0

0.0

0.0

0.0

5.8

0.0

0.0

62,6

864,

244

323

040

6072

6.17

1.4

4.0

8.6

4.9

294

91.2

122.

80.

039

.80.

00.

02.

20.

00.

13,

225

620

323

140

7064

4.21

3.4

9.1

11.9

10.6

412

94.9

18.2

0.0

42.5

0.0

0.0

11.1

0.0

0.0

81,1

307,

755

323

240

7071

8.35

1.0

3.7

4.8

4.8

264

117.

056

.10.

00.

00.

00.

00.

30.

00.

056

940

123

540

7064

7.48

0.7

3.0

4.4

4.4

326

106.

489

.60.

00.

00.

00.

02.

20.

00.

019

,735

6,83

43

237

4080

660.

410.

04.

14.

14.

143

273

.60.

00.

00.

00.

00.

00.

00.

00.

04,

195

347

3

min

0.0

0.0

0.0

0.0

0.0

0.0

0.0

0.0

0.0

0.0

0.0

0.0

0.0

00

0m

ax11

.123

.128

.622

.427

8.2

291.

418

0.0

112.

80.

50.

810

0.0

13.3

2.4

138,

571

32,6

352,

050

mea

n1.

43.

57.

86.

790

.880

.95.

628

.00.

00.

15.

70.

60.

29,

185

1,20

045

25%

-qua

ntile

0.0

1.0

2.9

2.2

79.6

44.9

0.0

0.0

0.0

0.0

0.1

0.0

0.0

30

1m

edia

n0.

82.

86.

35.

099

.887

.40.

018

.80.

00.

01.

20.

00.

015

29

375

%-q

uant

ile2.

44.

911

.99.

710

0.0

106.

70.

046

.20.

00.

14.

50.

20.

04,

203

313

4

28