COMEX: COMBINATORIAL AUCTIONS FOR THE

INTRA-ENTERPRISE EXCHANGE OF LOGISTICS SERVICES

Oleg Gujo

Institute of Information Systems, J. W. Goethe University

Mertonstrasse 17, D-60054 Frankfurt, Germany

Michael Schwind

Business Information Systems and Operations Research, Technical University Kaiserslautern

Erwin-Schr

odinger-Str. Geb. 42, D-67663 Kaiserslautern, Germany

Keywords:

Combinatorial exchange, logistic services, vehicle routing problem with time windows, proﬁt center.

Abstract:

The exchange of cargo capacities is an approach that is well established in the practice of logistics. Few

of these mostly web-based market places, however, are able to take synergies into consideration that can be

generated by the appropriate combination of the transportation lanes of different carriers. One way to achieve

this is to employ combinatorial auctions, that allow one to bid on bundles of lanes. This article describes a

combinatorial auction for the intra-enterprise exchange of logistic services. In the real world case considered

here, we implement and analyze such an exchange process in an enterprise that is related to the food sector

and organized in a proﬁt center structure. In the intra-enterprise exchange process, each proﬁt center is able

to release delivery contracts for outsourcing if the geographic location of a customer allows a reduced-cost

delivery by another proﬁt center in the neighborhood. The cost calculation is based on the results of an

integrated routing system, and the in and outsourcing process is managed by using the auction mechanism

ComEx. For the purpose of customer retention the delivery contracts are kept by the corresponding proﬁt

center, the incentive for exchanging the customers is achieved by a cost-savings distribution mechanism. After

a description of the web-based logistics auction together with the route optimization system DynaRoute, the

article describes the search for a cost optimizing strategy that bundles the appropriate delivery contracts.

1 INTRODUCTION

For several years combinatorial auctions (CA) have

been gaining increasing inﬂuence as an application

method in procurement and resource allocation pro-

cesses. Driven by the development of mechanisms

for the allocation of bandwidth in the frequency spec-

trum to telecommunication service providers in the

, Germany, and the US

, CAs came into the focus

of electronic market engineering (McMillan, 1995).

This article is concerned with an electronic CA for the

exchange of delivery contracts in a medium-sized en-

terprise which is organized in a proﬁt center structure.

The proﬁt centers, which are assigned to regional de-

livery areas are able to release a delivery contract to

http://www.ofcom.org.uk/consult/condocs/spectrum award/

http://www.fcc.gov

an adjacent proﬁt center if the delivery cost situation

is unfavorable for this contract. For the purpose of in-

centive compatibility, the cost savings achieved will

be distributed according to a previously deﬁned allo-

cation scheme. The issue of this work is the question

of the effective bundling and pricing of transportation

contracts, such that the interaction of the local route

optimization of the proﬁt centers and the CA will be

optimal with respect to the delivery time windows.

2 COMBINATORIAL AUCTIONS

IN LOGISTICS

One of the ﬁrst approaches to introducing the appli-

cation of CAs in the logistics sector was made by

(Caplice, 1996). (Caplice and Shefﬁ, 2003) com-

Gujo O. and Schwind M. (2007).

COMEX: COMBINATORIAL AUCTIONS FOR THE INTRA-ENTERPRISE EXCHANGE OF LOGISTICS SERVICES.

In Proceedings of the Ninth International Conference on Enterprise Information Systems - SAIC, pages 5-12

DOI: 10.5220/0002351500050012

 SciTePress

b) c)

Figure 1: Exchange of customers among two neighboring proﬁt centers.

bine a route planning process with the allocation of

transportation capacity by using a CA that selects the

cost-minimal combination of delivery contracts. A

related approach has been proposed by (Regan and

Song, 2003) who suggested a spot market for lo-

gistic services that are in excess supply or required

in the short-term. A model published by (Pankratz,

1999) leads in the same direction, however, the fo-

cus is more on incentive compatibility for the bidders.

Some providers of logistics software and operators

of freight exchanges have already introduced simple

CA mechanisms into their route planning and freight

allocation methods.

(Elmaghraby and Keskinocak,

2005) document a two-step procurement auction for

transportation capacities that has been organized by

the home improvement chain Home Depot to ensure

the logistics supply of about thousand stores. In co-

operation with i2-Technologies a ﬂexible auction soft-

ware was developed to support the bidders in formu-

lating the appropriate bid combinations which pro-

vide the optimal synergy effects between the routes.

The formulation of bid bundles is a critical point in

CAs, because it represents a combinatorial optimiza-

tion problem that cannot be solved without technical

support (Schwind, 2005). For this reason, the ComEx

system presented here is designed to make the in and

outsourcing decisions automatically in combination

with the bundling of the bids for transportation con-

tracts supported by the DynaRoute optimization sys-

tem provided by our start-up company VARLOG

3 FUNCTIONAL AND SYSTEM

DESCRIPTION OF COMEX

In the following we provide an overview of the func-

tional principle of the CA for logistics services ex-

change and present the interaction of the auction and

optimization process in the ComEx system. Fig. 1

depicts the exchange process between two proﬁt cen-

ters. The scenario is close to that presented in a recent

approach by (Krajewska and Kopfer, 2006), which is

also based on a proﬁt center structure and makes use

www.combinenet.com, www.i2.com, www.nex.com

www.varlog.de

of the matrix auction (Gomber et al., 1999).

Each proﬁt center disposes of an individual set

of customers. For simplicity, these customers have

already been assigned to a preliminary delivery route

(see Fig. 1a). Thereafter a decision is made, based on

the geographical location, which customers should

deﬁnitely remain in the delivery process of the out-

sourcing proﬁt center. These customers are assigned

to the ‘ﬁxed area’ (see Fig. 1b ellipses shaded in

gray). The remaining customers are merged into

clusters with respect to the geographical location and

the delivery time window (see Fig. 1b, cluster C

containing one customer, cluster C

containing two

customers). For each cluster, the cost that could be

saved by outsourcing all customer delivery contracts

to a neighboring proﬁt center, is calculated. Then

an information interchange is performed among

the proﬁt centers aimed at deﬁning the outsourcing

candidates of some neighboring proﬁt centers as

insourcing candidates of the other proﬁt centers.

Subsequently each proﬁt center investigates whether

the insourcing candidates ﬁt into the existing delivery

routes according to delivery time and geographical

position. If this is the case, the difference in delivery

cost is calculated by including and excluding the

set of customers that is added by insourcing the

candidate cluster. This difference in cost is then used

as the bid price of a cluster while performing the

ComEx auction. A bid in the ComEx auction can also

consist of more than one cluster. After each proﬁt

center has submitted its bids the CA takes place,

searching for the optimal allocation of bids to proﬁt

centers that minimizes the total delivery costs in

the enterprise. After the closing of the auction the

transportation routes will be recalculated based on

the information about the customers that have been

assigned to it in the ﬁnal allocation (see Fig. 1c).

The ComEx system consists of four components:

the ComEx server that controls the entire auction pro-

cess, the ComEx engine, which is responsible for

the calculation of the optimal allocation, the ComEx

clients that administer all customers in the delivery

area of a proﬁt center while formulating, submitting

and receiving in and outsourcing bids, and the Dy-

naRoute server that provides the optimal routing in-

formation and the associated cost of delivery. In the

ICEIS 2007 - International Conference on Enterprise Information Systems

following we describe the interaction of the system

components in the four phases of the ComEx auction:

• Initialization phase: Auction format settings, reg-

istration, and licensing data of the participants are

transmitted to the ComEx server ➀ using a special

XML format CAMeL (Schwind et al., 2004).

• Outsourcing: For each proﬁt center the customers

are grouped into the clusters taking the time win-

dows and the geographical position into account.

The outsourcing candidates are selected based on

their distance from the proﬁt center

B . For each

cluster a request is sent ➁ to the DynaRoute server

to determine the costs which can be saved by out-

sourcing the included customers

A . The infor-

mation about the outsourcing candidates is sent

by every proﬁt center to the ComEx server. The

proﬁt centers receive the list of clusters which are

designated to be outsourced ➂.

• Insourcing: Every ComEx client examines for

each proﬁt center whether the insourcing candi-

dates can be added to the existing set of customers

while taking their temporal and geographical rela-

tion to the remaining customers into account

C .

If insourcing is possible, another request to the

DynaRoute server ➃ is sent in order to calculate

the additional delivery costs

A . This information

is used as the bid price of a cluster. The ComEx

server collects these bids, which are passed over

by the clients ➄ and initiates the CA ➅. An op-

timal allocation is determined

D and the proﬁt

centers are informed about the results ➆.

• Final assessment phase: The route plans are

updated➇

A and the cost savings in the enterprise

are calculated➈.

3.1 DynaRoute Server

Each ComEx client has to solve an instance of

the Time Dependent-Stochastic Capacitated Vehicle

Routing Problem with Time Window (TD-CVRP-TW)

which extends the

N P -complete Vehicle Routing

Problems with Time-Windows by driving times that

depend on the time of day and different categories of

time windows. Case studies show that the time win-

dow’s start can vary to a certain degree while lead-

ing an acceptable optimization result. The degree

to which a tour does not comply with the time win-

dows is penalized in evaluation function. By using

DynaRoute the clients are able to optimize their in-

dividual tours (Wendt et al., 2005). The optimization

algorithm of the DynaRoute server is an extended ver-

sion of Cooperative Simulated Annealing (COSA), an

combination of the meta-heuristics Simulated Anneal-

ing (SA) and Genetic Algorithms (GA) proposed by

(Wendt, 1994). Due to the high complexity that re-

sults from the TD-CVRP-TW the COSA algorithm

is extended by several concepts to improve perfor-

mance:

• Compression of the solution space: Due to the

fact that the evaluation function of the TD-CVRP-

TW consumes considerable time cycles, the per-

formance of the algorithms can be increased by

excluding solution candidates by using stochas-

tic stop criteria when evaluating new candidates.

Moreover, DynaRoute identiﬁes patterns by using

algorithms similar to ant systems (Bullnheimer

et al., 1999) to direct the search to efﬁcient ‘re-

gions of the search space’.

• Tabu-lists: In the context of SA algorithms

the use of tabu lists resulted in a considerable

improvement of the performance (Li and Lim,

2003). By using tabu lists (Glover, 1986) Dy-

naRoute is able to temporarily exclude solutions

or solution regions from the search space. This

results in greater diversity of the search process

and thus in a higher probability of ﬁnding the

optimal solution.

• Approximation of the ﬁtness of inferior solutions:

By reducing the time required for the process of

solution evaluation the overall performance of

the algorithm can be considerably improved. An

exact evaluation of the candidates is not always

necessary but can be replaced by an approxima-

tion if the accuracy of the approximation is within

certain limits. The DynaRoute uses the dynamic

approximization rules, which adapt the accuracy

of the evaluation depending on the progress of

the optimization process.

In the context of the ComEx framework the Dy-

naRoute server optimizes the client speciﬁc TD-

CVRP-TW and approximates the cost of in and out-

sourcing customer clusters. The ComEx client spec-

iﬁes existing clusters of customers or provides new

ones. Based on the generated population of solu-

tions the DynaRoute server approximates the impact

of these clusters on the cumulative costs of the tour.

By running a short reactivation of the optimization

process it is possible to integrate or exclude the given

clusters in a new ‘near optimal’ solution. In this way,

the cost of adding a new cluster or the savings made

when removing a cluster can be very quickly calcu-

lated by comparing the original solution to the new

solution. The population-based approach allows the

server to check the k-best solutions of both popula-

tions, which provides a strong indicator of the opti-

COMEX: COMBINATORIAL AUCTIONS FOR THE INTRA-ENTERPRISE EXCHANGE OF LOGISTICS SERVICES

Customer orders n

ComEx Server ComEx Engine

DynaRoute Server

ComEx Client 1

Customer orders 1

CAMeL

ComEx Client n

CAMeL

LP SOLVE CPLEX

- Communication

- Process

Legend:

2 84

B C

31 5 9

Figure 2: Communication and process ﬂow in the ComEx system.

mality of the solution. This strongly reduces the risk

of wrong calculations by minimizing the error that

may occur for a single solution due to a ‘weaker opti-

mization’. This enables the DynaRoute server to deal

with the complex cost estimates for the clusters iden-

tiﬁed by the ComEx client.

3.2 ComEx Client

For each delivery contract each proﬁt center i decides

whether to keep the customer in the set of contracts

that should be served anyway or to release the cus-

tomer to the outsourcing process. In the former case

the customers are located geographically close to the

particular proﬁt center. As already mentioned above,

such customers deﬁne the ‘ﬁxed area’ P

. The re-

maining customers are assigned to the ‘outsourcing

area’ P

3.2.1 Clustering and Outsourcing

The clustering process aims to reduce the complexity

of the underlying combinatorial optimization problem

and to identify synergies amongst the particular de-

livery contracts. To achieve this, customers are ﬁrst

ranked according to their geographical position and

their delivery time window and subsequently clus-

tered. In our context a cluster C

deﬁnes an undirected

path, that can be part of a route, that will be deter-

mined by a planning algorithm later on. The cluster-

ing is independent of the route planning process and

aims to group customers that are located geographi-

cally close together such that they can be served by

a single delivery vehicle with respect to the delivery

time constraints. At the beginning of the clustering

process each path includes only one single customer.

In order to ﬁnd as many combinations as possible, two

strategies are employed in two consecutive phases.

During the ﬁrst phase the clustering algorithm tries

to extend a path at its beginning or end by attaching

another path to it. In the second phase the clustering

CLUSTERINGFIRSTPHASE(V, E)

1 E

′

:= {{a

, b

} |a

∈ C

, b

∈ C

, |{a

, b

}| ≤ r};

2 Sort E

′

according to increasing distance |{a

, b

}|;

3 for all {a

, a

} ∈ E

′

(* sorted *)

4 do if ((a

end node ∧ b

end node ) ∧

5 (a

∧ b

not in the same cluster ))

6 then bind both clusters with {a

, b

};

7 if (combined cluster is valid)

8 then save combined cluster C

;

Figure 3: Clustering algorithm: phase one.

algorithm tries to include a path into another. Let E be

the set of edges, V the set of vertices in an undirected

graph G while e

new

= {a

, b

} ∧ a

, b

∈ V is a pair of

customers a

and b

and |e

new

| deﬁnes the distance be-

tween them. To illustrate the algorithm of phase one,

Fig. 3 depicts the corresponding pseudo code.

Initially the algorithm identiﬁes the set of con-

nections that can be constructed by employing the

permutations of all pairs of customers a

and b

in the

proﬁt center area. This process considers only these

customers, that are located within a maximum dis-

tance r to each other (line 1). The customer pairs are

sorted according to the inclining distance (line 2). The

pairs will be considered by the algorithm following

this order (line 3). If the customers a

and b

of such a

pair are both terminal points of two disjoint clusters, a

cluster will be constructed by concatenating them by

a new edge (lines 4 to 6). Subsequently the validity of

the newly generated cluster is checked. The validity

of a route is exactly given if the customers of a cluster

can be served by using one delivery vehicle within

the given time windows by choosing at least one of

both possible directions. If validity is given, the route

will be saved (lines 7 and 8).

The algorithm of the second phase aims to incor-

porate existing clusters into others. This means that

the route of a cluster is interrupted to integrate the cus-

tomers of another cluster. Only if this has happened,

the processing of the original route is continued by

the algorithm. For the integration process of a cluster

ICEIS 2007 - International Conference on Enterprise Information Systems

b1 b2 b3

a1 a2 a3

b1 b2 b3

a1 a2 a3

b1 b2 b3

a1 a2 a3

new

keep

remove

new

remove

new

a) b) c)

Figure 4: Two variants of clustering: a) ﬁrst insertion variant b) initial situation c) second insertion variant.

into another cluster, two variants depicted in Fig. 4

are conceivable.

In the beginning a connection between an end

node of the ﬁrst cluster and an inner node of the sec-

ond cluster is established (see Fig. 4 b). Now two po-

tential connections are determined that allow the in-

clusion of the ﬁrst cluster C

into the second cluster

(Fig. 4 a, 4 c). In analogy to the ﬁrst phase, the

newly created path must not include a connection that

has a length greater than r.

Fig. 5 shows the second phase of the clustering al-

gorithm. Initially an empty set E

′′

is generated (line 1)

and successively ﬁlled by investigating each element

of E

′

to see whether it is capable of representing the

ﬁrst of two potential connections between both clus-

ters (lines 2 to 3). This is the case if customers a

and

are not included in the same cluster and exactly one

of the customers represents a terminal point of a clus-

ter (lines 4 and 5). After determining both variants for

the second connection (lines 6 and 7) the correspond-

ing tuple of edges is added to the set E

′′

(lines 8 to

10).

Subsequently the elements of the set E

′′

are or-

dered according to their total distance (line 11) and a

veriﬁcation is made whether the tuple of edges leads

to a valid integration of a cluster into another cluster

(lines 12 to 16). If the validity of a cluster is guaran-

teed the cluster is stored (line 17 and 18).

Let I ∈ N be the number of proﬁt centers in the

enterprise. For each proﬁt center i the customers are

sorted into two groups according to their geographi-

cal distance to the proﬁt center: a deﬁned percentage

of customers with a distance far from the proﬁt center

is assigned to the outsourcing area P

, the remaining

customers that are closer to the proﬁt center constitute

the ‘ﬁxed area’ P

. For each proﬁt center i a set C

formed that contains all clusters in P

that have been

formed for proﬁt center i. Each proﬁt center commu-

nicates the set C

to the ComEx server that groups the

sets to C

∗

i=1

. Finally the set C

∗

is communi-

cated to all proﬁt centers.

CLUSTERINGSECONDPHASE(V, E)

1 E

′′

2 E

′

= {{a

, b

}|a

∈ C

, b

∈ C

, |{a

, b

}| ≤ r};

3 for all e

new

= {a

, b

} ∈ E

′

4 do if ((a

inner node ∧ b

end node) ∧

5 (a

∧ b

not in the same cluster))

6 then for all e

keep

= {b

, b

∗

} ∧

7 l 6= l

∗

∧ b

∈ e

new

8 do determine e

∗

new

, e

remove

9 ∆ = |e

new

| + |e

∗

new

| − |e

remove

10 E

′′

← (e

new

, e

keep

)

11 sort(e

new

, e

keep

) ∈ E

′′

in according to ∆

12 for all (e

new

, e

keep

) = ({a

, b

13 {b

, b

∗

}) ∈ E

′′

(* sorted *)

14 do if (a

, b

∗

not in the same cluster)

15 then Integrate C

according to the

16 edges e

new

, e

∗

new

, and e

keep

into C

;

17 if (combined cluster is valid)

18 then save combined cluster C

;

Figure 5: Clustering algorithm: phase two.

3.2.2 Bid Formation and Insourcing

Relevant literature describes multiple variants of bid-

ding languages and related bidding languages (Nisan,

2000). The ComEx system uses a OR-of-XOR bid

logic. This means that XOR-bundles facilitate the for-

mulation of bid combinations that are exclusively se-

lectable, while OR bundles allow for the selection of

more than one alternative bid bundle.

Due to its size the number of elements in C

∗

must be reduced to limit computational effort. This

reduction is achieved by removing clusters with a

large distance from the insourcing proﬁt center from

the set C

∗

of customer candidates. Subsequently the

reduced set C

∗

is used by the insourcing proﬁt center

to construct new combinations while merging clusters

of its own customer set with clusters of the other

outsourcing proﬁt centers (this is done by using the

clustering method described in section 3.2.1). Finally

the bids of a proﬁt center are constructed as follows:

Let M

∈ N be the number of clusters C

∈ C

∗

that are

COMEX: COMBINATORIAL AUCTIONS FOR THE INTRA-ENTERPRISE EXCHANGE OF LOGISTICS SERVICES

ComEx-Server

ComEx Client 3

XOR XOR

ComEx Client 2ComEx Client 1

AtomicBid b

Figure 6: Bid formation logic of ComEx.

available for insourcing into proﬁt center i ≤ I, then

= C

∧C

∧. . . ∧C

, α 6= β 6= δ∧α, β, δ ≤ M

∧ j ∈

N is deﬁned as atomic bid of proﬁt center i. For each

atomic bid b

a request is sent to the DynaRoute

server to determine the delivery costs associated with

this bid. At this point, we will not discuss the game

theoretic implications concerning truthful bidding

and incentive compatibility, because the delivery

costs that provide the basis for the formation of the

bid prices in each proﬁt center are calculated solely

by the DynaRoute server and are therefore difﬁcult to

manipulate (Schwind, 2005).

Let B be the composition of all bids of a proﬁt cen-

ter that are linked by the logic operators OR or XOR.

Different atomic bids of a proﬁt center may include

the same clusters. In this case bids have to be linked

by XOR operators. If the subsets of atomic bids are

disjoint, the OR operator is employed to link the bids,

e.g. B

= b

∨ b

⊕ b

(Fig. 6).

The ComEx system provides a graphical user in-

terface to support the proﬁt centers in administering

the customers and visualizes the optimization result.

Each proﬁt center holds a table of customers that con-

tains the information required for their administration,

e.g. the geographical position of the customer. Cus-

tomers that should be deﬁnitely kept in the set of de-

livery contracts that are served by a proﬁt center (e.g.

for the reason of a high turnover), can be labeled and

thus assigned to the area of ﬁxed delivery contracts.

After initiating the route optimization process the re-

sults of the particular optimization steps are succes-

sively visualized by the system (e.g. the outsourc-

ing of released clusters, the bid prices calculated for

the released cluster bundles, or the delivery contracts

taken over from other proﬁt centers). After the opti-

mization process is ﬁnished, the ComEx system indi-

cates, how much cost has been saved within the enter-

prise by employing the CA exchange process, in re-

lation to the sole application of DynaRoute. The ﬁnal

route optimization result based on the new customer

allocation is visualized on a map.

3.3 ComEx Server

The ComEx server has the task of controlling the auc-

tion process. The ComEx server receives the bids

submitted by the proﬁt center via web-based requests

by using a servlet. The information interchange be-

tween the clients and the ComEx server is handled

using SOAP messages. Our system allows the trans-

mission of messages using the XML-based language

CAMeL, that has been designed for the standardized

submission of bids in CAs (Schwind et al., 2004). The

bids of all proﬁt centers are stored in a data structure

∗

= B

∨ B

∨ . . . ∨ B

(see Fig. 6). All SOAP mes-

sages are encrypted using SSL technology during the

transmission process in the web. The bids of all proﬁt

centers are stored in a corresponding data structure

and are transferred to the ComEx engine where the

calculation of the CAP is initiated. The resulting op-

timal allocation is reported to the ComEx clients via

the ComEx server. Additionally, the ComEx server

calculates the cost savings for the entire enterprise re-

sulting from the application of the CA.

3.4 ComEx Engine

As already mentioned, the ComEx server forwards the

bids B

of each ComEx client i to the ComEx engine

employing CAMeL messages. In order to ﬁnd the op-

timal allocation the ComEx engine solves the com-

binatorial auction problem (CAP) (Vries and Vohra,

2001). Let z be delivery costs in the enterprise, M

the set of all clusters to be allocated to the proﬁt cen-

ters, J

the number of all bids of proﬁt center i, and

i, j

) = 1 if atomic bid b

i, j

contains cluster C

, then

the CAP for the combinatorial exchange linked to the

construction of optimal tours can be formulated:

min z =

∑

i=1

∑

j=1

(1)

∑

i=1

∑

j=1

) x

= 1 ∀C

∈ C

∗

∧ x

∈ {0, 1} (2)

) =

(

1, if C

in b

0, otherwise

∀C

∈ C

∗

(3)

The acceptance variable x

indicates whether a

bid should be accepted or not and guarantees that

the integer condition is maintained. Equation (2) as-

sures that all clusters appear in the ﬁnal allocation.

ICEIS 2007 - International Conference on Enterprise Information Systems

The open source software package LP SOLVE 5.5 is

used to solve the CAP. The package provides an ex-

act solution of the

N P -complete problem for up to

hundreds of bids within acceptable computation time.

For higher problem complexities the use of heuristics

is recommendable (Schwind and Gujo, 2005).

4 THE IMPACT OF CLUSTERING

PROCESS ON DELIVERY

COSTS

The main objective of the ComEx system is to pro-

vide a signiﬁcant reduction of delivery costs for the

entire enterprise. This can only be achieved if a suit-

able strategy for the preparation of bids in the in- and

outsourcing of delivery contracts is applied. In our

framework two impact factors are of importance: The

maximum distance r between two neighboring cus-

tomers within a cluster that forms a possible section

of a delivery tour and the percentage a of clusters that

deﬁnitely remain under the service of a proﬁt cen-

ter. Additionally, the impact of the pricing method

for bids submitted by each proﬁt center is of interest

with respect to the total delivery costs. For the simula-

tions the parameters were varied within the following

limits:

• Maximum distance r between two customers in a

cluster: 2− 20 km (steps of 2 km)

• Percentage a of clusters in the outsourcing area:

20%− 100% (steps of 20%)

The simulations were repeated multiple times to

reduce the stochastic effects that result from the use

of the heuristic procedure in the DynaRoute tour opti-

mization process.

Fig. 7 shows that a maximum distance of r = 10

km between the customers in a cluster and a percent-

age a of 40% of clusters in the outsourcing area yields

the largest reduction in overall transportation costs. If

the maximum allowed distance r is set to a high level

in the bid preparation process, the single clusters con-

tain only few customers. This means that the transfer

of customers can only cause small synergy effects.

On the other hand, if the maximum allowed dis-

tance r in the clustering process is set to a high value,

very complex clustered tours are released into the out-

sourcing process. Such clusters are difﬁcult to inte-

grate into the existing tours of the other proﬁt centers.

The second simulation method strives to achieve

computational time savings by omitting the Dy-

naRoute-based calculation of delivery costs for each

outsourcing candidate. Instead, for each cluster the



 



 















 



 















 



 







Figure 7: Reduction of delivery costs by the application of

the DynaRoute-based stategy (ﬁrst simulation).



 



 



Figure 8: Reduction of delivery costs by the application of

the distance-based strategy (second simulation).

average of distances between the customers and the

proﬁt center is used to estimate the bid price. All

other simulation settings are kept identically to the

previous simulation. Fig. 8 shows that a maximum

distance of r = 10 km between the clients in a cluster

and a percentage a of 60% of clusters in the ﬁxed area

lead to the largest reduction in transportation costs

in the entire enterprise. The differences between

the simulations occur because in the ﬁrst case the

integratability of a cluster into an existing route is

crucial for the pricing of the bids. Obviously, pricing

using a distance estimate performs signiﬁcantly

for a large set of outsourcing candidates and small

distances in the clusters (Fig. 8 lower left). By

contrast, if the concatenation of customers to longer

clusters is allowed, the combinatorial complexity of

CA-based tour recombination is dominant. This leads

to a slightly better performance of DynaRoute-based

COMEX: COMBINATORIAL AUCTIONS FOR THE INTRA-ENTERPRISE EXCHANGE OF LOGISTICS SERVICES

pricing (Fig. 7 upper middle). The results indicate

that bid formation must be improved.

5 CONCLUSION

This paper presented the software system ComEx for

the auction-based exchange of transportation services

in connection with the route optimization system Dy-

naRoute. The evaluation focuses on the search for

an optimal mechanism that helps to formulate the ap-

propriate in and outsourcing bids for the transfer of

delivery contracts between neighboring proﬁt centers,

such that a transportation cost reduction results for the

entire enterprise. By using real world delivery data

from an enterprise in the food industry, an optimal

maximum distance of 10 km between two neighbor-

ing customers in a cluster and an optimal fraction of

40% (60% in case of the second simulation) of cus-

tomers kept in the ﬁxed delivery area have been iden-

tiﬁed. Further, our ComEx mechanism has a much

greater importance for the exchange of transportation

services between independent enterprises than for the

intra-enterprise sector, e.g. as a general combinato-

rial freight exchange. Such a realization of a logis-

tics marketplace, however, raises questions about an

incentive compatible method for evaluating the bids,

for example the Vickrey-Clarke-Groves mechanism.

In our case, this problem has been circumvented by

the use of identical automated bid construction mech-

anisms for all participants. However, the question of

a fair and incentive compatible distribution of the cost

savings between the proﬁt centers remains. Together

with the implementation of a general exchange for lo-

gistic services together with a major logistic provider,

and the improvement of the clustering and pricing

mechanism, this will be the next issue in our research.

REFERENCES

Bullnheimer, B., Hartl, R., and Strauss, C. (1999). An im-

proved ant system algorithm for the vehicle routing

problem. An. of Operations Research, 89:319–328.

Caplice, C. (1996). An Optimization Based Bidding Pro-

cess: A New Framework for Shipper-Carrier Rela-

tionships. PhD thesis, MIT, Massachusetts.

Caplice, C. and Shefﬁ, Y. (2003). Optimization-based pro-

curement for transportation services. Journal of Busi-

ness Logistics, 24(2).

Elmaghraby, W. and Keskinocak, P. (2005). Combinato-

rial auctions in procurement. In Harrison, T. P., Lee,

H. L., and Neale, J. J., editors, The Practice of Supply

Chain Management: Where Theory and Application

Converge, pages 245–258. Springer, Berlin.

Glover, F. (1986). Future paths for integer programming

and links to artiﬁcial intelligence. computers and Op-

erations Research, 13(5):533–549.

Gomber, P., Schmidt, C., and Weinhardt, C. (1999). Auc-

tions in electronic commerce: Efﬁciency versus com-

putational tractability. In Proceedings of the Interna-

tional Conference on Electronic Commerce 98, pages

43–48, Seoul, Korea.

Krajewska, M. A. and Kopfer, H. (2006). Collaborat-

ing freight forwarding enterprises. OR Spectrum,

28(3):301–436.

Li, H. and Lim, A. (2003). Local search with annealing-

like restarts to solve the VRPTW. European Journal

of Operational Research, 150:115–127.

McMillan, J. (1995). Why auction the spectrum? Telecom-

munications Policy, 19:191–199.

Nisan, N. (2000). Bidding and allocation in combinatorial

auctions. In Proceedings of the 2.nd ACM Conference

on Electronic Commerce (ACM EC’00), Minneapolis,

MN, pages 1–12. ACM.

Pankratz, G. (1999). Analyse kombinatorischer Auktio-

nen f

ur ein Multi-Agentensystem zur L

osung des

Groupage-Problems kooperierender Speditionen. In

Inderfurth, K., Schw

odiauer, G., Domschke, W.,

Juhnke, F., and W

ascher, G., editors, Operations Re-

search Proceedings, pages 443–448. Springer Verlag.

Regan, A. and Song, J. (2003). An auction based collabora-

tive carrier network. Transportation Research Part E:

Logistics and Transportation Review.

Schwind, M. (2005). Design of combinatorial auctions for

allocation and procurement processes. In 7th Interna-

tional IEEE Conference on E-Commerce Technology

2005, M

unchen, Germany, pages 391–395.

Schwind, M. and Gujo, O. (2005). An agent-based sim-

ulation environment for a combinatorial grid sched-

uler. Technical report, Institute of Information Sys-

tems, J. W. Goethe University, Frankfurt, Germany.

Schwind, M., Weiss, K., and Stockheim, T. (2004). CAMeL

- Eine Meta-Sprache f

ur Kombinatorische Auktionen.

Technical report, Institut f

ur Wirtschaftsinformatik,

Johann Wolfgang Goethe Universit

at, Frankfurt.

Vries, S. D. and Vohra, R. (2001). Combinatorial auc-

tions: A survey. INFORMS Journal on Computing,

15(3):284–309.

Wendt, O. (1994). Naturanaloge Verfahren zur approxi-

mativen L

osung kombinatorischer Optimierungsprob-

leme. PhD thesis, University of Frankfurt, Germany.

Wendt, O., Stockheim, T., and Weiss, K. (2005). Intelli-

gente Tourenplanung mit DynaRoute. Wirtschaftsin-

formatik, 47(2):135–140.

ICEIS 2007 - International Conference on Enterprise Information Systems