**For full functionality of ResearchGate it is necessary to enable JavaScript.**

Here are the

instructions how to enable JavaScript in your web browser .

Here are the

instructions how to enable JavaScript in your web browser .

See all ›

4 Citations

See all ›

81 References

See all ›

3 Figures

Download full-text PDF

# Dynamic Multi-Linked Negotiations in Multi-Echelon Production Scheduling Networks

**Conference Paper (PDF Available)**· January 2006 with 83 Reads

DOI: 10.1109/IAT.2006.56 · Source: DBLP

Conference: Conference: Proceedings of the 2006 IEEE/WIC/ACM International Conference on Intelligent Agent Technology, Hong Kong, China, 18-22 December 2006

Conference: Conference: Proceedings of the 2006 IEEE/WIC/ACM International Conference on Intelligent Agent Technology, Hong Kong, China, 18-22 December 2006

Hoong Chuin Lau

- 27.08
- Singapore Management University

Guan Li Soh

Wee Chong Wan

Abstract

In this paper, we are concerned with scheduling resources in a multi-tier production/logistics system for multi-indenture goods. Unlike classical production scheduling problems, the problem we study is concerned with local utilities which are private. We present an agent model and investigate an efficient scheme for handling multi-linked agent negotiations. With this scheme we attempt to overcome the drawbacks of sequential negotiations and negotiation parameter settings. Our approach is based on embedding a credit-based negotiation protocol within a local search scheduling algorithm. We demonstrate the computational efficiency and effectiveness of the approach in solving a real-life dynamic production scheduling problem which balances between global production cost and local utilities within the facilities.

## Discover the world's research

**15+ million**members**118+ million**publications**700k+**research projects

Join for free

Figures

Framework for early warning system in food supply networks

A simple case of the distribution network under investigation

Diagram of JAIST Nanatsudaki Model

VEAM and LOGISTICS AND SUPPLY

CHAIN MANAGEMENT Workshops

VEAM IFIP Working Group 7.6 Workshop on Virtual

Environments for Advanced Modelling

4th US-European Workshop on Logistics and Supply

Chain Management – An International Research

Perspective

June 6-9, 2006; Hamburg, Germany

Stefan Voß

Imke Sassen

(Eds.)

IWI

IWI

HAMBURG

Institute of

Information

Systems

University

of Hamburg

In partnership and

supported by Deutsche

Post World Net

International

Federation

for

Information

Processing

VI

All rights preserved. No part of this book may be reproduced, stored in a

retrieval system, or transmitted, in any form or by any means, without the

prior written permission of the publisher.

The publisher is not responsible for the use which might be made of the

information contained in this book.

Published by:

Institute of Information Systems

University of Hamburg

Von-Melle-Park 5

20146 Hamburg, Germany

http://iwi.econ.uni-hamburg.de

Printed in Germany

Preface

“Hamburg — Das Tor zur Welt.”

This volume includes contributions submitted for two workshops held back-

to-back in June 6-9, 2006 in Hamburg, Germany. The workshops are

•VEAM IFIP Working Group 7.6 Workshop on Virtual Environments for

Advanced Modelling

•4th US-European Workshop on Logistics and Supply Chain Management

– An International Research Perspective

In this volume we have put together the accepted contributions for those two

workshops. They constitute work of researchers and practitioners from more

than a dozen countries from all over the world. To ensure a fruitful collabo-

ration and to keep the workshop character with open discussion we have not

strived towards unifying the contributions regarding, e.g., style, length, etc.

They are provided in alphabetical order of the authors.

We greatly appreciate the ﬁnancial support from Deutsche Post World Net

which especially helped us to organize the 4th US-European Workshop on

Logistics and Supply Chain Management, and to have it in Germany for the

ﬁrst time. We also like to mention the Hamburger Hafen und Logistik AG

(HHLA), who made it possible to visit one of their container terminals within

this week. Moreover, we would like to acknowledge the assistance of everybody

at the Institute of Information Systems (IWI) at the University of Hamburg

for their valuable support.

We wish everybody successful and enjoyable days in Hamburg.

Stefan Voß

Imke Sassen

Hamburg, June 2006

VIII

Contents

Part I Contributions VEAM Workshop

Events in Modeling of Complex Systems

Janusz Granat …………………………………………… 3

SimConT: A Tool for Quick Layout and Equipment Portfolio

Evaluation and Simulation of Hinterland Container Terminal

Operations

Manfred Gronalt, Thouraya Benna, Martin Posset ………………. 7

Dynamic Multi-Linked Negotiations in Multi-Echelon

Production Scheduling Networks

Hoong Chuin Lau, Guan Li Soh, Wee Chong Wan ………………. 10

Modelling Classiﬁcation Analysis for Competitive Events

with Applications to Sports Betting

Stefan Lessmann, Johnnie Johnson, Ming-Chien Sung ……………. 14

Structured Modeling Technology: Recent Developments and

Open Challenges

Marek Makowski …………………………………………. 16

Production Planning with Load Dependent Lead Times and

Deterioration

Julia Pahl, Stefan Voß, David L. Woodruﬀ …………………….. 19

Guided Online Decision Making

J¨orn Sch¨onberger, Herbert Kopfer ……………………………. 22

Scalability of Three Parallel Direct Search Methods in

Simulation-Based Optimization

Frank Thilo, Manfred Grauer ……………………………….. 25

X Contents

Nanatsudaki Model of Knowledge Creation Processes

Andrzej P. Wierzbicki, Yoshiteru Nakamori ……………………. 31

The Use of Reference Proﬁles and Multiple Criteria

Evaluation in Knowledge Acquisition from Large Databases

Andrzej P. Wierzbicki, Jing Tian, Hongtao Ren ………………… 34

Convex Envelope for Medical Modeling

Fadi Yaacoub, Yskandar Hamam, and Charbel Fares ……………… 36

Applying Data Mining for Early Warning and Proactive

Control in Food Supply Networks

Li Yuan, Mark R. Kramer, Adrie J.M. Beulens …………………. 38

Part II Contributions Logistics and SCM Workshop

Optimizing Inventory Decisions in a Multi-Stage Supply

Chain Under Stochastic Demands

Ab Rahman Ahmad, M. E. Seliaman …………………………. 45

Impact of E-Commerce on an Integrated Distribution

Network

Daniela Ambrosino, Anna Sciomachen ………………………… 46

An Interval Pivoting Heuristic for Finding Quality Solutions

to Uniform-Bound Interval-Flow Transportation Problem

Aruna Apte, Richard S. Barr ……………………………….. 49

Managing the Service Supply Chain in the US Department of

Defense: Opportunities and Challenges

Uday Apte, Geraldo Ferrer, Ira Lewis, Rene Rendon ……………… 50

Analysis of Heuristic Search Methods for Scheduling

Automated Guided Vehicles

Thomas Bednarczyk, Andreas Fink …………………………… 52

Exact and Approximate Algorithms for a Class of Steiner

Tree Problems Arising in Network Design and Lot Sizing

Alysson M. Costa, Jean-Fran¸cois Cordeau, Gilbert Laporte ………… 54

Supply Chain Management in Archeological Surveys,

Excavations and Scientiﬁc Use

Joachim R. Daduna, Veit St¨urmer …………………………… 55

Real-World Agent-Based Transport Optimization

Klaus Dorer …………………………………………….. 56

Contents XI

Scheduling of Automated Double Rail-Mounted Gantry

Cranes

Ren´e Eisenberg ………………………………………….. 57

Solving Real-World Vehicle Scheduling and Routing Problems

Jens Gottlieb ……………………………………………. 59

Exact and Heuristic Solution of the Global Supply Chain

Problem with Transfer Pricing and Transportation Cost

Allocation

Pierre Hansen, S´ebastien Le Digabel, Nenad Mladenovi´c, Sylvain Perron 60

Planning Problems for Combined Pick-up Point Allocation,

Transportation, and Production Processes with Time-Varying

Processing Capacities

Christoph Hempsch……………………………………….. 62

Paradigm Shift in the Supply Chain – Is it Really Happening?

Britta Kesper, Yuriy Kapys ………………………………… 64

Support of Bid-Price Generation for International Large-Scale

Plant Projects

Dirk Mattfeld, Jiayi Yang ………………………………….. 65

Bid Querying Policies in Combinatorial Auctions for

Collaborative Transportation Planning

Giselher Pankratz ………………………………………… 68

Application of HotFrame on Tabu Search for the Multiple

Freight Consolidation Problem

Filip Rychnavsk´y …………………………………………. 71

Simulation Metamodeling of a Perishable Supply Chain

M.E. Seliaman, Ab Rahman Ahmad ………………………….. 74

Non-Cooperative Games in Liner Shipping Strategic Alliances

Xiaoning Shi, Stefan Voß ………………………………….. 75

Container Terminal Operation and Operations Research

Dirk Steenken, Stefan Voß, Robert Stahlbock ……………………. 77

Mixed Integer Models for Optimized Production Planning

Under Uncertainty

David L. Woodruﬀ ……………………………………….. 78

XII Contents

Part III Contributions Not Presented

Simulation Optimization of the Cross Dock Door Assignment

Problem

Uwe Aickelin, Adrian Adewunmi …………………………….. 81

Heuristics for the Multi-Layer Design of MPLS/SDH/WDM

Networks

Holger H¨oller, Stefan Voß ………………………………….. 84

Part I

Contributions VEAM Workshop

Events in Modeling of Complex Systems

Janusz Granat

National Institute of Telecommunications, Szachowa 1, 04-894 Warsaw and

Institute of Control and Computation Engineering, Warsaw University of

Technology, 00-665 Warsaw, Poland, [email protected]

Management and modeling of complex systems is a challenging area of re-

search. There are various approaches for modeling these systems. One of the

approaches is event driven modeling and management of complex systems.

The concept of application of events to systems modeling is not a new one.

It has been applied for modeling of the discrete systems, stochastic systems

etc. However, most of the existing modeling approaches use only information

about the type of event and the time when an event occurs. The information

systems store much richer information about events. This information might

be structured as well as unstructured. The structured information is stored

in databases in the form of tables. The unstructured information is stored in

various forms of textual information. It can be considered to use more infor-

mation about the events what advances events driven modeling approaches.

Figure 1 shows the basic components of the event driven modeling frame-

work: the system that is inﬂuenced by external as well as internal events, data

and textual information about the system as well as about the events, models,

algorithms, event detection algorithms, knowledge representation, description

of decision maker behavior and actions.

Internal

events

External events

SYSTEM

Actions

Data & text

Models

Event

detection

Algorithms

Decision

maker

Knowledge

representation

Fig. 1. Basic components of the modeling framework

4 Janusz Granat

In order to build models or algorithms we have to store the data about

the system and the events. The existence and the proper quality of data

are crucial to any further steps. We can distinguish primary data that are

stored in relational databases and preprocessed data that are prepared for

speciﬁc modeling tasks. The data can be stored in one central database or

can be stored in distributed databases. Moreover, the designers apply the

event based system design approach which leads to well structured databases

that contain information about events. There is also increased importance

of using textual information about events. Recently, the video sequences are

becoming important source of data for event discovery.

The models use mathematical formulas to describe the behavior of the

system. In case of the presented framework the models describe dependencies

between events and observable variables. Various models can be considered

like stochastic models, temporal relationships, temporal sequence associations

etc. The algorithms in Figure 1 are understood as algorithms that work with

analytical models as well as algorithms for event mining or event processing.

A key to understanding events is knowledge of what might have caused them

and having that knowledge at the time the events happen. Event mining is

one of the key approaches. Event mining can be deﬁned as a process of ﬁnding:

the frequent events, the rare events, unknown events (its occurrence can be

deduced from observation of the system), the correlation between events, the

consequences of the event and what caused the event. There is a special class

of algorithms for event detection. We distinguish two classes of algorithms,

event detection based on numerical and categorical data analysis and event

detection by analysis of the textual information. The results of algorithms,

data and textual information go to the block called Knowledge representa-

tion. In this block there is unifying representation of the results. However, the

results are a very simple form of the knowledge. Here, there is a place for in-

troducing contextual knowledge and more advanced algorithms that support

knowledge creation and management. Also the knowledge about the conse-

quences of events will be represented. The ability to track event causality

and consequences is an essential step toward on-line decision support and an

important challenge for new algorithms for event mining. The models and al-

gorithms as well as data provide the decision maker with important knowledge

about the system. Then the decision maker can specify various actions that

will be applied in the system and reduce the inﬂuence of events on the system.

The information about actions should be stored in computerized form. That

will help later in the evaluation of consequences of the chosen actions. In some

cases the results of the algorithms can be directly applied to the system (for

example the event based control algorithms).

Recently, the focus is on real-time decision support what requires a new

class of the data processing, the analytical algorithms as well as modeling ap-

proaches. The actions have to be taken immediately after the event occurred.

The delay may cause the fault of the system or signiﬁcant losses. It should be

stressed that we can distinguish a broad spectrum of various types of events. It

Events in Modeling of Complex Systems 5

will often require dedicated algorithms and approaches. However, the frame-

work will help in the generalization of the speciﬁed methods and algorithms.

Moreover, this framework may help in the integration of achievements in event

based modeling in diﬀerent scientiﬁc disciplines. At this time there are sepa-

rate developments in temporal data mining, stochastic systems, event based

control etc. The combination of these approaches might signiﬁcantly improve

the results of new algorithms.

The presented approach has various applications in business monitoring,

network management, intrusion detection, fault detection etc. In this section

we will present selected examples of event driven modeling: events monitoring,

event processing networks, events in environmental scanning, event based con-

trol, temporal sequence associations for rare events, event mining and events

in alerting systems. There is research on events monitoring in given environ-

ment. Sensor networks are applied for events monitoring. Sensor networks are

systems of many sensing elements endowed with computation, communication

and motion that can work together to provide information about events in an

environment. In this case we have information about the type of event, the

time and location of events. The control algorithms are used for positioning

mobile sensors in response to a series of events. Many monitoring problems

can also be stated as the problem of detecting a change in the parameters of

a system called event detection. Another important concept are Event Pro-

cessing Networks (EPN). Such networks consist of Event Processing Agents

called event sources, event processors and event viewers. EPN have been ap-

plied for computer network monitoring. The events sources were middleware

sniﬀers. The aggregated information about events has been displayed by view-

ers and additionally has been used for event mining. This concept has also

been applied for solving business problems. The organizations are working on

improvement of the analysis of the external environment and inﬂuence of this

environment on the performance of the organization. Environmental scan-

ning is a new term and it means the acquisition and use of the information

about events, trends, and relationships in an external environment. In this

case the methods of dealing with unstructured information about events are

especially important. In event based control the sampling is event-triggered

instead time-triggered. The event-based PID controller can be built. Such an

approach reduces CPU utilization. The event-triggered PID controller is a

nonlinear system of hybrid nature. In many cases we have to monitor and

analyze rare events like credit card frauds, network faults etc. However, if we

store the data about the system in the database it is very diﬃcult to identify

rare events. In this case the events are characterized by the type of event and

the time of occurrence of the event. Temporal sequence associations for rare

events can be applied to solve this problem.

There are new opportunities that come from the large amount of data

that is stored in various databases. Event mining becomes a challenging area

of research. In this subsection we will focus on formulating the event mining

tasks that consider observations of the system as well as internal and external

6 Janusz Granat

Events

t

Alarm

?

Observations

Fig. 2. Events and observations

events. Figure 2 shows interrelations between events, observation of the system

that is given in form of time series and alarms. Sometimes, it is impossible

to observe the events directly. In such cases the data are stored in databases

in form of time series. This data represents observations of the system in

selected points. The observations are analyzed by the system and alarms are

generated in case of abrupt changes in the values of observations. In the next

step another algorithm ﬁnds the events that caused the changes in the system.

The following algorithms can be considered:

•For signiﬁcant change of observation ﬁnd events that are the reasons of

this change

•Prediction of future events by analyzing the changes of observations

•Prediction of changes of observations after the event occurs

There are various applications of event based modeling. These approaches

use various methodologies. The presented modeling framework might help in

developing future event driven modeling. We have stressed the new direction

of research called event mining.

SimConT: A Tool for Quick Layout and

Equipment Portfolio Evaluation and Simulation

of Hinterland Container Terminal Operations

Manfred Gronalt, Thouraya Benna, and Martin Posset

University of Natural Resources and Applied Sciences Vienna, Department of

Economic and Social Science, Institute of Production and Logistics,

Feistmantelstraße 4, A-1180 Vienna, Austria,

[manfred.gronalt,thouraya.benna,martin.posset]@boku.ac.at

Hinterland container terminals (HCTs) are important hubs in modern logistic-

networks that ensure eﬃcient and frictionless intermodal (rail, truck, ship)

container turnover which has to be planned and coordinated. The increasing

number of HCTs along the Danube and within vital industrial regions shows

their signiﬁcant and leading role in the development of European hinterland

container traﬃc. In contrast to open sea container terminals, hinterland con-

tainer terminals face other challenging optimization issues. Open sea container

terminals typically handle mainly two types of containers (20 feet and 40 feet).

Diﬀerent container types can be stored in separated storage blocks which can

furthermore be separated into import and export blocks. Although HCTs are

usually constrained in their storage capacity, they are faced with a bigger

container diversity, nearly no predictable delivery and pickup time windows

and smaller turnover. Consequently containers have to be stored within mixed

yard blocks. In addition, most of the operation activities are triggered by rail-

way processes.

Keeping these characteristics in mind, eﬃciently planning of extensions

and rebuilding of HCTs has to be done very carefully. Therefore, dynamic

analyses of the maximum storing positions as well as modelling of the in-

bound and outbound ﬂows are necessary in order to determine the resulting

capacity requirements and the infrastructure needed (railways and road in-

frastructure). Numerous restrictions must be met and kept in mind during

the planning of new and extended inland terminals to avoid costly problems

in daily terminal operation. Hence a comprehensive analysis of equipment uti-

lization and detailed terminal infrastructure planning is becoming necessary

to ensure an eﬃcient HCT operation. The simulation of functions of container

terminals is an approach for eﬃcient resource-planning and eﬀective capacity

analysis of HCTs, which is based on modern simulation techniques. By means

of simulation diﬀerent material handling technologies, shift patterns, resource

8 Manfred Gronalt, Thouraya Benna, and Martin Posset

scheduling and infrastructure capacity are analysed. Further, optimization is

used in order to ﬁnd optimal conﬁguration parameters.

The goal of our research is to minimize the risk of bad investments and

stranded costs when planning and (re)building the infrastructure and capacity

of HCT. The SimConT Simulation environment is based on a modular concept

which supports a potential user with on-time available and reliable results

and reports for eﬃcient planning of capacity and infrastructure for inbound

and outbound ﬂows of a hinterland container terminal. The integration of

inbound and outbound ﬂows, which enables the evaluation of infrastructure

requirements for train, trucks and vessels is a further essential facet of our

research. SimConT was developed in a modular design including terminal-

conﬁguration, simulation and report generator. The modules of the simulation

environment have to be passed through in a sequential way and end up in a

clear and comprehensive reporting.

System

Configuration

Order

Configuration

Data

Generator

Simulation Report

Generator

Fig. 1. Sequential components of SimConT

The terminal conﬁguration enables the creation of an artiﬁcial image of

a potential terminal layout by determining all parameters and consists of

three modules, including system conﬁguration, order conﬁguration and data

generator. All interfaces for parameter entry are designed in a self-explanatory

way to avoid extensive training measures and additional consulting. System

and order conﬁguration modules include the conﬁguration of all necessary

layout, operation and order related data and oﬀer intelligent input data error

avoiding interfaces. Hereafter the deﬁned parameters are transmitted to the

data generator, where lists of inbound and outbound containers (which will

be used as input data for the simulation) are produced and edited.

The subsequent simulation works with exact number and identiﬁcation of

containers and records all container movements, equipment allocations and

storage positions exactly. The scheduling of the terminal equipment is done

according to shortest path, predeﬁned priority rules and the availability of con-

tainer related information. The goal of the simulation is to provide guidelines

for improving HCTs layout and equipment conﬁguration with consideration

of transportation lead time reduction and number of container lifting. Finally,

the simulation results are processed by the report generator to oﬀer a clear

and comprehensive overview of the main performance indicators of HCTs. All

SimConT: A Tool for Quick Layout and Equipment Portfolio Evaluation 9

results can be accessed within a cockpit in an intuitive way with extensions

of graphical support and logical aggregations.

Our work is done in close cooperation with an Austrian HCT operat-

ing company and an Austrian rail infrastructure operator. This ensures the

integration of practice based data and know-how. Typically, HCT provide ca-

pacity in a range of of 600 to 1.500 TEU with a medium turnover of 100.000

to 150.000 TEU per year. They serve two to six tracks and the operation is

predominantly done by transtainers and reach stackers.

The results are a preliminary work for creating a prototype of a dedicated

HCT simulation-environment. Within our research distinctive functions of

HCTs are investigated and set in a modular relation, which can be calibrated

and scaled, to create the ability of simulating HCTs in many possible shapings

and sizes.

Keywords: Intermodal terminals, hinterland terminals, container terminal

management, simulation, optimization

Dynamic Multi-Linked Negotiations in

Multi-Echelon Production Scheduling

Networks

Hoong Chuin Lau1, Guan Li Soh2, and Wee Chong Wan2

1School of Information Systems Singapore Management University, Singapore,

2The Logistics Institute – Asia Paciﬁc National University of Singapore,

Singapore, [tlisgl, tliwwc]@nus.edu.sg

Introduction

In a multi-agent system, agents negotiate to ﬁnd a mutually acceptable so-

lution to a problem. More often than not, agent negotiations are performed

independently of one another. In this paper, we apply the concept of multi-

linked negotiations [1] (where negotiations exert inﬂuences on one another)

in a dynamic system that needs to generate solutions quickly to satisfy de-

mand across a multi-echelon production scheduling network. Multi-linked ne-

gotiations occur in situations where a task requires further sub-tasks to be

completed, and also when the existence of many such tasks results in compe-

tition for a common resource. An example of this is in a manufacturing supply

chain network that consists of entity nodes and linkages deﬁning contractor-

contractee relationships. Very often such relationships do not extend beyond

a direct relationship. Suppliers upstream usually have no information about

their customer’s customer, and the converse is true. Agent implementations

usually simulate a single tier commodity market without such multi-tier rela-

tionships. Applying single-tier agent negotiation strategies to multi-tier sys-

tems brings up the questions of negotiation ordering and the parameters of

the negotiations.

We consider a dynamic (as opposed to anticipatory) multi-echelon pro-

duction scheduling network involving the production, assembly and trans-

portation of multi-indenture goods that arrive dynamically. A ﬁnished good

undergoes component production and diﬀerent levels of assembly at diﬀerent

facilities. Diﬀerent facilities have diﬀerent capabilities and capacities in pro-

viding the various operations required. A request is composed of a number

of goods at a number of locations by a certain time. It is handled by a man-

agement (contractor) agent who generates a schedule that optimizes the total

production cost. The facilities and transportation services are represented by

Dynamic Multi-Linked Negotiations 11

contractee agents who negotiate for the available jobs according to their lo-

cal utilities. Contractor and contractee agents have no visibility over each

other’s agenda. The problem is to generate a schedule that maximizes the

contractee agents’ local utilities while minimizing the total production cost.

Requests are processed one after another as they enter the system, and no

decommitment is allowed. Due to the directly and indirectly linked relation-

ships [1] brought about by a multi-echelon network and multi-indenture goods,

approaches based on single independent negotiations will not work well. To

overcome this shortcoming, we apply the concept of negotiation ordering and

the feature assignment as described in [1]. Our work diﬀers from [1] in the

problem solved; [1] outputs a negotiation ordering and corresponding feature

assignments for a subsequent negotiation phase. In our work, we assume that

the actual time of negotiations is negligible, hence enabling us to embed the

negotiations within the scheduling algorithm. This integrated approach allows

us to do away with the uncertainties of negotiation outcomes and generate a

ﬁnal production schedule eﬃciently.

Problem Formulation and Modeling

We deﬁne our problem in a military context. We view a request as a mission

order (MO) and a production schedule is needed to fulﬁll the MO. A request

comprises a list of ﬁnished goods required at speciﬁed locations by a stipulated

time. This list of ﬁnished goods requires diﬀerent levels of assembling from

components at diﬀerent locations, and also their transportation between these

locations. An MO can hence be deﬁned as a hierarchy of mjobs and sub-jobs,

both assembly and transportation, that need to be completed to fulﬁll the

request. A production schedule that fulﬁlls the MO is deﬁned by (a) the

facility where each job is executed; (b) the start and ﬁnish times of the jobs;

(c) transport assignment at pickup and drop locations. For simplicity in this

paper, we will ignore production and transportation, and focus on assembly.

Our model comprises two diﬀerent types of agents: the management agent

(ma) and nfacility agents (fa). The ma receives the MO and seeks a schedule

that minimizes the sum of production costs Pifor all jobs i=1 to m. Each

fa represents a facility that maintains its own local schedule of jobs assigned.

Each fa jwishes to maximize its local utility function Ujthat models internal

preferences (that may be in conﬂict with the ma’s objective). In our problem,

Ujmodels the facility’s foreknowledge and experience in handling future as-

signments (bearing in mind that mission orders arrive dynamically but are

fulﬁlled one after another). The goal in this paper is to ﬁnd a schedule χthat

minimizes the objective function Z=

m

P

i=1

Pi(χ)−

n

P

j=1

Uj(χ). We propose a

negotiation scheme where a series of negotiations will proceed sequentially in

a given order (to be explained below), each involving the ma with a partic-

ular fa on a particular sub-job, and the outcome of an early negotiation will

12 Hoong Chuin Lau, Guan Li Soh, and Wee Chong Wan

become a constraint for later negotiations. All fas are assumed to be truth-

telling and non-collusive. The challenge is to ﬁnd a negotiation ordering such

that negotiations proceeding along that ordering will produce a schedule that

minimizes Z.

Three-Phase Solution Approach

Our algorithm is a nested search comprising of three sub-phases: (a) ﬁnd

a facility assignment, (b) ﬁnd a negotiation ordering, and (c) generate the

best schedule that maximizes Z. The initial facility assignment is generated

greedily based on capacity, utilization, and distance while the initial negotia-

tion ordering is generated in a bottom-up lexicographical manner of the job

hierarchy.

Facility Assignment ϕ

This phase is concerned with allocating a fa to each job. Here we apply a

heuristic local search algorithm to generate diﬀerent facility assignments.

Negotiation Ordering φ

Given a ﬁxed facility assignment ϕ, this phase ﬁnds the best negotiation order-

ing. Using simulated annealing, successive orderings are generated. For each

φ, a project scheduling algorithm that considers only the total production cost

is then used to determine an optimal schedule χPwhich consists of the start

time and deadline for each job in the MO. Each (ϕ,φ,χP) triplet is the input

for the next phase.

Agent Negotiation

For each triplet (ϕ,φ,χP), negotiations between ma and the respective fa’s in

ϕproceeds in the order deﬁned in φ. The timings in χPis used as a proposed

timing from ma to the fa. We propose a reward scheme where units of credits

ﬂow among fa’s through the ma. Initially, the ma possesses all credits. Based

on its utility function Uj,fa jgenerates an internal local schedule χC(i.e., its

own set of timings for the assigned job i). The local schedule χCand current

credit standing will be used to negotiate with the ma as follows:

i. fa accepts χPunconditionally, if Uj(χP)≥Uj(χC)

ii. fa counter-proposes χCand gives up Uj(χC)−Uj(χP) credit units, if

Uj(χP)< Uj(χC) and fa has suﬃcient credits

iii. fa accepts χPwith an increase of Uj(χC)−Uj(χP) units credited, if

Uj(χP)< Uj(χC) and fa has insuﬃcient credits

Dynamic Multi-Linked Negotiations 13

In case of ii, the ma will give a rough commitment if through accommodat-

ing χC, it is able to generate a feasible schedule, followed by a full commitment

if the total production cost is increased by an amount no more than the net

number of credits received from all fas at the end of all negotiations. In case

of iii, the ma accedes if it has enough credits. Otherwise, the negotiation

is unsuccessful. The result of a successful negotiation will in turn become a

constraint for the subsequent negotiations in the ordering. A new schedule

is formed when all negotiations are successful. The ﬁtness of this schedule χ

is measured with the objective function Z. Our algorithm seeks to ﬁnd the

triplet (ϕ,φ,χ) that minimizes Z.

Experimental Results

This will be provided in the full-length version of the paper.

Acknowledgment

This work is supported by the Singapore Ministry of Defense.

References

1. Zhang, X., Lesser, V., Abdallah, S.: Eﬃcient Management of Multi-Linked Ne-

gotiation Based on a Formalized Model. Autonomous Agents and Multi-Agent

Systems 10(2), 165–205 (2005).

Modelling Classiﬁcation Analysis for

Competitive Events with Applications to

Sports Betting

Stefan Lessmann1, Johnnie Johnson2, and Ming-Chien Sung2

1University of Hamburg, Institute of Information Systems, Von-Melle-Park 5,

20146 Hamburg, Germany, [email protected]

2Centre for Risk Research, School of Management, University of Southampton,

UK

Classiﬁcation analysis involves interfering a functional relationship between

independent variables and a discrete target variable from a set of example

patterns. Subsequently, the captured relationship facilitates predicting the

value of target variables when only the values of the independent variables are

known. While multivariate statistical methods like logistic regression or dis-

criminant analysis are well established several empirical benchmarks give rise

to the suspicion that novel machine learning techniques like artiﬁcial neural

networks or support vector machines are capable of providing more accurate

predictions. Hence, such techniques are predominantly used in contemporary

application of classiﬁcation analysis; e.g., the support of managerial decision

making, medical diagnosis, speech and image recognition or text mining.

Competitive events diﬀer from ordinary classiﬁcation analysis in the sense

that certain patterns compete against each other for a speciﬁc target value in

a certain context. That is, the functional relationship the classiﬁer or learn-

ing machine has to derive from the set of example patterns does not only

depend on the independent variables of one pattern but as well on those of

some interlinked patterns. We refer to this linking as contextual information.

Consider the case of horseracing as an example. A large number of indepen-

dent variables can be used to build a prediction model facilitating winner

versus non-winner classiﬁcation, e.g., measurements of the horses and jockeys

past performance. However, the likelihood of a particular horse winning a race

does not only depend on its individual skill and past performance but also on

those of its competitors in a given race. In fact, this information seems crucial

and omitting it can be expected to be highly detrimental for predictive per-

formance. As a result, the literature on modelling competitive events focuses

on statistical methods that are based on maximum likelihood estimation and

capable of taking this contextual information into account.

Modelling Classiﬁcation Analysis for Competitive Events 15

We strive to adapt machine learning methods to competitive settings by

ﬁnding an appropriate representation of the data. The idea to account for

competition purely by data modelling is appealing since it requires no algo-

rithmic modiﬁcation of the classiﬁer therewith facilitating the application of

several standard learning techniques for comparison purpose. Therefore, we

develop three diﬀerent modelling techniques of diﬀerence to best coding, pair-

wise matching and race to example modelling and compare them with a stan-

dard classiﬁcation setting. Preliminary results are derived for a horseracing

data set using the support vector machine classiﬁer. The horseracing domain

is selected due to the large body of literature within this ﬁeld and the fact that

the varying number of competitors within a race imposes additional constrains

on the applicability of traditional classiﬁcation analysis. The selection of the

support vector machine is motivated by its excellent empirical performance

in several benchmark studies and its solid mathematical underpinnings.

Keywords: Classiﬁcation, competitive events, horseracing, support vector

machines

Structured Modeling Technology: Recent

Developments and Open Challenges

Marek Makowski

International Institute for Applied Systems Analysis, A-2361 Laxenburg, Austria,

[email protected], http://www.iiasa.ac.at/~marek

Mathematical modeling of a complex problem is actually a network of ac-

tivities involving interdisciplinary teams collaborating closely with experts in

modeling methods and tools; often however new methods and/or software

need to be developed, especially in the case of:

•Models with a complex structure using large amounts of diversiﬁed data,

possibly from diﬀerent sources.

•The need for robust strategies to account for a proper treatment of spatial

and temporal distributional aspects, vulnerabilities, inherent uncertainty

and endogenous risks aﬀecting large communities and territories.

•Demand for integrated model analysis, which should combine diﬀerent

methods of model analysis for supporting a comprehensive examination of

the underlying problem and its alternative solutions.

•Stronger requirements for the whole modeling process, including quality

assurance, replicability of results of diversiﬁed analyses, and automatic

documentation of modeling activities.

•Requirement of a controlled access through the Internet to modeling re-

sources (composed of model speciﬁcations, data, documented results of

model analysis, and modeling tools).

•Demand for large computing resources (e.g., large number of computa-

tional tasks, or large-scale optimization problems, or large amounts of

data).

Use of established modeling methods and general-purpose modeling tools

cannot adequately meet requirements of such modeling activities. Thus we

need to advance modeling methodology to address these requirements.

Geoﬀrion presented in [2] a detailed speciﬁcation of a modeling cycle. Here,

we discuss the modeling cycle composed of more aggregated elements which

correspond to the elements of the Structured Modeling Technology (SMT)

outlined below:

Structured Modeling Technology 17

•Analysis of the problem, including the role of a model in the corresponding

decision-making process; and the development of the corresponding model

speciﬁcation.

•Collection and veriﬁcation of the data to be used for the calculation of the

model parameters.

•Deﬁnition of various model instances (composed of a model speciﬁcation,

and a selection of data deﬁning its parameters).

•Diversiﬁed analyses of the instances.

•Eﬃcient use of computational grids for large volume of computations.

•Documentation of the whole modeling process.

SMT is a Web-based application supporting the whole modeling process.

Users do not use any modeling language; model speciﬁcation is done through

several simple forms composed of choice lists and text ﬁelds (only for spec-

iﬁcation of relations basic knowledge of LaTeX is required). All persistent

elements of the whole modeling process are stored in an automatically gener-

ated data warehouse. This approach supports modeling work by teams located

in distant locations.

While the basic functionality of SMT has been developed, and is used for

large and complex models (having more than a million variables and rather

complex indexing structure) there are several open challenging problems, solu-

tions of which are needed for extending functionality of SMT. These problems

include:

•an eﬃcient implementation of handling of complex measurement units (a

key attribute of each SMT entity),

•methods for eﬀective numerical experiments (analysis of a large number of

solutions in order to automatically generate new, possibly also large, sets

of computations).

The current state of the SMT development, and of on-going research activ-

ities related to the open problems will be discussed. More information about

SMT can be found in [4], and at URL http://www.iiasa.ac.at/∼marek.

Acknowledgment

Several ideas exploited in the SMT have resulted from many discussions

and joint activities of the author with A. Beulens, A. Geoﬀrion, J. Granat,

H. Scholten, H-J. Sebastian and A.P. Wierzbicki. The user and DBMS in-

terfaces of SMT has been designed and implemented by colleagues from

the National Institute of Telecommunications, Warsaw, Poland: M. Majdan,

C. Chudzian, B. Kozlowski.

18 Marek Makowski

References

1. Geoﬀrion, A.: An introduction to structured modeling. Management Science

33(5), 547–588 (1987).

2. Geoﬀrion, A.: Integrated modeling systems. Computer Science in Economics

and Management 2, 3–15 (1989).

3. Geoﬀrion, A.: Indexing in modeling languages for mathematical programming.

Management Science 38(3), 325–344 (1992).

4. Makowski, M.: Structured modeling technology, European Journal of Opera-

tional Research 166(3), 615–648 (2005).

5. Makowski, M., Wierzbicki, A.: Modeling knowledge: Model-based decision sup-

port and soft computations. In: Yu, X., Kacprzyk, J. (eds.) Applied Decision

Support with Soft Computing, Vol. 124 of Series: Studies in Fuzziness and

Soft Computing, Springer, Berlin, 3–60 (2003). Draft version available from

http://www.iiasa.ac.at/∼marek/pubs/prepub.html.

6. Wierzbicki, A., Makowski, M., Wessels, J. (eds.): Model-Based Decision Support

Methodology with Environmental Applications. Series: Mathematical Modeling

and Applications, Kluwer, Dordrecht (2000).

Production Planning with Load Dependent

Lead Times and Deterioration

Julia Pahl1, Stefan Voß1, and David L. Woodruﬀ2

1University of Hamburg, Institute of Information Systems, Von-Melle-Park 5,

20146 Hamburg, Germany, [email protected],

2Graduate School of Management, UC Davis, Davis CA 95616, USA

Summary. As organizations move from creating plans for individual production

lines to entire supply chains it is increasingly important to recognize that decisions

concerning utilization of production resources impact the lead times that will be

experienced. In this paper we give some insights into why this is the case by looking

at queuing that results in delays. We use these insights to brieﬂy survey and sug-

gest optimization models that take into account load dependent lead times. Related

“complications”consider the relationship and inﬂuence between deterioration or per-

ishable items and load dependent lead times in the framework of tactical production

planning.

Keywords: Supply chain management, lead times, tactical planning, deteri-

orating items, perishability, rework

The increased globalization forces companies to compete on an expanding set

of criteria. One key criterion is the lead time which is deﬁned as the time

between the release of an order to the shop ﬂoor or to a supplier and the

receipt of the items. Lead time considerations are essential with respect to

the global competitiveness of supply chains, because long lead times impose

high costs due to rising work in process (WIP), inventory levels as well as

larger safety stocks caused by increased uncertainty about demand. Shorter

lead times permit the increase of eﬃciency by, e.g., enabling companies to

quote faster deliveries to customers and reducing the uncertainty of demand

forecasting. However, in intermittent production systems manufacturing lead

times tend to be long and variable with only a fraction of time being due to

value added processing times and the rest of the time being a result of wait-

ing in the system. Large planning models typically treat lead times as static

input data, but in most situations, the output of a planning model implies

capacity utilizations which, in turn, imply lead times. Despite this, consider-

20 Julia Pahl, Stefan Voß, and David L. Woodruﬀ

ations about lead times dependent upon resource utilization (load dependent

lead times: LDLT) are rare in the literature. The same is valid for models

linking order releases, planning and capacity decisions to lead times, and tak-

ing into account factors inﬂuencing lead times such as the system workload,

batching, sequencing decisions, WIP levels or rework due to deterioration or

perishability of products. Furthermore, in Supply Chain Management (SCM)

and production planning models nonlinear dependencies, e.g., between lead

times and the workload of a production system or a production resource, are

usually omitted. This happens even though there is empirical evidence that

lead times increase nonlinearly long before resource utilization reaches 100%,

which may lead to signiﬁcant diﬀerences in planned and realized lead times.

Certain characteristics of production materials, e.g., deterioration or per-

ishability can necessitate rework of production materials: when passed a spe-

ciﬁc maximum lifetime items have to be replaced or rerouted thus consum-

ing capacity and augmenting utilization. The perishability or deterioration of

goods is regarded as the process of decay, damage or spoilage of items in such

a way that they cannot be used for their original purpose anymore, viz. they

go through a change in storage and loose their utility partially or completely.

This could be a continuous process so that such items have a stochastic life-

time in contrast to perishable goods which are considered as items with a

ﬁxed, maximum lifetime. The latter is true for products which become ob-

solete at some ﬁxed point in time, because of various reasons, e.g., change

in style or technological developments. The higher the grade of deterioration,

the more rework time (if possible) and rework costs are necessary in order to

recover the item and to bring it back to a good quality state [1].

Rework can be economically attractive if rework times are much smaller

than the initial production times and if the value of the reworkable item is

substantial due to, e.g., expensive input materials. Additionally, producers can

be obliged by legislation and disposal bans to rework their defective products.

Other producers may want to take environmental responsibility and conse-

quently rework their defective, perished and/or returned products. According

to this, the integration of deterioration and perishability into models with

LDLT is very interesting especially in regard of lead time behavior. There-

fore, in this paper we extend our previous work on load dependent lead times

(see [2, 3, 4]) by considering them in the context of “complications” investigat-

ing the relationship and inﬂuence between deteriorating or perishable items

and those lead times in the framework of tactical production planning.

References

1. Inderfurth, K., Lindner, G., Rachaniotis, N.P.: Lot Sizing in a Production Sys-

tem with Rework and Production Deterioration. International Journal of Pro-

duction Research,43, 1355–1374 (2005).

2. Voß, S., Woodruﬀ, D.L.: A Model for Multi-Stage Production Planning with

Load Dependent Lead Times. In: Sprague, R.H. (ed.) Proceedings of the 37th

Production Planning with LDLT and Deterioration 21

Annual Hawaii International Conference on System Science, IEEE Piscataway,

DTVEA03 1–9 (2004).

3. Pahl, J., Voß, S., Woodruﬀ, D.L.: Load dependent lead times – From empir-

ical evidence to mathematical modeling. In: Kotzab, H., Seuring, S., M¨uller,

M., Reiner, R. (eds.) Research Methodologies in Supply Chain Management,

Physica, Heidelberg, 539–554 (2005).

4. Pahl, J., Voß, S., Woodruﬀ, D.L.: Production Planning with Load Dependent

Lead Times. 4OR: A Quarterly Journal of Operations Research,3, 257–302

(2005).

Guided Online Decision Making

J¨orn Sch¨onberger and Herbert Kopfer

Chair of Logistics, Faculty of Business Studies and Economics

Wilhelm-Herbst-Straße 5, 28359 Bremen, Germany, [sberger,

Kopfer]@logistik.uni-bremen.de

The derivation and the formulation of formal decision models in terms of

mathematical expressions are necessary prerequisites for the successful ap-

plication of automatic algorithms to support decision making to manage the

ﬂows of goods and resources through a logistical network. Numerous generic

and special models, often of optimization type, have been developed and inves-

tigated in this context as well as powerful algorithms to solve them. During

the last ﬁve decades, the quality as well as the quantity and scope of the

models have been successively extended by considering more and more rele-

vant problem parameters and possible decisions. However, in the last decade, a

new stream of interest has been established: the consideration of dynamic and

non-predictable future data. Within this contribution, we will apply state-of-

the-art methods as well as new ideas to model such a decision problem arising

from the reactive routing of a ﬂeet of service teams which response quickly

after technical failures or machine breakdowns have been reported.

We consider the deployment problem of the dispatching unit of a ﬂeet

of service teams visiting customer sites after these customers have reported

technical failures that require a solving by a technician who is able to rem-

edy the malfunctions or breakdowns. Each technician is equipped with a van

that carries all tools necessary to solve technical failures immediately at the

corresponding customer site. As soon as additional customers call for tech-

nical support, the associated customer site visits have to be integrated into

the existing schedule in order to meet service time windows at the customer

sites agreed between the dispatching unit and the customers (online decision

making). As long as the number of additionally incoming requests stays below

the maximal possible workload of the service ﬂeet then nearly all requests can

be served without signiﬁcant loss of punctuality and at reasonable costs. In

case that the number of additionally received demands for customer site visits

exceeds the maximal possible work-load of the service ﬂeet (a demand peak)

then the quality of service decreases: the percentage of in-time customer site

visits declines and the amount of penalty payments for late arrivals increases.

Guided Online Decision Making 23

Selected requests are allowed to be subcontracted to other service partners

that are more expensive but ensure an in-time task completion.

Initially, we investigate the application of a pure cost-based myopic deci-

sion strategy for routing the service teams. Every time a new request arises,

the so far followed schedule is replaced by a new one. This new schedule is

the solution of a scheduling problem in which the costs for the required new

schedule are minimized. This means only the costs for fulﬁlling the current

set of open and uncompleted customer site visits are considered. Since the

penalty amount payable for an arrival after the closure of a customer site

time window is relatively low compared to the costs for the incorporation of

an external service partner, out-of-time-window visits will be preferred espe-

cially in situations in which a spontaneous and unpredictable demand peak

occurs. Within some numerical simulations we show that the punctuality de-

clines dramatically and for a too long time if such a demand peak arises.

The fallen punctuality represents a loss of service quality and has nega-

tive impacts for the survivability of the service company. In order to ensure

a continuous and high service performance, the management of the service

company has deﬁned and published a policy which speciﬁes that at least 90%

of the customer sites must be visited within the agreed time windows on av-

erage. Such a policy can be understood as a guideline to which the repeated

schedule derivation has to adapt in the sense that the generated schedules

must fulﬁl the properties speciﬁed in the policy and if the properties are not

fulﬁlled then the re-achievement of the property fulﬁlment must be re-achieved

immediately.

In the pure myopic dispatching strategy mentioned above the policy re-

quirements are not fulﬁlled at all. To remedy this problem, we propose to

decompose the dynamic decision problem into two separate but interacting

decision problems. One of these two problems is dedicated to ensure a high

punctuality even in situations with very high system load and is handled as

the superior problem. Solving the superior problems means to modify the

scheduling rules with respect to the current system punctuality. The resulting

scheduling rules are then used to solve the second, inferior, decision problem

in which the minimization of schedule costs is requested.

The adaptation of the scheduling rules is equivalent to the modiﬁcation of

the corresponding short-term scheduling problem and therefore results in the

variation of a mathematical optimization problem of the current dispatching

instance consisting of an objective function, a set of constraints and a col-

lection of domains for the considered decision variables. The scheduling rule

adaptation can be realized by re-deﬁning the objective function, one or more

constraints or by shrinking and relaxing the decision variable domains. We

report about our experiments and results with the modiﬁcation of the search

direction and the decision variable domains.

We have investigated the adaptation of the search direction by modifying

the objective function of the mathematical schedule optimization problem. It

is intended to bias the search so that it becomes more promising to externalize

24 J¨orn Sch¨onberger and Herbert Kopfer

requests if the punctuality is low and to prevent expensive externalization if

the punctuality is already high. We have introduced one coeﬃcient for the

part of externalization costs and one coeﬃcient for the part of travel costs in

the objective function. By adjusting the coeﬃcients, we can shift the relative

weight from externalization to self-entry or vice versa and can therefore guide

the search in a particular direction. We report about numerical experiments

with this kind of scheduling rule adaptation. The observed results show that

the adaptation is possible but hardly controllable.

In order to keep as much control as possible about the adaptation of the

optimization model, we have decided to guide our attention to the adaptation

of the domains of the decision variables. In a preliminary step, we reformulate

the short-term decision model and introduce a binary decision variable for each

request that indicates whether this request has to be externalized or not. The

adaptation of the optimization model works as follows: In case that the average

punctuality of the most recently completed and scheduled requests declines

below the value indicated in the policy, the binary indicator decision variables

for all additionally released requests are set to “1” so that all additionally

released requests are externalized. As soon as the punctuality re-increases,

the binary indicator variables for the new requests are set to ‘0’ so that all

recently received requests are allowed to be sourced out as well as to be served

by an own team. Within several numerical experiments we prove the general

applicability and controllability of this approach to adapt the myopic decision

model.

We terminate our contribution with the report of some numerical experi-

ments in which the costs for the application of the scheduling rule adaptation

are estimated.

Scalability of Three Parallel Direct Search

Methods in Simulation-Based Optimization

Frank Thilo and Manfred Grauer

Information Systems Institute, University of Siegen, H¨olderlinstr. 3, D-57068

Siegen, Germany, [email protected],[email protected]

Introduction

Optimization tasks in multidisciplinary optimization include design problems

in manufacturing in the aircraft or automotive industry, alloy casting processes

and metal-sheet forming. Typically, solving these problems involves running

many complex simulations which are implemented in commercial software

packages. In general, the optimization algorithm has to treat the simulation

system as a black box. To solve such a simulation-based optimization prob-

lem, many hundreds or thousands of simulations are necessary, each of which

is computationally expensive. This results in extremely high computational

demands which can only be met by distributing the computation over many

CPUs.

Basically, there are two diﬀerent approaches how parallel or distributed

computing can be used in this scenario: First, the time needed for a single

simulation run can be reduced by parallelizing the simulation software itself,

e.g., by partitioning a FEM mesh and assigning each partition to a diﬀerent

CPU. Second, the optimization algorithm can request the evaluation of mul-

tiple scenarios at the same time, so that many (sequential) simulations are

executed simultaneously. It is also possible to combine the two approaches.

The simulation and optimization of complex products in the area of virtual

prototyping aims at reducing the time to market for new innovative products

and creates an ever-increasing demand for computational power which can

only be met by utilizing larger numbers of CPUs. This requires scalable hard-

ware architectures and algorithms. This paper focuses on algorithmic scala-

bility; in particular, three scalable optimization algorithms are presented. To

allow a meaningful scalability analysis, a suitable performance and eﬃciency

metric for heuristic, parallel optimization is developed. This is then used to

analyze the characteristics of the algorithms for solving a benchmark problem

as well as real-life industrial problems from metal sheet forming [1] and cast-

ing processes [9, 4] on a compute cluster of up to 300 CPUs. Furthermore, the

26 Frank Thilo and Manfred Grauer

scalability of parallel versions of the simulation packages, which are used for

the optimization problems, is examined for diﬀerent networking technologies.

The Scalability Concept

The term scalability is used in diﬀerent contexts to express that a computer

system or an algorithm is able to solve a given problem faster or can cope with

an increased workload when resources are added. Here, we are concerned with

adding additional nodes to a compute cluster or a computational grid. Scal-

ability analysis can be divided into algorithmic and architectural scalability.

While the ﬁrst focuses on attributes of an algorithm, i.e., the algorithm’s se-

quential portion, its inherit concurrency limits and synchronization costs, the

latter examines hardware related aspects as processing capacity, information

capacity and connectivity. To predict real runtimes of a parallel algorithm

on a given hardware architecture, both kinds of analyses must be taken into

account.

The main metric to quantify scalability behavior is speedup, which com-

pares the computation times of a parallel algorithm for diﬀerent numbers of

CPUs [7]. In the case of heuristic, parallel optimization algorithms, there is no

sharply deﬁned goal for which each algorithm’s elapsed time can be compared.

Instead, both the time needed and the quality of the solution must be con-

sidered, e.g., by examining the progress of the achieved solution quality over

time. However, a target solution quality can be deﬁned and the time needed

to reach this level be used as the basis to calculate speedup and eﬃciency

values.

Distributed Simulation-Based Optimization

Optimization problems which arise in the ﬁeld of computational engineering

usually cannot be formulated analytically because of their complexity. Instead,

a model of the real problem is created, typically in the form of a data set for

a simulation software package. For optimization, the model is parameterized

by some variables which can be chosen within lower and upper bounds. The

goal is to ﬁnd the best set of ndecision variables as deﬁned by some objective

function which has to be minimized. Furthermore, the set of possible solutions

can be limited by arbitrary constraints.

During the course of the optimization, thousands of solution candidates

must be evaluated, requiring a costly simulation run each time. While a sin-

gle simulation requires several minutes up to several hours or even days of

processing time, the amount of computation in the optimization algorithm

itself is several orders of magnitude lower. Also, the amount of data that is

exchanged between the search method and the simulations is very low. Thus,

the algorithms’ scalability is mainly limited by their inherent concurrency

limits, synchronisation points and drop in eﬀectiveness when increasing the

degrees of parallelism and not by communication costs. Experiments show

Scalability of Three Parallel Direct Search Methods 27

that for this type of simulation-based optimization problems the relation of

local computation time to communication time is at least 1000 to 1.

The simulation-based nature of the problem means that in general no

derivative information is available and no assumptions can be made about

the nature of the problem space. This prohibits the use of linear program-

ming techniques or gradient-based optimization algorithms. One class of al-

gorithms which can be applied are so-called direct search methods [6]. There

is no exact deﬁnition of direct search, but important characteristics are that

these methods do not explicitly use derivative information nor build a model

of the objective function. Instead, the basic operation relies on direct com-

parison of objective function values. To utilize parallel computing resources,

parallel direct search methods are needed, which can evaluate several solu-

tion candidates simultaneously. Three such search methods are compared:

The Distributed Polytope Search (DPS), a parallel implementation (PSS) of

the meta-heuristic scatter search [8], and asynchronous parallel pattern search

(APPS) [3].

DPS belongs to the class of simplex-based search methods. It generates

new solution candidates by applying geometrical operations to a set of pre-

viously calculated solutions. The initial set of 2nfeasible solutions is gener-

ated by a parallel random search. During the main exploration phase, new

points are generated by reﬂecting or contracting existing solutions relative to

the weighted center of gravity. The number of operations in each iteration

depends on parameters which can be adjusted to the number of CPUs. In-

feasible points (i.e., points which violate the constraints) are modiﬁed by a

binary search repair strategy. The algorithm terminates when the standard

deviation of the objective values drops below a threshold.

PSS can be viewed as an evolutionary approach. At the heart of the al-

gorithm is the reference set which is initialized by a diversiﬁcation algorithm.

The size of the set is ﬁxed and between 10 and 20. In each iteration, all pairs

and 3-tuples containing a new solution are combined to create new candidate

solutions. The combination is a linear combination with a diﬀerent random

factor for each decision variable. This typically results in several hundred new

points which are then evaluated simultaneously. The new reference set is built

by choosing both the best and most diverse solutions while moving infeasi-

ble solutions towards a known solution in a path relinking step. When the

standard deviation of the objective function values reaches a threshold, new

random solutions are created to increase diversity. After a ﬁxed number of

these steps, the algorithm terminates.

The third algorithm, APPS, belongs to the group of pattern search op-

timization algorithms. By default, APPS uses two search directions for each

decision variable. Thus, 2nnew points are created and evaluated in parallel.

Infeasible points are discarded. If there is a best new point, it is chosen as

the new starting point for the next iteration. Otherwise the step length is de-

creased. The algorithm terminates when this length drops below a threshold.

APPS does not work iteratively, but asynchronously, i.e., it does not wait for

28 Frank Thilo and Manfred Grauer

all points to be evaluated before it creates new candidate solutions, but can

continue as soon as one evaluation has ﬁnished and has resulted in a new best

solution. In this respect, it diﬀers from the synchronous approaches of both

DPS and PSS.

Computational Results

In the following, ﬁrst results from solving multidisciplinary optimization prob-

lems are presented. The computations have been performed on the Rubens

SLES Linux cluster of the University of Siegen on up to 300 CPUs.

To allow an extensive scalability analysis over a wide range of problem di-

mensions and numbers of CPUs, a mathematical test problem is deﬁned based

on the well-known Rosenbrock function. To mimic the temporal behaviour of

a real, simulation-based problem, an event-based simulation is used which

keeps track of virtual wall clock time where each evaluation of the function

is assigned a given amount of virtual time. To validate the results, a subset

is compared with those of a real distributed optimization on the compute

cluster.

Figures 1 and 2 show the relative speedup of DPS and PSS for diﬀerent

problem dimensions and numbers of CPUs. Figures 3 and 4 depict the algo-

rithms’ average solution quality over time for 10 and 100 CPUs, respectively.

The results indicate that each algorithm exhibits diﬀerent scalability charac-

teristics. Scatter search has the highest eﬃciency for large numbers of CPUs,

in particular for problems with few decision variables. However, it takes more

absolute time to reach a given solution quality than the other two algorithms

when using only a small number of CPUs.

The algorithms have been used to solve several real-world optimization

problems. They have been integrated into the OpTiX optimization environ-

ment which presents an abstract interface of the problem to the algorithms

and handles the distribution and scheduling of the distributed simulation runs.

It is currently being transformed into a service-oriented architecture [2]. The

past problems include a design problem of an aircraft wing and several design

and hybrid control problems in groundwater and pollution management. We

plan to present scalability results for two new problem domains and compare

them to the ﬁndings of the Rosenbrock test problem: The ﬁrst optimization

problem is a multi-stage problem in metal-sheet forming where an initial steel

blank repeatedly undergoes a deep drawing process with diﬀerent tools and

forces in each stage. Possible decision variables are the geometrical parameters

(e.g., diameters, radii) of the tools or the blank, or other process parameters,

e.g., blank holder forces. In addition to these continuous variables, the algo-

rithm must also ﬁnd the optimal number of stages. Thus, the task can be

classiﬁed as a mixed-integer, non-linear optimization problem. The INDEED

[5] software package is used to simulate the forming process and the CATIA

CAD software is interfaced for geometry generation.

Scalability of Three Parallel Direct Search Methods 29

Figure 1: Relative speedup of DPS for solving 10-

to 100-dimensional Rosenbrock problems on 1 to

200 CPUs (average of 200 runs)

Figure 2: Relative speedup of PSS for solving 10-

to 100-dimensional Rosenbrock problems on 1 to

200 CPUs (average of 200 runs)

Figure 3: Comparison of the solution quality over

virtual time; 10-dimensional Rosenbrock problem,

solved using 10 CPUs (average of 200 runs)

Figure 4: Comparison of the solution quality over

virtual time; 10-dimensional Rosenbrock problem,

solved using 100 CPUs (average of 200 runs)

The second problem is one in the domain of alloy casting processes. A hot,

liquid alloy is cast into a molding shell which deﬁnes the ﬁnal shape of the

desired object. This is a multi-scale problem which covers the complex motion

of the liquid alloy as it ﬂows into the shell and its thermodynamic behavior on

the large scale and eﬀects of the solidiﬁcation process of the material’s evolving

microstructure on the very small scale. Possible goals of the optimization is to

enhance the quality of solidiﬁed material and to decrease the time needed for

the whole process. Decision variables include temperatures, the ﬁlling speed

and geometrical parameters. The software packages CASTS [4] and MICRESS

are used to simulate the casting process and the microstructure attributes.

For both the casting and the metal-sheet forming problem, it is possible

to utilize parallelism for a single simulation. For the parallel INDEED vari-

ant FETI-INDEED as well as for CASTS, their scalability behavior is being

analyzed when utilizing diﬀerent networking technologies from Fast Ethernet

to Myrinet. Furthermore, the combination of parallelization on the simulation

and optimization level is being examined. Based on the scalability analysis of

each, an optimal allocation of resources can be determined by estimating the

combined speedup for a given number of CPUs and allocation strategy.

30 Frank Thilo and Manfred Grauer

References

1. Grauer, M., Barth, T.: About Distributed Simulation-based Optimization of

Forming Procesess Using a Grid Architecture. In: Ghosh, S., Castro, J.M.,

Lee, J.K. (eds.) Materials Processing and Design: Modeling, Simulation and

Applications (NUMIFORM), Springer, 2097–2102 (2004).

2. Foster, I.: Service-Oriented Science. Science 308, 814–817 (2005).

3. Hough, P., Kolda, T.G., Torczon, V.: Asynchronous Parallel Pattern Search for

Nonlinear Optimization. SIAM Journal of Scientiﬁc Computing 23(1), 134–156

(2001).

4. Jakumeit, J., Barth, T., Grauer, M., Reichwald, J.: Grid Computing for Casting

Simulations. In: Proc. Modeling of Casting, Welding and Advanced Solidiﬁca-

tion Processes XI MCWASP, to be published (2006).

5. Kessler, L., Weiher, J., Roux, F.-X., Diemer, J.: Forming simulation of high-

and ultrahigh-strength steel using INDEED with the FETI method on a work-

station cluster. In: Mori, K.-I. (ed.) Simulation of Materials Processing, Proc.

of NUMIFORM 2001. Balkema Publishers, 399–404 (2001).

6. Kolda, T.G., Lewis, M.R., Torczon, V.: Optimization by Direct Search: New

Perspectives on Some Classical and Modern Methods. SIAM Review 45(3),

385–482 (2003).

7. Kumar, V., Grama, A., Gupta, A., Karpysis, G.: Introduction to Parallel Com-

puting. Benjamin Cummings (1994).

8. Marti, R., Laguna, M., Glover, F.: Principles of Scatter Search. European Jour-

nal of Operational Research 169(2), 359–372 (2006).

9. Stefanescu, D.M.: Computer simulation of shrinkage related defects in metal

castings – a review. International Cast Metals Research 18, 129–143 (2005).

Nanatsudaki Model of Knowledge Creation

Processes

Andrzej P. Wierzbicki12 and Yoshiteru Nakamori1

1Center for Strategic Development of Science and Technology, Japan Advanced

Institute of Science and Technology, Asahidai 1-1, Nomi, Ishikawa 923-1292,

Japan

2National Institute of Telecommunications, Szachowa 11, 04-894 Warsaw, Poland,

In the book Creative Space [1], we have shown that there are many spirals of

knowledge creation, some of them of organizational character, typical for mar-

ket innovations and practice-oriented organizations, some of normal academic

character, typical for research organizations.

The normal academic research combines actually three spirals: hermeneu-

tics (gathering scientiﬁc information and knowledge from literature, web

and other sources and reﬂecting on these materials), called by us the EAIR

(Enlightenment-Analysis-Immersion-Reﬂection)Spiral;debate (discussing in

a group research under way), called by us the EDIS (Enlightenment-Debate-

Immersion-Selection)Spiral;experiment (testing ideas and hypotheses by

experimental research), called by us the EEIS (Enlightenment-Experiment–

Interpretation-Selection)Spiral. Since all of these spirals begin with having

an idea, called the Enlightenment (illumination,aha,eureka) eﬀect, they can

be combined into a Triple Helix of normal knowledge creation, typical for

academic work.

These three spirals contained in the Triple Helix do not exhaustively de-

scribe all what happens in academic knowledge creation, but they describe

most essential elements of academic research: gathering and interpreting in-

formation and knowledge, debating and experimenting. However, these spirals

are individually oriented, even if a university and a laboratory should support

them; e.g., the motivation for and the actual research on preparing a doctoral

thesis is mostly individual. Moreover, the Triple Helix only describes what

researchers actually do, it is thus a descriptive model. Obviously, the model

helps in a better understanding of some intuitive transitions in these spirals

and makes possible testing, which parts of these spirals are well supported in

academic practice and which require more support; but it does not give clear

conclusions how to organize research.

However, there are also several other creative spirals described and ana-

lyzed in the book Creative Space. One is the ARME Spiral of revolution-

32 Andrzej P. Wierzbicki and Yoshiteru Nakamori

ary knowledge creation; however, revolutionary knowledge creation occurs

rarely and in unexpected places. But three others are important for prac-

tical knowledge creation, for innovations, particularly in industry and other

purpose-oriented organizations. These are the organizational creative spirals,

motivated by purposes of a group and aimed at using the creative power

of the group, while an individual plays here the role of a member of the

group, not of an individual researcher. One of them is the widely known SECI

(Socialization-Externalization-Combination-Internalization)Spiral; another,

actually older but formulated as a spiral only recently, is the brainstorming

DCCV (Divergence-Convergence-Crystallization-Veriﬁcation)Spiral; still an-

other, the Occidental counterpart of the SECI Spiral (which is of Oriental ori-

gin), is the objective setting OPEC (Objectives-Process-Expansion-Closure)

Spiral.

Each of these spirals has a diﬀerent role and can be applied for diﬀerent

purposes, but all have their strengths. Unfortunately, they cannot be easily

combined into a multiple helix like the Triple Helix, because they do not share

the same elements. However, the main challenge is not only to combine these

spirals between themselves, but also with the spirals of academic knowledge

creation. This general challenge is diﬃcult, but such a combination would be

important for several reasons:

•Combining these spirals might strengthen academic knowledge creation,

because it would increase in it the role of the group supporting the indi-

vidual research;

•Combining these spirals might strengthen also industrial innovation and

knowledge creation, because it always contains also some individual ele-

ments that should be explicitly accounted for;

•Combining these spirals might help in the cooperation of industry with

academic institutions in producing innovations, because it could bridge

the gap between the diﬀerent ways of conducting research in academia

and in industry.

With these purposes, we present in this paper the JAIST Nanatsudaki

Model – an exemplar (serving as an example to follow, a normative model) of

a process of knowledge and technology creation. It consists of seven creative

spirals; and each of these spirals might be as beautiful and unpredictable in its

creativity, as water whirls in the seven waterfalls (nanatsudaki) on Asahidai

close to JAIST. The seven spirals include the three academic and the three or-

ganizational mentioned above, but are supplemented by a planning roadmap-

ping spiral based on the I-System (the pentagram of Nakamori). The model

is build following the assumption that its applications will concern technol-

ogy or material science development, thus the application phase consists of

experimental work.

Although the model could start with any constitutive spiral, we assume

that it starts with objective setting (thus uses part or entire of the OPEC

Nanatsudaki Model of Knowledge Creation Processes 33

Spiral) and ends with the applications, experimental work, here represented

by the EEIS Spiral.

There can be two interpretations of the JAIST Nanatsudaki Model. One

is that each constitutive spiral of this septagram should be completed, i.e.,

at least one cycle of the spiral should be realized. This is, however, a rather

constraining interpretation, since creative spirals should start and end at any

of their elements, without a prescribed number of cycles. Thus, we describe

the model while using a diﬀerent interpretation: we might use any number

of the elements (transitions) of the spirals, as necessary, sometimes without

completing even one cycle, sometimes repeating more than one cycle.

Beside the detailed description of the model, the paper presents its in-

tended applications and comments on the comparison of importance of all its

constitutive spirals, based on a survey of opinions conducted at JAIST.

Fig. 1. Diagram of JAIST Nanatsudaki Model

References

1. Wierzbicki, A., Nakamori, Y.: Creative Space, Springer, Berlin (2006).

The Use of Reference Proﬁles and Multiple

Criteria Evaluation in Knowledge Acquisition

from Large Databases

Andrzej P. Wierzbicki12 , Jing Tian1, and Hongtao Ren1

1School of Knowledge Science, Japan Advanced Institute of Science and

Technology (JAIST), Asahidai 1-1, Nomi, Ishikawa 923-1292 Japan

2National Institute of Telecommunications, Szachowa 11, 04-894 Warsaw, Poland,

When analyzing complex data sets in a large database, the problem of knowl-

edge acquisition can be posed as ﬁnding such data sets that either correspond

best to the expectations of a user (client, decision maker, etc.) or, contrari-

wise, correspond worst to such expectations. We suggest in this paper that

such expectations should be described by a set of criteria and by a reference

proﬁle of the desired values of such criteria. The reference point method can

be applied then to ﬁnd the data sets that correspond either best or worst to

the expectations.

The paper describes such an approach ﬁrst in an abstract way and sug-

gests that this approach can be applied to diverse aims, such as ﬁnding critical

aspects of a complex logistics and supply chain management problem. How-

ever, the actual motivation of this approach was the interpretation of data

of a survey of conditions and problems of scientiﬁc creativity at JAIST. This

original application is presented in more detail in the paper.

The purpose of the survey was to ﬁnd what aspects of knowledge creation

processes are evaluated by graduate students (preparing for a master or doc-

toral degree) as either most critical or most important. A long questionnaire

was prepared and answered by over 120 students; the questions were of three

types. The ﬁrst type was assessing importance of a given subject; the most

important questions might be considered as those that correspond best to

a reference proﬁle. The other type was assessing the situation between stu-

dents and at the university; the most critical questions might be selected as

those that correspond worst to a reference proﬁle. The third type was testing

the answers to the ﬁrst two types by indirect questioning revealing student

attitudes.

It was found that most critical questions of the second type (worst nega-

tively evaluated by students) are related to not good enough situations con-

cerning:

The Use of Reference Proﬁles and Multiple Criteria Evaluation 35

1. Critical feedback, questions and suggestions in group discussions;

2. Organizing and planning research activities;

3. Preparing presentations for seminars and conferences;

4. Designing and planning experiments;

5. Generating new ideas and research concepts.

These are actually elements of four spirals of knowledge creation: Inter-

subjective EDIS (Enlightenment-Debate-Immersion-Selection) Spiral – items

1) and 3); Experimental EEIS (Enlightenment-Experiment-Interpretation-

Selection) Spiral – item 4); Hermeneutic EAIR (Enlightenment-Analysis-

Immersion-Reﬂection) Spiral – item 5); and Roadmapping (I-System) Spiral

of planning knowledge creation processes – item 2). The importance of these

spirals is also stressed by the positive evaluation of the importance of other

elements of these spirals in response to questions of the ﬁrst type:

1. Learning and training how to do experiments;

2. Help and guidance from the supervisor and colleagues;

3. Frequent communication of the group.

The analysis also has shown that language barriers are considered most

critical for good research, which is an expected result, but also indicated some

unexpected results, such as that research competition and personal shyness

do not essentially prevent an exchange of ideas.

A general conclusion is that the use of a multiple criteria formulation and

reference proﬁles for knowledge acquisition from complex data sets gives very

promising results and should be applied more broadly.

Convex Envelope for Medical Modeling

Fadi Yaacoub, Yskandar Hamam, and Charbel Fares

ESIEE, Lab. A2SI, Cit´e Descartes, BP 99, 93162 Noisy-Le-Grand, France,

[f.yaacoub, y.hamam, c.fares]@esiee.fr

The environment simulation is widely used nowadays. Training in many ﬁelds

such as medicine and architecture heavily depends on virtual reality tech-

niques. Since objects in real life do not have a deterministic shape it is not

possible to have a geometric equation that might model them. Convex en-

velopes are a must in simulation. The need of such envelopes rises with the

intention of having realistic scenes with collision detection between objects.

In this paper three methods for generating the convex envelope are compared.

Then a combination of those methods is shown in order to reduce the time of

execution yielding into a hybrid method for convex envelope generation.

The convex hull or convex envelope of a ﬁnite set Sof npoints in the Eu-

clidean space <dof dimension ddenoted as CH (S) is deﬁned by the smallest

convex set containing all the points or simply the intersection of all half-spaces

containing the set S. The convex hull in <dis the set of solutions to a ﬁnite

system of linear inequalities in d-variables:

CH (S) = x∈ <d:Ax ≤b(1)

where A∈ <n∗dand b∈ <n.

In this paper three methods for computing the convex hull are shown:

Brute Force [1], Gift Wrapping (Jarvis March) [2] and QuickHull [3]. Finally

a hybrid technique that combines those algorithms is shown.

The Brute Force algorithm begins by taking a random point piand con-

siders three other diﬀerent points as a facet (pj,pk,pl). It checks if the point

piis counterclockwise with respect to this facet and continues with another

facet (pj,pk,pl+1) and so on by checking all the facets made by all points

other then pi. If piis counterclockwise with respect to all facets, it is on the

convex hull. Otherwise, if the point is clockwise with one of the facets, it is

not on the convex hull.

The Gift Wrapping algorithm acts as follows: ﬁrst it ﬁnds a starting edge

(a,b) by using the 2D algorithm on the projection of the points on the XY

plane; it pivots a plane around the edge of the hull; it ﬁnds the smallest angle

of a plane picontaining the starting edge (a,b) and a point pi; it replaces piby

Convex Envelope for Medical Modeling 37

cand forms a triangular face containing (a,b,c). All points now lie to the left

of triangle (a,b,c). Finally the algorithm repeats the same process recursively

for the edges (a,c) and (b,c) by ﬁnding other triangles adjacent to those edges.

The QuickHull ﬁnds the convex envelope by recursively partitioning the

given set of points. It begins by dividing the set of points into two subsets

with respect to a plane formed by three points: the vertices corresponding

to the minimum (xmin) and maximum (xmax) abscise, and the vertex cor-

responding the maximum distance from the line joining (xmin,xmax ). From

this initial plane, QuickHull creates a cone of new facets (called visible facets)

by calculating the point that has the maximum distance with respect to the

plane. Therefore, QuickHull builds new sets of points from the outside set of

points of the visible facets. If a point is above multiple new facets, one of the

new facets is selected. If it is below all the new facets, the point is inside the

convex hull and can consequently be discarded. Partitioning also records the

furthest points of each outside set. Table 1 shows the calculation time for the

algorithms shown above when applied on three diﬀerent wrist bones.

Since the running time of each algorithm depends on the number of points

in the object, our objective is to reduce the number of iterations to speed up

the algorithm. Therefore, a hybrid algorithm based on QuickHull and Gift

Wrapping is proposed. It consists ﬁrst of using the QuickHull and then ap-

plies the Gift Wrap on each subset of points obtained. In the full paper the

algorithms will be discussed in detail and numerical simulations will be given.

3D Model Original Model Convex Hull Brute Force Gift Wrap QuickHull

#vertices/#facets #vertices/#facets time(sec) time(sec) time(sec)

3rdMeta- 675 1272 150 296 2330.7 0.26 0.21

carpal

Hamat 2812 5620 394 784 19231.2 0.89 0.62

Ulna 977 1864 312 620 12153 0.41 0.37

Table 1. Execution time for computing the 3D convex hull between Brute Force,

Gift Wrap and QuickHull

Keywords: Convex envelope, medical modeling, computational geometry,

bounding volumes, collision detection, virtual reality

References

1. Breg, M., Schwarzkopf, O., Kreveld, M., Overmars, M.: Computational Geome-

try: Algorithms and Applications, 2nd ed., Springer (2000).

2. O’Rourke, J.: Computational Geometry in C, Cambridge University Press, New

York (1994).

3. Barber, C., Dobkin, D., Huhdanpaa, H.: The QuickHull Algorithm for Convex

Hulls. ACM Transactions on Mathematical Software 22(4), 469–483 (1996).

Applying Data Mining for Early Warning and

Proactive Control in Food Supply Networks

Li Yuan, Mark R. Kramer, and Adrie J.M. Beulens

Information Technology Group, Wageningen University Dreijenplein 2, 6703 HB,

Wageningen, The Netherlands, [Yuan.Li, Mark.Kramer, Adrie.Beulens]@wur.nl

European consumers are highly conscious of food quality and safety. This

concern has been strengthened by a series of food safety crises in the recent

past such as Bovine Spongiform Encephalopathy (BSE), dioxin contamina-

tion, Foot and Mouth Disease (FMD), Nitrofen, etc. [1]. Recall announce-

ments can be found in newspapers almost weekly as a reaction to deﬁciencies

in food products. In response further, food supply networks implement sys-

tems to improve quality of food products and to guarantee food safety. In order

to prevent problems in food quality and improve eﬃciency and eﬀectiveness of

operations, early warning and proactive control systems are required in food

supply networks.

Early Warning and Proactive Control

Early warning systems are well known in natural sciences. These systems,

based on historical monitoring, local observation or computer modelling, pre-

dict and help to prevent or reduce the impact of natural disasters. They are

typically used to monitor potential disasters relating to meteorology, geology

(e.g., earthquakes and volcanoes) [3] or technology (e.g., nuclear safety). Early

warning is being extended to other application areas as well. For example, [2]

presented a prototype sensor system for the early detection of microbially

linked spoilage in stored wheat grain. The early warning system we intend

to build should not only predict potential food quality problems, but also

help identify relations between determinant factors and quality attributes of

food products. Ultimately, the knowledge about these relations and the de-

cision varieties associated with these factors will enable proactive control to

prevent those problems. A proactive control system can adjust corresponding

determinant factors to prevent quality problems.

Applying Data Mining in Food Supply Networks 39

Data Mining

The application of early warning and proactive control requires predictive

models of the object system (i.e., the food supply network being controlled).

However, to construct such a model, we would normally require detailed in-

sight into processes involved. Processes that determine quality of food prod-

ucts are not completely understood and there are many unknown interactions

between quality attributes. However, in current food supply networks, large

amounts of data about business operations and transactions are recorded ev-

ery day. So an alternative approach is to use data mining to infer a model

from available data. Data mining (DM) is the process of extracting valid,

previously unknown, comprehensible and actionable information from large

databases and using it to make crucial business decisions [4]. Application

of data mining in food supply networks is cheap and ﬂexible when domain

knowledge is scarce [5].

Framework for Early Warning System in Food Supply Networks

Based on the objectives of early warning and proactive control we designed a

framework for early warning system in food supply networks. An important

component of this framework is the knowledge base. This knowledge base

contains information needed to implement early warning and proactive control

in food supply networks. For each type of process that we intend to control,

the following information is stored: variables involved, control limits for all

variables, data availability and the time required to gather data, decision

variety and inﬂuence of decisions on subsequent processes.

Fig. 1. Framework for early warning system in food supply networks

The knowledge base also serves to extend the early warning and proactive

control systems. As new relations are discovered, these relations and variables

40 Li Yuan, Mark R. Kramer, and Adrie J.M. Beulens

involved will be recorded in the knowledge base. So it is necessary to accu-

mulate the knowledge we discovered along the way and organize it with a

systematic, ontological approach. Our knowledge base will contain

•types of determinant factors,

•types of deviations,

•types of relations between determinant factors and performance,

•suitable data mining techniques for discovering these types of relations,

•and instantiations of those types of relations in food supply networks.

Managers can easily beneﬁt from this knowledge base by either directly ap-

plying similar relations to their cases or by employing suggested data min-

ing techniques to solve their problems. The template approaches will provide

managers with a guidebook to help classify their problems into appropriate

types and to select proper data mining techniques for relation discovery and

prediction.

Case Study

Our case study in a chicken supply network has already shown the advantage

of using data mining in food supply networks. In order to identify relations

between Death On Arrival (DOA) and its determinant factors, we used three

diﬀerent types of data mining techniques (decision tree, neural networks, and

nearest-neighbours methods) in the large volume of data stored at various

stages of the supply network. Results of this research have already been con-

ﬁrmed by domain experts in food supply networks.

Concluding Remarks

The combination of serious eﬀects of food safety problems, the abundance of

recorded data and potential beneﬁts of preventing food quality problems in

food supply networks provided the motivation for this research eﬀort. Our next

step in this ongoing research project is to construct a knowledge base with the

associated ontology using the knowledge we obtained in case studies. Further,

we also build an early warning system with the knowledge base and apply it

to novel cases in order to verify the usability and validity for predicting food

quality problems.

References

1. Beulens, A.J.M.: Transparency Requirements in Supply Chains and Networks:

Yet another challenge for the Business and ICT Community. Herausforderungen

der Wirtschaftsinformatik in der Informationsgesellschaft. Wissenschaftsverlag

Edition am Gutenbergplatz, Leipzig (2003).

Applying Data Mining in Food Supply Networks 41

2. De Lacy Costello, B.P.J., Ewen, R.J., Gunson, H., Ratcliﬀe, N.M., Sivanand,

P.S., Spencer-Phillips, P.T.N.: A prototype sensor system for the early detection

of microbially linked spoilage in stored wheat grain. Measurement Science &

Technology 14(4), 397–409 (2003).

3. Grijsen, J.G.S., Snoeker, X.C., Vermeulen, C.J.M.: An information system for

ﬂood early warning. Pres. at the 3rd International Conference on Floods and

Flood Management, Florence, Italy, 24–26 November 1992 (1993).

4. Simoudis, E.: Reality Check for Data Mining. IEEE EXPERT 11(5), 26–33

(1996).

5. Verdenius, F., Hunter, L.: The power and pitfalls of inductive modelling. In:

Tijskens, L.M.M., Hertog, M.L.A.T.M., Nicolai, B.M. (eds.) Food Process Mod-

elling. Woodhead Publishing Limited, 105–136 (2000).

Part II

Contributions Logistics and SCM Workshop

Optimizing Inventory Decisions in a

Multi-Stage Supply Chain Under Stochastic

Demands

Ab Rahman Ahmad1and M. E. Seliaman2

1UTM, Johor, Malaysia, [email protected]

2King Fahd University of Petroleum and Minerals, Dhahran 31261, KSA,

Supply chain management can be deﬁned as a set of approaches utilized to

eﬃciently integrate suppliers, manufacturers, warehouses, and stores, so that

merchandise is produced and distributed at the right quantities, to the right

locations, and at the right time, in order to minimize system-wide cost while

satisfying service level requirements. Recently numerous articles in supply

chain modeling have been written in response to the global competition. How-

ever, most supply chain inventory models deal with two-stage supply chains.

Even when multi-stage supply chains are considered, most of the developed

models are based on restrictive assumptions. Therefore, there is a need to

analyze models that relax the usual assumptions to allow for a more realistic

analysis of the supply chain.

In this paper we consider the case of a three-stage supply where a ﬁrm can

supply many customers. This supply chain system involves suppliers, manu-

facturers, and retailers. Production and inventory decisions are made at the

suppliers and manufacturers levels. The production rates for the suppliers

and manufacturers are assumed ﬁnite. In addition the demand for each ﬁrm

is assumed to be stochastic. The problem is to coordinate production and

inventory decisions across the supply chain so that the total cost of the sys-

tem is minimized. For this purpose, we develop a model to deal with diﬀerent

inventory coordination mechanisms between the chain members. Numerical

examples will be presented and simulation experiments will be used to vali-

date the model.

Impact of E-Commerce on an Integrated

Distribution Network

Daniela Ambrosino and Anna Sciomachen

DIEM – Universit´a di Genova Via Vivaldi, 5, 16126 Genova, Italy, [ambrosin,

sciomach]@economia.unige.it

E-Commerce (EC) provides new channels for the distribution of goods; it

represents an opportunity to improve the ﬂows in the supply chain, and con-

sequently, to reduce the inventory level in the whole network [4]. But on the

other hand, EC imposes to use fast and accurate information systems and

communication; in these new systems collaboration and coordination become

a critical issue and integration represents the only way to survive [5].

Motivated by the above considerations, in this work we will devote our

attention to the integration in the inventory management in a multi-echelon,

multi-channel distribution system; in particular, we will analyse the manage-

ment of inventories of ﬁnal goods in a distribution system where products are

available in diﬀerent supply channels: a traditional channel in which the prod-

ucts are distributed through depots, a direct channel and an Internet-enabled

direct channel.

The distribution network we are involved with is made up of three levels:

central depots (CD), peripheral depots (D) and customers (clients C, big

clients BC and e shopping clients e C). The channels for supplying goods

are the following: a traditional channel where peripheral depots (supplied by

central ones) serve clients (C); a direct channel for serving big clients (BC),

i.e., clients characterized by a large demand are served directly by CD; the

Internet-enabled channel where depots at the top echelon (CD) serve e clients

(e C). The CD represent the supply points of the network and play a dual

role: they supply peripheral depots and serve customers. Customers served by

CD are big-clients and e-clients of the direct and Internet-enabled channel,

respectively. The assignment of peripheral depots and both big clients and

e clients to central depots is known and also the assignment of clients C to

the peripheral depots. In order to better understand the ﬂows of goods in

the network under investigation, we report in Figure 1 a simple distribution

system with two central and three peripheral depots, four big clients and a

set of customers and e-clients.

Impact of E-Commerce on an Integrated Distribution Network 47

Fig. 1. A simple case of the distribution network under investigation

We assume that the demand of customers is known and expressed in terms

of units of a single representative commodity. Inventories can be stocked both

at CD and D.

Note that in the Internet-enabled channel the manufacturer receives orders

directly from e-clients (via Internet) and ships the product directly to them.

This supply chain strategy has the same structure as a single echelon tradi-

tional supply chain; however, the problems arising in this new channel are

completely new and in part connected with the increasing customer service

expectations [3].

Given the network described above and a time horizon Tsplit into periods,

the problem we deal with is to determine the optimal inventory level for central

and peripheral depots within each time period Tin order to minimise ordering,

inventory, stock out, e commerce and transportation costs, whilst satisfying

capacity and requirements constraints and granting a certain customer service

level. The inventory policy is based on a periodic review policy (e.g., every

day) in which goods are ordered when inventories are under a given level

called ordering point; the quantity to order is deﬁned for restoring inventories

while minimising the logistic costs; it depends on both the existing stock in

the whole system and the inventory strategy.

We present a three phase algorithm: the pre-processing phase is devoted

to the deﬁnition of the order point for each stock point of the network (i.e.,

CD and D), the ﬁrst phase determines the optimal inventory policy by solving

a Mixed Integer Linear Programming model (MIP); ﬁnally, a second phase,

denoted “integration” phase, deﬁnes the “current stock situation” in the whole

48 Daniela Ambrosino and Anna Sciomachen

network and identiﬁes accordingly the best transferring policy for managing

the ﬂow of goods in the network and improving the customer service level.

Note that a particular stock situation in a part of the network can require

to modify the optimal inventory allocation in the whole distribution system;

the aim of this integration phase is to avoid to have unbalanced inventory

levels in the diﬀerent echelons when the network or a part of it is suﬀering

or is risking to suﬀer a shortage (a stock out). For doing this we introduce

some controls on local and global stock out and deﬁne the best transferring

strategy for granting an inventory balance in accordance with the rationing

strategy [2] and the base stock policy modiﬁcation [1].

The proposed three phase solution approach is used for evaluating new

distribution strategies for an Italian food company. We simulate diﬀerent sce-

narios by assuming diﬀerent initial stock situations, diﬀerent customers’ de-

mands and diﬀerent percentage of demand devoted to e-commerce (20, 40, 60

and 80% of the demand of C).

We evaluate the impact of EC on both the distribution costs and the in-

ventory levels in the network. Preliminary results show that if a part of clients

C, usually served by peripheral depots, choose the Internet channel, the inven-

tory level at the peripheral echelon of the network gets lower. Remembering

the dual rule of central depots, we can note that EC has a positive eﬀect

also on inventories taken at the central depots for supplying the depots at

the lower level of the network. In our case the Internet-enabled direct channel

enables the company to obtain an average cost reduction of 15%.1

References

1. Chen, F.: Optimal policies for multi-echelon inventory problems with batch

ordering. Operations Research 48(3), 376–389 (2000).

2. Diks, E.B., De Kok, A.G.: Optimal control of a divergent multi-echelon inven-

tory system. European Journal of Operational Research 111, 75–97 (1998).

3. Disney, S.M., Naim, M.M., Potter, A.: Assessing the impact of e-business on

supply chain dynamics. International Journal of Production Economics 89,

109–118 (2004).

4. Gunasekaran, A., Marri, H.B., McGaughey, R.E., Nebhwani, M.D.: E-commerce

and its impact on operations management. International Journal of Production

Economics 75, 185–197 (2002).

5. Manthou, V., Vlachopoulou, M., Folinas, D.: Virtual e-Chain (VeC) model for

supply chain collaboration. International Journal of Production Economics 87,

241–250 (2004).

1Partially supported by MIUR PRI N 2005012452-003 project.

An Interval Pivoting Heuristic for Finding

Quality Solutions to Uniform-Bound

Interval-Flow Transportation Problem

Aruna Apte1and Richard S. Barr2

1Graduate School of Business and Public Policy, Naval Postgraduate School,

Monterey CA 93943, USA, [email protected]

2Department of Engineering Management, Information, and Systems, Southern

Methodist University, Dallas, TX 75275, USA, [email protected]

We present interval-ﬂow networks, network ﬂow models in which the ﬂow on

an arc may be required to be either zero or within a speciﬁed range. The ad-

dition of such conditional lower bounds creates a mixed-integer program that

captures such well-known restrictions as time windows and minimum load

sizes. This paper describes the mathematical properties of interval-ﬂow net-

works as the basis for an eﬃcient new heuristic approach that incorporates the

conditional bounds into the simplex pivoting process and exploits the eﬃcient,

specialized pure-network simplex technologies. The algorithm was applied to

interval-ﬂow transportation problems with a uniform conditional lower bound

and tested on problems with up to 5000 nodes and 10000 arcs. Empirical

comparisons with CPLEX demonstrate the eﬀectiveness of this methodology,

both in terms of solution quality and processing time.

Managing the Service Supply Chain in the US

Department of Defense: Opportunities and

Challenges

Uday Apte, Geraldo Ferrer, Ira Lewis, and Rene Rendon

Graduate School of Business and Public Policy, Naval Postgraduate School, 555

Dyer Street, Monterey, CA 93943, USA, [email protected]

The services acquisition volume in the US Department of Defense (DoD) has

continued to increase in scope and dollars in the past decade. Between FY 1999

to FY 2003, DoD’s spending on services increased by 66%, and in FY 2003, the

DoD spent over $118 billion or approximately 57% of total DoD’s procurement

dollars on services. In recent years, DoD has spent more on services than on

supplies, equipment and goods, even considering the high value of weapon

systems and large military items. These services belong to a very broad range

of activities ranging from grounds maintenance to space launch operations.

The major categories include professional, administrative, and management

support; construction, repair, and maintenance of facilities and equipment;

information technology; research and development, and medical care.

As DoD’s services acquisition volume continues to increase in scope and

dollars, the agency must keep greater attention to proper acquisition plan-

ning, adequate requirements deﬁnition, suﬃcient price evaluation, and proper

contractor oversight. In many ways, these are the same issues aﬀecting the

acquisition of physical supplies and weapon systems. However, the unique

characteristics of services and the increasing importance of services acquisi-

tion oﬀer a signiﬁcant opportunity for conducting research in the management

of the service supply chain in the Department of Defense.

The objectives of the exploratory research presented in the paper are to

1. analyze the size, structure and trends in DOD’s service supply chain,

2. understand the challenges faced by contracting oﬃcers, program managers

and end users in services acquisition,

3. develop a conceptual framework for understanding and analyzing the sup-

ply chain in services, and

4. provide policy recommendations that can lead to more eﬀective and eﬃ-

cient management of DOD’s spending on services.

In addition to the analysis of service acquisition related data and theory de-

velopment, this research also includes empirical work in terms of site visits

Managing the Service Supply Chain in the US Department of Defense 51

and interviews at Navy, Army and Air Force bases. Addressing issues related

to both theory and practice, this paper makes a modest contribution towards

more eﬀective and eﬃcient management of service acquisition in the Depart-

ment of Defense.

Keywords: Service supply chain, outsourcing, contract management

Analysis of Heuristic Search Methods for

Scheduling Automated Guided Vehicles

Thomas Bednarczyk and Andreas Fink

Chair of Information Systems, Department of Economics,

Helmut-Schmidt-University / Universit¨at der Bundeswehr Hamburg,

Holstenhofweg 85, 22043 Hamburg, Germany, [thomas.bednarczyk,

andreas.fink]@hsu-hamburg.de

We consider the problem of scheduling automated guided vehicles (AGVs) for

processing elementary transportation jobs. This problem, which is a special

case of the general pickup and delivery problem [1, 5], arises, e.g., on seaport

container terminals, where AGVs may be employed to transport containers

from quay cranes that load and unload ships to storage locations on the termi-

nal yard and vice versa [3, 6, 7]. That is, containers are to be moved between

the ship area and the yard using a ﬂeet of vehicles, each of which can carry

one container at a time.

We look at the problem of dispatching AGVs to the transportation jobs

in order to minimize the total time it takes to serve a given set of jobs (e.g.,

minimizing the total time it takes to unload and upload a given set of con-

tainers to a ship and from a ship, respectively). Essentially, given a set of

resources (AGVs) and jobs (transportation requests) with processing times

for the jobs and sequence-dependent setup times for subsequently processed

jobs (driving times between the destination location of a job and the origin

point of the subsequent job), we aim for determining a sorted assignment of

jobs to resources such that the maximum completion time for the resources is

minimized. This problem is N P -hard, which motivates the use and analysis

of heuristics.

Since problems from practice, such as the AGV scheduling problem de-

scribed above, mostly embrace distinctive characteristics, applying heuristics

may imply a costly development of specialized algorithms which hinders the

application of such methods in the real world. On the one hand, this problem

might be partly solved by applying metaheuristics, which are generic with re-

gard to the type of problem and the respective solution space. In practice one

would like to apply metaheuristics by reusing suitable software components

which have to be adapted to the speciﬁc problem at hand in some well-deﬁned

manner. HotFrame [2] provides such metaheuristics software components. On

the other hand, there is the question which metaheuristic and which conﬁg-

uration of some selected metaheuristic may provide best results for speciﬁc

Analysis of Heuristic Search Methods 53

problem instances for some considered problem scenarios. Therefore, we are

interested in analyzing connections between search landscape characteristics

and the performance of heuristic search methods.

We focus on metaheuristics that are based on the local search paradigm:

A greedy local search strategy such as steepest descent means selecting and

performing in each iteration of a search process a best move (i.e., an apparently

most promising change of the current solution); the search stops at a locally

optimal solution with no better neighboring solution. As the solution quality

of such local optima may be unsatisfactory we consider an iterated steepest

descent approach where the local search process restarts after a local optimum

has been obtained by means of some randomized perturbation scheme that

generates a new initial solution. Moreover, we use simulated annealing and

diﬀerent tabu search approaches which employ more intelligent concepts to

overcome local optimality.

By means of search landscape analysis, in particular considering ﬁtness

distance correlations [4], we examine the relationship between the solution

quality and the distance between solutions within a given search landscape.

This provides information about the diﬃculty of problem instances, e.g., in

connection with the valley structure of the underlying search landscape, which

can be used to select and conﬁgure elements of search methods.

References

1. Cordeau, J.-F., Laporte, G., Potvin, J.-Y., Savelsbergh, M.W.P.: Transporta-

tion on demand. Working Paper, CRT-2004-25 (2004).

2. Fink, A., Voß, S.: HotFrame: A heuristic optimization framework. In: Voß, S.,

Woodruﬀ, D.L. (eds.) Optimization Software Class Libraries. Kluwer, Boston,

81–154 (2002).

3. Grunow, M., G¨unther, H.-O., Lehmann, M.: Dispatching multi-load AGVs

in highly automated seaport container terminals. OR Spectrum 26, 211–235

(2004).

4. Hoos, H.H., St¨utzle, T.: Stochastic Local Search, Foundations and Applications.

Morgan Kaufmann, San Francisco (2005).

5. Savelsbergh, M.W.P., Sol, M.: The general pickup and delivery problem. Trans-

portation Science 29, 17–29 (1995).

6. Steenken, D., Voß, S., Stahlbock, R.: Container terminal operation and opera-

tions research – a classiﬁcation and literations review. OR Spectrum 26, 3–49

(2004).

7. Vis, I.F.A., Harika, I.: Comparison of vehicle types at an automated container

terminal. OR Spectrum 26, 117–143 (2004).

Exact and Approximate Algorithms for a Class

of Steiner Tree Problems Arising in Network

Design and Lot Sizing

Alysson M. Costa, Jean-Fran¸cois Cordeau, and Gilbert Laporte

Centre for Research in Transportation and Canada Research Chair in Distribution

Management, HEC Montr´eal, 3000 chemin de la Cˆote-Sainte-Catherine, Montr´eal,

Canada H3T 2A7, [alysson, cordeau, gilbert]@crt.umontreal.ca

Several network design problems can be modeled as Steiner tree problems

with additional constraints, including budget constraints (imposing an upper

bound on the total network cost), and hop constraints (imposing that the path

from the root to any vertex in the solution has a maximum of hhops). Budget

constraints are frequently encountered in the design of distribution or telecom-

munication networks where the goal is to obtain a minimizing-cost network

connecting certain vertices. Hop constraints are also encountered in telecom-

munications, where they are used to model the network reliability or impose

limits on the transmission delays. Moreover, for certain classes of lot-sizing

problems modeled as Steiner tree problems, the addition of hop constraints

enables the consideration of time-capacity constraints. In this work, we deal

with a variation of the Steiner tree problem where, besides costs associated

with the arcs, one also has revenues associated with the vertices. The goal is

to maximize the sum of the collected revenues while respecting both hop and

budget constraints. We propose several mathematical formulations for this

problem and use them to develop branch-and-cut algorithms which are tested

on middle-sized instances. Computational results show that the choice of the

best formulation/algorithm strongly depends on the number of allowed hops.

We also propose a destroy-and-repair heuristic capable of obtaining very good

approximations within short computational times.

Keywords: Prize collecting, network design, Steiner tree problem, budget,

branch-and-cut, hop constraints, lot-sizing, time-capacity constraints.

Supply Chain Management in Archeological

Surveys, Excavations and Scientiﬁc Use

Joachim R. Daduna1and Veit St¨urmer2

1Berlin University of Applied Business Administration, Badensche Straße 50–51,

D-10715 Berlin, Germany, [email protected]

2Winckelmann Institut f¨ur Klassische Arch¨aologie, Humboldt-Universit¨at, Unter

den Linden 6, D-10099 Berlin, Germany

A fundamental problem in archaeology is the recording and management of

large numbers of objects, which represent the basis for the study, evaluation

and reconstruction of excavation results as well as their scientiﬁc presentation.

Here the use of methods in Supply Chain-Management (SCM) when plan-

ning archaeological processes can oﬀer a solution, in particular with regard to

the mandatory management of information. The archaeological processes of

achievement, excavation (procurement), evaluation and reconstruction (pro-

duction) as well as provision for preservation and/or (public)presentation

(distribution) constitute the ﬁrst part in this complex. The second part, man-

agement of evaluation, (in the sense of ‘after sales-achievements’) encompasses

essentially the administration,conservation,restoration and presentation of

excavated objects. Thereby the goal is to install eﬃcient processes in the eval-

uation of material and storage through the use of logistic concepts and tech-

niques in the physical organisation of processes as well as in the development

of a comprehensive management of information. Beginning with a description

of the present structure of processes, a SCM-based concept is presented that

should reveal new possibilities in the area of archaeology.

Real-World Agent-Based Transport

Optimization

Klaus Dorer

Senior Researcher, Whitestein Technologies GmbH (previously Living Systems

GmbH), [email protected]

As with many industries and markets, the logistics sector faces extensive and

fundamental challenges associated with globalization. With shrinking margins

and, in many cases, just barely coping with immense cost pressures, companies

are being driven to substantially revise their product and service oﬀerings,

business processes, and levels of operational excellence. It is widely recognized

that the optimal utilization of available capacity is the single most important

critical success factor for logistics operations.

Whitestein Technologies oﬀers, with its Living System R

°Adaptive Trans-

portation Networks (LS/ATN) software, a sophisticated IT solution that ad-

dresses the needs of logistics companies operating in a dynamic and unpre-

dictable business world. LS/ATN focuses on the management and dispatching

of transportation orders and the optimization, execution and monitoring of

capacities (e.g., trucks).

In this talk we present the agent architecture on which the LS/ATN

bottom-up optimization is based. Agents interact to solve subproblems of

transporting orders that, when consolidated, result in an optimized solution

to the overall problem. Similar to human decision-making, solutions to prob-

lems arise from the interaction of individual decision makers (represented by

software agents), each with their own local knowledge. We present results

obtained by running LS/ATN on real world data of big logistics companies.

Finally we present LS/ATN in a live demo.

Scheduling of Automated Double

Rail-Mounted Gantry Cranes

Ren´e Eisenberg

University of Hamburg, Institute of Information Systems, Von-Melle-Park 5, 20146

Hamburg, Germany, [email protected]

Double rail-mounted gantry cranes (DRMG) depict one of the latest devel-

opments in container terminal handling equipment and have been put into

operation on the Container Terminal Altenwerder at the port of Hamburg in

2002. In a setup like this two rail-mounted gantry cranes of diﬀerent sizes can

serve any stack of a single container block since the super-sized crane is able

to pass the standard crane even when loaded. Containers have to be relocated

to and from vehicles in transfer areas on both sides of a block and if required

within the block to free blocked up from-bin containers. Hence, ﬁve job types

are distinguished. Especially for the latter type ping-pong or cyclic restacking

has to be avoided.

In order to synchronize the cranes with adjacent transportation equipment,

for each container move a transfer date is given and has to be met in order to

avoid delay in horizontal transportation. These transfer dates are determined

by a superordinate planning component covering the whole terminal. Hence,

productivity maximization is limited by the deﬁned transfer dates. In the

oﬄine case, e.g., all information is available in advance, the DRMG scheduling

problem may be seen as a multiple travelling salesman problem with time

windows and, thus, is N P-hard when minimizing the idle movements and

maximizing productivity, respectively.

In practice not all information is present or correct in the ﬁrst place. Also,

in the course of operations especially on the land side of a container block,

where cranes are remotely controlled manually by fewer crane operators than

cranes, transfer times are delayed stochastically. Because both cranes may

serve the whole stacking block, inter crane interferences might lead to longer

crane movements. These conditions cause uncertainty and make this an online

optimization problem and its data may change any time.

In this paper we present simple priority-rules and metaheuristics, but

also a simple branch-and-bound approach considering reduced problem sizes.

Limited time availability in real-time and the online problem characteristic

make algorithm design challenging. Algorithms are tested against a stochastic

discrete-event simulation model of a single block with realistic container load

58 Ren´e Eisenberg

which was implemented by the Hamburger Hafen und Logistik AG (HHLA).

The performance of the algorithms is measured in comparison to a simple

FIFO heuristic with respect to minimum delay of the horizontal transport

and minimum idle crane movements.

Keywords: Container transport, online optimization, travelling salesman

problem with time windows, discrete event simulation

Solving Real-World Vehicle Scheduling and

Routing Problems

Jens Gottlieb

SAP AG, Walldorf, Germany, [email protected]

This talk introduces the vehicle scheduling and routing problem (VSRP), for

which an optimization algorithm is oﬀered in SAP’s supply chain management

solution. The algorithm for the VSRP is sketched, followed by a discussion of

its application to typical real-world scenarios from SAP’s customer base.

Exact and Heuristic Solution of the Global

Supply Chain Problem with Transfer Pricing

and Transportation Cost Allocation

Pierre Hansen1, S´ebastien Le Digabel2, Nenad Mladenovi´c3, and Sylvain

Perron4

1GERAD and HEC Montr´eal, [email protected]

2´

Ecole Polytechnique de Montr´eal, [email protected]

3Brunel University and GERAD, [email protected]

4HEC Montr´eal, [email protected]

In this paper, we consider one of the most important issues for multination-

als, i.e., the determination of transfer prices (prices that a buying subsidiary

of a ﬁrm has to pay to a selling subsidiary of the same ﬁrm). More speciﬁ-

cally, we consider a multinational corporation that attempts to maximize its

global after tax proﬁts by determining the ﬂow of goods, the transfer prices,

and the transportation allocation between each of its subsidiaries. Vidal and

Goetschalckx [4] have formulated this problem as a Bilinear Program (BLP)

where each bilinear term corresponds to the product of two decision variables

representing the ﬂow of goods and the transfer price between two subsidiaries,

respectively. These authors have proposed solving the BLP with an alternate

heuristic algorithm where an initial solution is obtained by linearization. This

local method consists in successively ﬁxing one set of variables and solving

the remaining LP for the other set. The process can be terminated when the

change in the objective function is negligible. Under given conditions, the

solution obtained by the alternate heuristic corresponds to a local optimum.

In this paper, we propose an eﬃcient way of using a new heuristic based on

Variable Neighbourhood Search (VNS) [3]. This algorithm has already been

used for the Pooling Problem [1], which may also be formulated as a BLP.

VNS consists of successively repeating two steps: (i) perturbing the current

solution within a neighbourhood of length k(initially set to 1); and (ii) from

this perturbed point, ﬁnding a new point with a local search. If this new local

optimum is better, it becomes the new current point, and the kparameter is

set again to 1. If this new local optimum is not better, the original current

point is kept, and the kparameter is increased for a bigger perturbation in

step (i). In our implementation of VNS, the local search is performed by the

alternate heuristic method proposed by Vidal and Goetschalckx. The pertur-

Transfer Pricing and Transportation Cost Allocation 61

bation procedure is performed by moving to a feasible extreme point in the

neighbourhood of the current solution.

Since BLP is a particular case of a nonconvex quadratic program with

nonconvex constraints (QP), an exact solution method designed for QP may

be applied to solve BLP. We therefore propose to use the branch and cut

algorithm of Audet et al. [2] for that purpose. This algorithm provides a

globally optimal solution (within given feasibility and optimality tolerances) in

ﬁnite time. The basic idea of this algorithm is to estimate all quadratic terms

by successive linearizations (outer approximations) within an enumeration

tree using Reformulation-Linearization Techniques (RLT). For the BLP case,

using RLT means replacing each bilinear term by a linear variable and adding

linear constraints to force the linear variable to approximate the bilinear term.

The three solution methods (Alternate, VNS, and branch and cut) are

tested on random instances.

References

1. Audet, C., Brimberg, J., Hansen, P., Le Digabel, S., Mladenovi´c, N.: Pooling

Problem: Alternate Formulations and Solution Methods. Management Science

50(6), 761–776 (2004).

2. Audet, C., Hansen, P., Jaumard, B., Savard, G.: A branch and cut algorithm

for nonconvex quadratically constrained quadratic programming. Mathematical

Programming 87(1, Ser. A), 131–152 (2000).

3. Hansen, P., Mladenovi´c, N.: Variable neighborhood search. In: Glover, F.,

Kochenberger, G.A. (eds.) Handbook of Metaheuristics, Kluwer, Boston, 145–

184 (2003).

4. Vidal, C.J., Goetschalckx, M.: A global supply chain model with transfer pricing

and transportation cost allocation. European Journal of Operational Research

129(1), 134–158 (2001).

Planning Problems for Combined Pick-up

Point Allocation, Transportation, and

Production Processes with Time-Varying

Processing Capacities

Christoph Hempsch

Deutsche Post Endowed Chair of Optimization of Distribution Networks, RWTH

Aachen University, Templergraben 64, D-52062 Aachen, Germany,

Some providers of postal or parcel services promise high levels of service to

their customers, e.g., next day delivery, resulting in tight lead times. To meet

these high service levels, complex logistics networks need to be planned and

operated for collection and delivery of mailings. Within such a supply chain

the interaction of processing at and transportation between diﬀerent kinds of

facilities play a vital role. Therefore, postal and parcel services are a good

example for an application where supply chain planning needs to integrate a

wide range of heterogeneous decisions.

The talk starts with a discussion of strategic and operational decisions

within an example postal supply chain. Strategic planning decisions include

the location of distribution centers and the allocation of customers to them

while operational planning tasks include transportation of mailings or routing

of vehicles. Customer allocation may as well be planned within an operational

planning horizon. Also in postal or parcel logistics networks, complex sorting

processes at the distribution centers need to be considered. The amount of het-

erogeneous decisions within this virtual supply chain reﬂects the complexity of

problems operations research is facing nowadays. To reduce this complexity, in

literature problems tend to be decomposed into well known standard problems

such as facility location problems, vehicle routing problems and production

planning problems. Yet, models incorporating heterogeneous decisions are in-

creasingly found in literature. An example are inventory routing problems,

which combine decisions on vehicle routing and required inventory at remote

customer facilities.

Motivated by the described supply chain and the discussion above, a model

covering a comprehensive part of the supply chain is introduced. The model

combines decisions in pick-up point allocation and transportation from pick-

up points to production facilities. Within the use case of a postal service

Planning Problems for Combined Pick-up Point Allocation 63

provider, the pick-up points can represent letter boxes or corporate clients,

while the production facilities may represent sorting centers. Pick-up points

contain information on quantities as well as time windows. The model also

includes production processes at the sorting centers as time-varying processing

capacities. The processing capacities induce input requirements. To model the

processing capacities or input requirements, respectively, the planning horizon

is divided into a discrete set of time intervals.

The resulting model incorporates decisions from diﬀerent logistics func-

tions within a supply chain. Solutions of the model are interesting for decision

support in strategic planning of supply chains like those of postal or parcel

service providers.

The model is formulated as a mixed integer problem. A problem formula-

tion as well as the underlying assumptions are presented. A software prototype

was implemented to solve test instances. Small and medium sized instances

are solved with standard solver software ILOG CPLEX. Allocations of cus-

tomers to production facilities as well as input distributions at the facilities

are visualized by the software for validation and analysis of results. The talk

concludes with examples of research perspectives.

Paradigm Shift in the Supply Chain – Is it

Really Happening?

Britta Kesper and Yuriy Kapys

DHL Solutions GmbH, Godesberger Allee 83-91, 53175 Bonn, Germany

Logistics landscape and requirements have drastically changed over the last

ten years. Manufacturing companies compete more and more on core com-

petencies which are mainly development, product design and production and

marketing. Meanwhile the companies start to cooperate in the ﬁeld of sup-

ply chain management. Such developments increase the industry demand for

knowledge based, adaptive, ﬂexible and collaborative business models. The

LLP/4PL business model is one of them.

As supply chains become more complex, solutions need to be more ad-

vanced. One of the services that is part of the LLP/4PL business model is

Supply Chain Consulting. Its objective is to support customers during their

decision making process at strategic and tactical levels to identify the best

supply chain structure.

DHL Supply Chain Consulting is focusing on three areas to create ad-

vanced logistics solutions, which are network design, carrier selection and

transport optimization. This workshop will use a reference case where the

beneﬁts of network design and transport optimization are presented. Three

customers are competing in their product segments and geographical markets

but cooperate in the ﬁeld of logistics with the support of DHL Exel Supply

Chain.

Support of Bid-Price Generation for

International Large-Scale Plant Projects

Dirk Mattfeld and Jiayi Yang

Technische Universit¨at Braunschweig, Abt-Jerusalem-Str. 4, 38106 Braunschweig,

Germany, [email protected]

Summary. We propose a mathematical model for the support of sourcing and

scheduling decisions for large-scale industrial plant projects. Local content require-

ments and limited production capacity constrain the problem of determining a lower

bound on the bid-price for an industrial plant project. The talk will describe the

problem domain paying particular attention to the rapid development of the East-

Asian market. Challenges for Western industrial engineering vendors are discussed.

According to the special interest group on large-scale plant engineering, a di-

vision of VDMA (Verband Deutscher Maschinen- und Anlagenbau), 80% of

large-scale industrial orders placed in Germany in 2004 and 2005 came from

foreign countries. As this trend will continue, competition in the interna-

tional markets for industrial plants will intensify, causing additional pressure

on prices, increase of the local content requirement and decrease of project

duration.

Local content requirements (LCR) are set up for multiple reasons, such

as supporting domestic industry, developing domestic technological capacity

and ensuring protection for the domestic workforce. In order to satisfy a given

LCR, the German vendor has to decide on the part of the plant to be produced

in the buyers country. This may cause an outﬂow of engineering know-how

and may also increase costs. Furthermore, producing abroad will conﬂict with

a short expected project leadtime.

Most of the articles published on local content rules are more or less the-

oretical treatments in the economics literature, e.g., Grossman [7], Hollander

[8] and Richardson [15]. These studies focus on macroeconomic production

and welfare eﬀects of local content policies. Only few papers look at this topic

from a business management point of view, e.g., by Munson and Rosenblatt

[11]. An overview of LCR for large-scale plant projects is given by Petersen

[13].

Most literature focusing on large-scale plant projects originates from en-

gineering disciplines. Exceptions are Backhaus [1] considering marketing as-

pects, Reiner [14] evaluating price management and Schiller [16] treating com-

66 Dirk Mattfeld and Jiayi Yang

petence management. These studies focus on qualitative approaches whereas

quantitative approaches on project scheduling typically refrain from an appli-

cation viewpoint, e.g., Kolisch [10], Klein [9], Neumann et al. [12], Zimmer-

mann et al. [18].

Strategic network planning in the context of supply chain management

has a strong link to large-scale plant project management, as decisions on

international facility location are to be taken; see, e.g., Goetschalckx and

Fleischmann [5], Vidal and Goetschalckx [17], Geoﬀrion and Powers [4], Cohen

and Lee [3]. These approaches support the strategic planning for sustainable

production within one period.

In conclusion, a deﬁciency on the strategic and tactical planning level

is seen for operations in large-scale industrial plant projects. In particular,

quantitative decision support approaches, addressing the interdependencies

between local content requirement, international facility location decision and

project scheduling are desirable for the phase of bid placement. To ﬁle a tender,

a lower bound on the project’s bid price has to be determined. The bid price

largely depends on sourcing decisions for the project components involved.

Decisions concerning the production location of project components are to

be taken such that total costs of production and transport are minimized.

These decisions are constrained by the LCR. On the other hand, activities

associated with the production of project components are to be scheduled

under limited resource capacities so that a predetermined project due date

is met. The duration of activities largely depends on the location decisions

taken. On the basis of this interrelation a mathematical optimization decision

support model is proposed. This model combines the international facility

location problem and the resource-constrained project scheduling problem.

The optimal solution obtained considers constraints such as the local content

requirement, resource capacities and the expected project lead time.

Support of Bid-Price Generation 67

References

1. Backhaus, K.: Industrieg¨utermarketing, 7th ed., M¨unchen (2003).

2. Burghardt, M.: Projektmanagement: Leitfaden f¨ur die Planung, ¨

Uberwachung

und Steuerung von Entwicklungsprojekten, 6th ed., Erlangen (2002).

3. Cohen, M.A., Lee, H.L.: Resource Deployment Analysis of Global Manufactur-

ing and Distribution Networks. European Journal of Manufacturing and Oper-

ations Management 2(2), 81–104 (1989).

4. Geoﬀrion, A.M., Powers, R.F.: Twenty years of strategic distribution system

design: An evolutionary perspective. Interfaces 25(5), 105–128 (1995).

5. Goetschalckx, M., Fleischmann, B.: Strategic Network Planning. In: Stadtler,

H., Kilger, C. (eds.) Supply Chain Management and Advanced Planning: Con-

cepts, Models, Software and Case Studies. Berlin (2005).

6. Gottwald, K., Stroh, V., Waldmann, T.: Rekord im Ausland – Investitions-

schw¨ache im Inland, Lagebericht der Arbeitsgemeinschaft Großanlagenbau. Ar-

beitsgemeinschaft Großanlagenbau, Lyoner Straße 18, 60528 Frankfurt am Main

(2004).

7. Grossman, G.M.: The Theory of Domestic Content Protection and Content

Preference. The Quarterly Journal of Economics 96(4), 583–603 (1981).

8. Hollander, A.: Content Protection and Transnational Monopoly. Journal of

International Economics 23(3/4), 283–297 (1987).

9. Klein, R.: Scheduling of Resource-Constrained Projects, Boston (1999).

10. Kolisch, R.: Project Scheduling under Resource Constraints, Heidelberg (1995).

11. Munson, C.L., Rosenblatt, M.J.: The Impact of Local Content Rules on Global

Sourcing Decisions. Production and Operations Management 6(3), 277–290

(1997).

12. Neumann, K., Schwindt, C., Zimmermann, J.: Project Scheduling with Time

Windows and Scarce Resources, 2nd ed., Berlin (2003).

13. Petersen, J.: Local Content-Auﬂagen, betriebswirtschaftliche Relevanz und

Handhabung am Beispiel des internationalen Großanlagenbaus, Wiesbaden

(2004).

14. Reiner, N.: Preismanagement im Anlagengesch¨aft: Ein entscheidungs-

orientierter Ansatz zur Angebotspreisbestimmung, Wiesbaden (2002).

15. Richardson, M.: The Eﬀects of a Content Requirement on a Foreign Duopsonist.

Journal of International Economics 31(1/2), 143–155 (1991).

16. Schiller, T.: Kompetenz-Management f¨ur den Anlagenbau, Ansatz, Empirie und

Aufgaben, Wiesbaden (2000).

17. Vidal, C., Goetschalckx, M.: Strategic production-distribution models: A criti-

cal review with emphasis on global supply chain models. European Journal of

Operational Research 98(1), 1–18 (1997).

18. Zimmermann, J., Stark, C., Rieck, J.: Projektplanung: Modelle, Methoden,

Management, Berlin (2005).

Bid Querying Policies in Combinatorial

Auctions for Collaborative Transportation

Planning

Giselher Pankratz

Faculty of Economics and Business Administration, FernUniversit¨at – University of

Hagen, Proﬁlstraße 8, 58084 Hagen, Germany,

Combinatorial Auctions for Collaborative

Transportation Planning

Due to their autonomy-preserving properties, auctions are considered to be

suitable coordination mechanisms in loosely-coupled collaborative systems. In

particular, combinatorial reverse auctions have been proposed in the literature

as an appropriate means for task reallocation in the ﬁeld of collaborative

transportation planning (see, e.g., [4, 3]). This is because combinatorial reverse

auctions oﬀer the bidders the possibility to express valuation dependencies

among transportation requests, thus allowing a more economically eﬃcient

allocation of the requests.

However, combinatorial reverse auctions impose several problems which

up to now impede their dissemination in practice. On the one hand, the auc-

tioneer faces a N P -hard optimization problem when searching for an optimal

allocation of the requests. This problem has been introduced as the winner

determination problem [5] and has been well studied in the literature (see,

e.g., [6]). On the other hand, the exponential bid space loads the bidders

with heavy computational burden: given mrequests to be allocated, there are

2mcombinations of requests a bidder may have to submit bids for. In the

transportation domain, this would require the bidder to solve an individual

N P -hard optimization problem for each and every bundle in order to provide

all valuations.

Recently, preference elicitation has been proposed as an approach for tak-

ing some of the strain oﬀ the bidders [1]. Generally speaking, preference elicita-

tion aims at signiﬁcantly reducing the number of valuations explicitly revealed

by the bidders through an intelligent process of stepwise querying conducted

by the auctioneer who systematically exploits implicit information contained

in previously revealed bids and strictly focuses on relevant information while

querying.

Bid Querying Policies 69

Proposed Bid Querying Policies

Most of the work on preference elicitation presented in the literature deals

with fairly generalized combinatorial auction scenarios. In our contribution,

we propose rather specialized elicitation policies which are more tailored to

collaborative planning situations in transportation. In particular, we make

use of several characteristic properties of the transportation domain in order

to make the bid querying process as eﬃcient as possible. Such properties are,

among others:

1. Valuation dependencies in the transportation domain involve both sub-

additivity and super-additivity. Sub-additivity means that a bidders cost

of a given combination of transportation requests is less than the sum

of the cost of the individual requests due to complementarities between

the requests. Super-additivity, which means that the costs of the bundle

exceeds the sum of its individual requests, occurs, e.g., when two requests

cannot be transported on the same truck due to mutually excluding time

requirements, thus giving rise to additional costs for an extra vehicle.

2. The free disposal assumption holds in the transportation domain. In the

context of transportation reverse auctions, free disposal basically means

that for any transportation request, the cost of a bundle including this

request is never below the cost of the same bundle without that request,

which appears quite reasonable in the transportation domain.

Based on these and other observations, we have developed two diﬀerent bid

querying approaches:

1. The ﬁrst approach takes up and extends research presented by the au-

thors in [3] and [2]. A two-phase querying policy is established: in the

initial bidding phase, the bidders are required to place bids on all sub-

additive bundles, i.e., bundles for which synergies can be realized between

the transportation requests contained. During the second phase, the auc-

tioneer systematically constructs promising allocations from the set of all

submissions. If an allocation contains two or more bids of the same bidder,

the bidder is requested to place a supplementary bid for the set union. As

per deﬁnition of the initial bidding phase, the bidders evaluation of such

supplementary bids must be super-additive. This approach strongly rests

upon the observation that in practice often only a small fraction of all

possible bundles exhibit complementarities. On the other hand, by com-

mitting the bidders to submit all sub-additive bids during the ﬁrst phase,

the auctioneers search process can be organized very eﬃciently.

2. The second approach takes a further step towards relieving the bidders

of unnecessary computational burden by abandoning the requirement to

identify and evaluate all sub-additive bundles in advance. Unlike the ﬁrst

approach, during the ﬁrst phase the auctioneer only requests bids on

rather small bids (e.g., containing up to three transportation requests)

70 Giselher Pankratz

which can be easily evaluated by the bidders. Similarly to the ﬁrst ap-

proach, if the auctioneer identiﬁes a promising allocation in the second

phase which contains one or more bundles for which the exact evaluation

is unknown, the auctioneer requests supplementary bids on these bundles.

In order to keep track of the auctioneers cumulative knowledge about the

bidders valuations, a constraint network is used [1]. Exploiting the free

disposal property, lower and upper bounds on a bidders true value of a

given bundle can be easily determined based on the corresponding values

known for its subbundles and superbundles, respectively [1]. If the valu-

ation of a bundle is updated, e.g., on receiving a supplementary bid, this

information is propagated through the network, thus tightening the lower

and upper bounds involved.

Both approaches have been implemented in a simulation environment using

the .NET environment. At the moment, the approaches are subject to in-

tensive tests using diﬀerent sets of randomly generated problem instances.

Preliminary results have shown a good performance of the approaches. The

approaches and their computational results will be presented at the conference

in detail.

References

1. Hudson, B., Sandholm, T.: Eﬀectiveness of Query Types and Policies for Prefer-

ence Elicitation in Combinatorial Auctions. In: Proceedings of the International

Joint Conference on Autonomous Agents and Multiagent Systems (AAMAS),

New York, 386–393 (2004).

2. Kopfer, H., Pankratz, G., Gehring, H.: Combinatorial Auctions for a Gener-

alized Cooperative Transportation Problem. Extended abstract of a scientiﬁc

talk given at the ODYSSEUS 2000 conference, Chania, Greece (2000).

3. Pankratz, G.: Analyse kombinatorischer Auktionen f¨ur ein Multi-

Agentensystem zur L¨osung des Groupage-Problems kooperierender Spedi-

tionen. In: Inderfurth, K., Schw¨odiauer, G., Domschke, W., Juhnke, F.,

Kleinschmidt, P., W¨ascher, G. (eds.): Operations Research Proceedings 1999,

Springer, Berlin, 443–448 (2000).

4. Sandholm, T.: An Implementation of the Contract Net Protocol Based on

Marginal Cost Calculations. In: Proceedings of the 11th Nat. Conf. on Arti-

ﬁcial Intelligence (AAAI-93), Washington D.C., 256–263 (1993).

5. Sandholm, T.: Algorithm for Optimal Winner Determination in Combinatorial

Auctions. Artiﬁcial Intelligence 135, 1–54 (2002).

6. Sandholm, T., Suri, S., Gilpin, A., Levine, D.: CABOB: A Fast Optimal Al-

gorithm for Winner Determination in Combinatorial Auctions. Management

Science 51, 374–390 (2005).

Application of HotFrame on Tabu Search for

the Multiple Freight Consolidation Problem

Filip Rychnavsk´y

University of Bremen, Schwachhauser Ring 60, 28209 Bremen, Germany,

Consolidation Problem

There are situations when forwarders do not do the carriage on their own and

use services of partners [1]. Reasons can be, e.g., a temporary insuﬃciency of

capacity or a disadvantageous scatter of orders. A forwarder can save a return

to the depot and the scatter can be more suitable for a partner with another

depot.

A freight fee is to be paid for the partner services. It is a result of mar-

ket negotiations and can depend on the actual situation. It does not depend

directly on costs caused at the executing carrier. Factors can be of natural

origins like distance, weight or time. Nonlinearity of fees in relation to the

scale of orders is expected [2]. A possible objective function is the minimiza-

tion of the sum of freights for each shipment. It could result in a lower freight

payment for a bundle of orders than for sending these orders separately. The

costs caused on the side of the executing carrier do not aﬀect establishing the

price (freight fee) for the service. His costs are not known to other members

of the market.

It is supposed that each vehicle has a given weight capacity. The total

sum of the weights of all orders is higher than the capacity of one vehicle.

It means that additional vehicles of subcontractors have to be hired. Orders

are to be allocated (bundled) to the vehicles of subcontractors. The bundle is

shipped to the unload node of some order and the node’s order is unloaded.

The remaining bundle is shipped to the next unload node or it can be split

to two or more subbundles. This represents that the subcontractor would add

one or more additional vehicles.

The real (physical) fulﬁllment of the supply is in the hands of the executing

carrier. The task is to propose the cheapest ﬂow, where all orders reach the

customers and transported weight on each arc does not exceed capacity of

a vehicle. The multiple freight consolidation problem can be interpreted as

ﬂow-oriented problem with capacity restrictions and nonlinear goal function.

72 Filip Rychnavsk´y

The Initial Solution

A starting algorithm was developed that gives us a feasible solution in a non-

combinatorial way. The main aspect is that we can not compute freights before

we know the structure of a ﬂow (used arcs and transported weights on them).

These weights are not known before setting them on arc’s ancestors.

The central idea is to bundle nodes (customers) that are near to each other

and whose orders are of signiﬁcant weights. It can be described as building

compact clusters. The sum of weights of nodes in one cluster may not exceed

vehicle capacity. A modiﬁed gravitation index is computed for each pair of

nodes. Based on the mutual gravity, member nodes of a cluster are selected.

When members of a cluster are set, spanning trees in the clusters can be

constructed. Arcs from the depot to a core node of a cluster are established

ﬁrst. Free nodes are added to the used ones according to their gravitation

power. Nodes with light unload weights are added on the end to prevent the

transporting of huge amounts on small distances.

The last part of the opening procedure is setting weights on used arcs.

The path of each order from the depot to an unload node is known now. So

the unload weight of the order is added to each arc on its path.

Tabu Search

The tabu search was chosen because of the possibility to integrate a lot of

problem speciﬁc knowledge. The HotFrame framework is a template for using

local search methods. The user has to implement its components and set

parameters. The program calls for attributes of solutions and neighbors and

decides about further steps. In our case, a solution is represented as a binary

matrix. Rows symbolise starting nodes and columns goal nodes. This matrix

represents a ﬂow; weights and freights can be derived from it.

Neighborhoods

Two kinds of neighborhoods have been implemented. They both need a special

structure made by the start solution, where the ﬂows consist mostly from

nodes connected after each other. Proposed neighbors are based on a modiﬁed

2-opt-move, where used arcs will be deleted and other arcs set.

•Tail swap move

Each node can have some followers. Let us call the following spanning tree

a tail. If a node changes its antecessors with another node the antecessors

receive new tails. This kind of move can cause large changes in the whole

structure of the solution.

•Two nodes swap move

An arc from the antecessor of a node can be connected directly to the

followers of this node. This node is added to another node, where the

Application of HotFrame 73

same procedure has been done. So two nodes in a solution are reallocated.

The structure of the solution does not change so much.

Special modiﬁcations are to be done, if the swap should be done with nodes

where one belongs to the antecessor line of the second.

To reduce the number of iterations, a selection of pairs of nodes is done.

The idea is to change important nodes that are close to each other. A modiﬁed

gravitation index is used. Not the unload weight of a node, but the weight of

incoming orders is considered.

Implementation

HotFrame [3] requires the deﬁnition of the problem’s speciﬁc components.

Attributes are represented with deleted and built arcs. Tabu status checks if

arcs deleted by the actual neighbor would be the same as the new built arcs of

the move in the tabu list. Tabu threshold is involved in the decision about a

tabu status. Some parts of the original HotFrame had to be changed, because

it is not possible to decide about feasibility of a move before constructing a

solution based on this move. By the selection of the best admissible neighbor

only a restricted set of neighbours is inspected.

Conclusions

A construction algorithm has been programmed and tabu search components

speciﬁc for this problem have been proposed. The algorithm has been imple-

mented in HotFrame. Test problems have been generated because of the lack

of real data. Initial solutions have been improved in the range of 5% to 15%.

The full version of this paper will include technical details of the implemen-

tation.

References

1. Pankratz, G.: Speditionelle Transportdisposition: Modell- und Verfahrensent-

wicklung unter Ber¨ucksichtigung von Dynamik und Fremdvergabe, PhD Thesis,

University of Hagen (2002).

2. Kopfer, H., Rychnavsk´y, F.: Freight optimization problem with approximated

fee function. Presented at the 22nd International Conference of Mathematical

Methods in Economics, Brno (2004).

3. Fink, A., Voß, S.: HotFrame: A Heuristic Optimization Framework. In: Voß, S.,

Woodruﬀ, D.L. (eds.) Optimization Software Class Libraries, Kluwer, Boston,

81–154 (2002).

Simulation Metamodeling of a Perishable

Supply Chain

M.E. Seliaman1and Ab Rahman Ahmad2

1King Fahd Univeristy of Petroleum and Minerals, Dhahran 31261, KSA,

2UTM, Johor, Malaysia, [email protected]

Perishable goods are those goods, which have a ﬁxed or speciﬁed lifetime after

which they are considered unusable, i.e., they cannot be utilized to meet the

demand. The planning and control of supply chain of perishable goods is im-

portant because in real life products like milk, blood, drug, food, vegetables

and some chemicals do have ﬁxed life times after which they will perish. The

presence of these kinds of products after their lifetime will not only occupy

space of the store but also eﬀect the lifetime (damage) of the neighboring

items. In some cases of perishable goods, which consume electricity for their

storage, the loss is greater. The determination of the ordering and replenish-

ment policies, to meet the demand of these types of goods across the supply

chain hence becomes very crucial. The problem becomes diﬃcult when there

are stochastic demands and lead times.

This paper proposes the use of regression metamodels in simulation to

support transportation-inventory decisions within a supply chain of perish-

able products. The supply chain consists of a single production facility and

multiple retailers. (Daily replenishment policy is followed.) We consider a per-

ishable product which has a common deterministic lifetime and units of the

same age will fail together if they are not taken by demands. We assume that

the demand is a random variable. We further assume that demands are in

batches, with random batch sizes and inter-demand times with back orders.

The objective is to determine the optimum ordering plan that minimizes the

expected total cost across the supply chain. The total cost includes the or-

dering costs, inventory holding costs, transportation costs, shortage cost and

the cost due to outdated inventories. The developed simulation model rep-

resents the described supply chain. After careful veriﬁcation and validation,

post-simulation regression analysis is used to determine the optimum oper-

ating conditions for this perishable supply chain system. Data from a local

distribution supply chain will be used to demonstrate the model.

Non-Cooperative Games in Liner Shipping

Strategic Alliances

Xiaoning Shi1and Stefan Voß2

1Department of International Shipping Management Shanghai, Jiao Tong

University, 1954 Hua Shan Road, Shanghai 200030, P. R. China,

2University of Hamburg, Institute of Information Systems, Von-Melle-Park 5,

20146 Hamburg, Germany, [email protected]

Nowadays, there is a trend to establish new business linkages and alliances

within the shipping industry together with customers, suppliers, competitors,

consultants, and other companies. A number of studies have attempted to ex-

plain this phenomenon occuring in the liner shipping industry using a variety

of conceptual and theoretical frameworks. This paper focuses on liner ship-

pings strategic alliances and their establishment and transformation within

the framework of non-cooperative game theory. The concepts developed and

improved by Nash, Selten and Harsanyi should be considered as eﬀective and

capable tools to analyse motivations, competitive structures, strategies and

potential pay-oﬀs in the turbulent liner shipping industry.

Not only a liner shipping company could be regarded as a player in shipping

alliances, but also a liner shipping strategic alliance itself could be viewed as

a player when it competes with other alliances. However, in this paper, we

pay more attention to the former model assuming those liner companies are

unable to make enforceable contracts through outside parties. The aims of

this paper are to

•indicate the motivations of short-run cooperation among several liner car-

riers;

•analyse pros and cons of being members in liner shipping strategic al-

liances;

•explain the departure of a player when it faces turbulence and unpre-

dictable shipping circumstances;

•advise ways to contain long-run alliances stability by increasing beneﬁts

while decreasing drawbacks.

Among those four main points, the diﬀerences between short term cooperation

and long term alliance are the amounts of sub-games and the potential pay-oﬀ

in future. Consequently, we set up speciﬁc models based on non-cooperative

games and repeated games to give those diﬀerences clear explanations. The

76 Xiaoning Shi and Stefan Voß

outcome of this paper shall be helpful for those liner shipping carriers who

attempt to succeed in the shipping industry with greater eﬃciency, better

customer service and lower cost.

Keywords: Game theory, non-cooperative, shipping, strategic alliance

Container Terminal Operation and Operations

Research

Dirk Steenken1, Stefan Voß2, and Robert Stahlbock2

1Former: HHLA, IS – Information Systems/Equipment Control, 20457 Hamburg,

Germany, [email protected]

2University of Hamburg, Institute of Information Systems, Von-Melle-Park 5,

20146 Hamburg, Germany, [email protected]

Containers came into the market for international conveyance of sea freight al-

most ﬁve decades ago. The breakthrough was achieved with large investments

in specially designed ships, adapted seaport terminals with suitable equip-

ment, and availability of containers. Today over 60 % of the world’s deep-sea

general cargo is transported in containers, whereas some routes are even con-

tainerized up to 100 %. International containerization market analysis still

shows high increasing rates for container freight transportation in the future.

This leads to higher demands on seaport container terminals, container logis-

tics and management as well as on technical equipment, resulting in an in-

creased competition between seaports. The seaports mainly compete for ocean

carrier patronage and short sea operators as well as for the land-based truck

and railroad services. The competitiveness of a container seaport is marked

by diﬀerent success factors, particularly the time in port for ships, combined

with low rates for loading and discharging. Therefore, a crucial competitive

advantage is the rapid turnover of the containers, which corresponds to a re-

duction of a ship’s time in port and of the costs of the transshipment process

itself.

The objective of this paper is to provide a survey and a classiﬁcation of

container terminal operations. Moreover, examples for applications of opera-

tions research models – including exact methods, heuristic methods as well as

simulation based approaches – are mentioned. For a detailed description and

a comprehensive list of references see [1].

References

1. Steenken, D., Voß, S., Stahlbock, R.: Container terminal operation and opera-

tions research – a classiﬁcation and literations review. OR Spectrum 26, 3–49

(2004).

Mixed Integer Models for Optimized

Production Planning Under Uncertainty

David L. Woodruﬀ

Graduate School of Management, UC Davis, Davis CA 95616, USA,

We concern ourselves with the process of making optimized production plan-

ning and inventory decisions in the face of low frequency, high impact uncer-

tainty, which takes the form of a small number of discrete scenarios. In this talk

we will describe general formulations as well as the general solution method,

progressive hedging. Computational results for a particular real-world, mixed

integer, inventory problem that is very large will be described.

Part III

Contributions Not Presented

Simulation Optimization of the Cross Dock

Door Assignment Problem

Uwe Aickelin and Adrian Adewunmi

University of Nottingham, Jubilee Campus, Wollaton Road, Nottingham, NG8

1BB, UK, [uxa,aqa]@cs.nott.ac.uk http://www.cs.nott.ac.uk/~uxa/

Summary. We present the Cross Dock Door Assignment Problem. This involves

assigning destinations to outbound dock doors of Cross Dock centres, such that the

total cost by material handling equipment is minimized. Proposed is a two fold solu-

tion; simulation and optimization of the simulation model – simulation optimization.

The novel aspect of our approach is that we intend to use discrete event simulation

to simulate the arrangement and assignment of destinations to dock doors. We will

include some random variability in the discrete simulation model, i.e., variation in

freight ﬂow within the Cross Dock centre. The purpose of applying discrete event

simulation to the Cross Dock assignment problem is to derive a more realistic ob-

jective function. Furthermore, we intend to minimise the realistic objective function

derived by the discrete event simulation. This will be achieved using Memetic algo-

rithms. The advantage of using Memetic algorithms is that it combines Local Search

with Genetic Algorithms. The Cross Dock Door Assignment Problem is a new ap-

plication domain to Memetic Algorithms and as such will prove to be challenging

and interesting research.

Keywords: Cross dock door assignment problem, discrete event simulation,

optimization, genetic algorithms

Introduction

Traditionally, warehousing companies have had the following functions: receiv-

ing, storage, order picking and shipping. They have found storage and order

picking cost intensive; in order to abate cost, a strategy of Cross Docking was

implemented. The goal of Cross Docking is to sort, consolidate and transfer

incoming freight onto outgoing trailers for delivery to pre-determined desti-

nations [2]. Presently, each incoming trailer is assigned an available inbound

door as soon as it arrives and each outbound trailer is assigned a speciﬁc

single outbound door. The eﬃciency of the Cross Dock centre is dependent

on factors which include, i.e., an optimal scheduling of dock doors, the reduc-

tion of Cross Dock congestion and a minimum travelling distance for material

82 Uwe Aickelin and Adrian Adewunmi

handling equipment. We are interested in investigating the Cross Dock Door

Assignment Problem.

Cross Dock Door Assignment Problem

The Cross Dock Door Assignment Problem is related to the Dock Door As-

signment Problem, ﬁrst formulated by [5]. The Cross Dock Door Assignment

Problem objective is to ﬁnd the optimal arrangement of a Cross Dock centre’s

inbound and outbound doors and the most eﬃcient assignment of destina-

tions to outbound doors, such that the distance travelled by material han-

dling equipment is minimized. It is assumed that there are Iinbound doors,

Joutbound doors, Morigins and Ndestinations for the Cross Dock centre,

I≥Mand J≥N. Let Xim = 1 if origin mis assigned to inbound door

i, Xim = 0 otherwise. Let Ynj = 1 if destination nis assigned to outbound

door j, Ynj = 0 otherwise. Let dij represent the distance between inbound

door iand outbound door j. Let wmn represent the number of trips required

by the material handling equipment to move items originating from mto the

Cross Dock door where freight destined for nis being consolidated. A math-

ematical formulation for the Cross Dock Door Assignment Problem based on

work by [5] is presented below:

Minimize:

I

P

i=1

J

P

j=1

M

P

m=1

N

P

n=1

dij wmnxmi ynj

+ Constraints

Proposed Plan of Work

Discrete Event Simulation

In order to ﬁnd solutions to problems, representation by mathematical models

has been a reasonable approach. These mathematical relationships (i.e., equa-

tions, inequalities, etc.) have a parallel correlation with relationships that exist

within problems. However, mathematical models take a standard static form,

which can make modelling certain aspects of a problem diﬃcult. [1] consider

objective functions as “Black Boxes”. The reason is that there are peculiar

problems like inventory management that need simulation runs to obtain a

globally optimal design. By using discrete event simulation to simulate the

dock door assignment, we will assess the performance of input parameters in

these relationships and gain a better understanding of the inherent relation-

ships that exist in the Cross Dock Door Assignment Problem. These relation-

ships in the objective function, which are not visible by a simple mathematical

formulation, will become clearer. Thus we will derive a more realistic objec-

tive function. Amongst others, we will simulate the ﬂow of freight between

inbound and outbound doors and as well as the arrangement and destination

assignment of the Cross Dock doors.

Simulation Optimization of the Cross Dock Door Assignment Problem 83

Simulation Optimisation

As well as simulating the diﬀerent possible Cross Dock Door Assignments,

we intend to ﬁnd the optimal door to freight and trailer to door assignment

using Memetic Algorithms. This can be achieved by optimizing the door as-

signment from the discrete event simulation models that performed the best

against predetermined criteria. In essence, the results of the discrete event sim-

ulation without the models inherent stochastic noise will be used as the ﬁtness

function for the Memetic Algorithm optimiser. [4] present two simulated an-

nealing algorithms that are designed to handle noisy objective functions; we

are interested in comparing the performance of Memetic algorithms to other

popular heuristics in relation to noisy objective functions [3]. The objective

is to demonstrate the ability of Memetic Algorithms to produce high quality

solutions with minimal computational expense.

Summary

As iterated, the purpose of this research is to present a novel search method

for solving the Cross Dock Door Assignment Problem, it emphasises on using

discrete event simulation to simulating various dock door destination assign-

ments, taking into consideration the stochastic nature of the problem, in order

to obtain a more realistic objective function. Furthermore, the pursuit of an

optimal solution will focus on exploring a global search for promising solutions

within the whole feasible region, while exploiting local searches for optimal

solutions to the Cross Dock Door Assignment Problem.

References

1. Baumert, S., Smith, L.R.: Pure Random Search Noisy Objective Functions.

The University of Michigan, Technical Report (2002).

2. Li, Y., Lim, A., Rodrigues, B.: Cross Docking – JIT scheduling with time win-

dows. Journal of the Operational Research Society 55, 1342–1351 (2004).

3. Merz, P., Freisleben, B.: A comparison of memetic algorithms, tabu search,

and ant colonies for the quadratic assignment problem. Proceedings of the 1999

Congress on Evolutionary Computation, 2063–2070 (1999).

4. Prudius, A.A., Andrad´ottir, S.: Two simulated anealing algorithms for noisy

objective functions. In: Kuhl, M.E., Steiger, N.M., Armstrong, F.B., Joines,

J.A. (eds.) Proceedings of the 2005 Winter Simulation Conference (2005).

5. Tsui, Y.L., Chang, C.-H.: A microcomputer based decision support tool for

assigning dock doors in freight yards. Computers & Industrial Engineering 19

(1–4), 309–312 (1990).

Heuristics for the Multi-Layer Design of

MPLS/SDH/WDM Networks

Holger H¨oller and Stefan Voß

University of Hamburg, Institute of Information Systems, Von-Melle-Park 5, 20146

Hamburg, Germany,

Current high-speed networks are mainly based on Synchronous Digital Hi-

erarchy (SDH) or its American equivalent Synchronous Optical Network

(SONET), Wavelength Division Multiplex (WDM) and Internet Protocol /

Multi Protocol Label Switching (IP/MPLS). The multi-layer network design

problem, as treated, e.g., in [3] is to decide which combination of equipment

and routing will be able to carry the given (protected) demands with the lowest

investment in new equipment. The models presented here rely on a speciﬁc set

of common network components: switches, cross-connects, multiplexers, port-

cards and so on. The diﬀerent layers considered are the ﬁber-layer, WDM,

SDH and IP/MPLS. Some of these layers might also have diﬀerent line speeds,

e.g., 2.5Gbit/s and 10Gbit/s. However, we do not consider native packet pro-

cessing but restrict ourselves to IP/MPLS traﬃc engineering scenarios with

dedicated (though unidirectional) label switched paths (LSPs).

To solve such multi-layer network design problems we have designed a

network optimizer that works according to the following general outline based

on [2]. Starting from a feasible solution, e.g., a shortest path routing, demands

are consecutively rerouted until no more reduction in the overall invest for the

network infrastructure can be achieved.

During the last years, several ideas from diﬀerent metaheuristic concepts

have contributed to the current state of our network optimizer and some more

are envisaged for the future. This ranges from simple random Multi-Start

over the Greedy Randomized Adaptive Search Procedure (GRASP), aspects

similar to Variable Neighborhood Search (VNS) up to the Pilot Method. Some

of these concepts have become an integral part while some others are modules

that can be used optionally.

The random components serve primarily as a means for diversiﬁcation,

while the GRASP ideas are used to intensify the search in promising parts

of the solution space. Similar to the VNS, the neighborhoods that are used

during the search can be changed. We do not limit these changes to the size

and position of the neighborhood, but we also change its inner structure. This

Heuristics for the Multi-Layer Design of MPLS/SDH/WDM Networks 85

might be a change in the granularity, e.g., from bundle rerouting to single

demand rerouting or some more fundamental changes. Also, the Pilot Method

might be used to evaluate which neighborhood will be the new choice.

While we have incorporated ideas from many well-known metaheuristics,

we do not always use the respective concepts exactly in their original sense.

Instead, we try to combine and modify ideas in maybe new ways inspired

by the speciﬁc needs of our problem and our past experience in the ﬁeld of

network planning. References to the underlying original metaheuristics can be

found, e.g., in the following publications. A general introduction to GRASP

is given in [4] and a bibliography can be found in [5]. Details with regard to

VNS can be found, e.g., in [1], while the Pilot Method is described in [6].

A mixed integer programming formulation solved by CPLEX serves as a

benchmark for the quality of the heuristics. However, due to the complexity of

the problem, it is only feasible for small to medium sized problem instances,

also strongly depending on the details of the equipment modeling and the

freedom of choice for the layers at intermediate nodes.

Keywords: Network design, SDH, WDM, GRASP, VNS, pilot method

References

1. Hansen, P., Mladenovi´c, N.: Variable neighborhood search. In: Glover, F.,

Kochenberger, G.A. (eds.) Handbook of Metaheuristics, Kluwer, Boston, 145–

184 (2003).

2. H¨oller, H., Voß, S.: A heuristic approach for combined equipment-planning and

routing in multi-layer SDH/WDM networks. European Journal of Operational

Research 171 (3), 787–796, 2006.

3. Melian, B., Laguna, M., Moreno-Perez, J.A.: Capacity expansion of ﬁber optic

networks with WDM systems: Problem formulation and comparative analysis.

Computers & Operations Research 31 (3), 461–472, 2004.

4. Pitsoulis, L.S., Resende, M.G.C.: Greedy randomized adaptive search

procedures, 2001. http://www.research.att.com/∼mgcr/papers.html, state:

31.05.2006.

5. Festa, P., Resende, M.G.C.: An updated bibliography of GRASP, 2003. http:

//www.research.att.com/∼mgcr/doc/graspbib.pdf, state: 31.05.2006.

6. Voß, S., Fink, A., Duin, C.: Looking ahead with the pilot method. Annals of

Operations Research 136, 285–302, 2005.

Heuristics for the Multi-Layer Design of MPLS/SDH/WDM Networks 87

- … Agents have their own beliefs about and preferences over the status of their environment and have particular sets of actions to change it. As a distributed problem-solving paradigm, an agent-based approach breaks complex problems into small and manageable subproblems [13,6,32,16]. Due to these properties, an agent-based scheduling model can operate in environments that are partly unknown and unpredictable. To address the need of reduced complexity and increased fault tolerance and flexibility, a continuous feedback control approach has been developed in distributed manufacturing applications [23]. …Distributed adaptive control of production scheduling and machine capacityArticle
- Apr 2007
- J MANUF SYST

- Sohyung Cho
Vittaldas V. Prabhu

- Organizacje komercyjne i niekomercyjne wobec wzmożonej konkurencji oraz wzrastających wymagań konsumentów (Tom 8)BookFull-text available
- Jan 2009

Anna Ujwary-Gil

- Adam Nalepka

- Decentralized Resource Allocation and Scheduling via Walrasian Auctions with Negotiable AgentsConference PaperFull-text available
- Aug 2010

- HuaXing Chen
Hoong Chuin Lau

- Land Use Changes in Shendong Coal Mining AreaArticle
- Oct 2010

- Guoliang Chen
Yunjia Wang

## Recommended publications

Discover more publications, questions and projects in Negotiation

Conference Paper

Probabilistic Analysis of Local Search on Random Instances of Constraint Satisfaction.

January 1996

Read more

Article

Approximation of Constraint Satisfaction via local search

August 1995

Without Abstract

Read more

Conference Paper

Probabilistic Analysis of Local Search and NP-Completeness Result for Constraint Satisfaction (Exten…

January 1996

Read more

Conference Paper

Parameters Optimization of Resources in a Container Terminal

July 2013

The world seaborne trade has been developing considerably in the last decade, mainly due to globalization and continued development of emerging countries. This world growth has an influence on the development of ports and maritime terminals. It is always a hot topic about how to make the port run in the maximum productivity with minimum cost. But in order to ameliorate the productivity in a… [Show full abstract]

Read more

Discover more

About

News

Company

Careers

Support

Help center

FAQ

Business solutions

Recruiting

Advertising

© ResearchGate 2018 . All rights reserved.

- Imprint
- Terms
- Privacy

or

Discover by subject area

Join for free

Log in