Professional Documents
Culture Documents
Models Methods For Project Selection: Concepts From Management Science, Finance and Information Technology
Models Methods For Project Selection: Concepts From Management Science, Finance and Information Technology
Models Methods For Project Selection: Concepts From Management Science, Finance and Information Technology
PROJECT SELECTION
Concepts from Management Science, Finance and
Information Technology
INTERNATIONAL SERIES IN
OPERATIONS RESEARCH & MANAGEMENT SCIENCE
Frederick S. Hillier, Series Editor Stanford University
by
Samuel B. Graves
Boston College
Jeffrey L. Ringuest
Boston College
with
Andres L. Medaglia
A C.I.P. Catalogue record for this book is available from the Library of Congress.
Graves, Samuel B. & Ringuest, Jeffrey L. / MODELS & METHODS FOR PROJECT
SELECTION: Concepts /rom Management Science, Finance & Information Technology
All rights reserved. No part ofthis work may be reproduced, stored in a retrieval
system, or transmitted in any form or by any means, electronic, mechanical,
photocopying, microfilming, recording, or otherwise, without the written
permission from the Publisher, with the exception of any material supplied
specifically for the purpose of being entered and executed on a computer system,
for exclusive use by the purchaser of the work.
DEDICATION v
TABLE OF CONTENTS VB
PREFACE Xl
CHAPTER!
THE LINEAR MULTIOBJECTIVE PROJECT SELECTION
PROBLEM
1.1 Introduction 1
1.2 An Example from the Literature 3
1.3 Towards a More General Multiobjective Formulation 7
1.4 A Second Example 9
1.5 Summary and Conclusions 11
References 15
CHAPTER 2
EVALUATING COMPETING INVESTMENTS
2.1 Introduction 19
2.2 Adjusting for Time Alone 19
2.3 Adjusting for Time and Risk 22
2.4 Conclusions 27
References 30
CHAPTER 3
THE LINEAR PROJECT SELECTION PROBLEM: AN
ALTERNATIVE TO NET PRESENT VALUE
3.1 Introduction 31
3.2 An Example 32
3.3 The Behavioral Implications ofNPV 33
3.4 Multiple Objective Decision Methods 35
3.5 Conclusions 38
References 40
CHAPTER 4
CHOOSING THE BEST SOLUTION IN A PROJECT SELECTION
PROBLEM WITH MULTIPLE OBJECTIVES
4.1 Introduction 41
4.2 Some Early Approaches 42
4.3 A Matching and Grouping Approach 46
4.4 A Stochastic Screening Approach 54
4.5 Conclusions 62
References 64
CHAPTERS
EVALUATING A PORTFOLIO OF PROJECT INVESTMENTS
5.1 Introduction 65
5.2 Examples 67
5.3 Conclusions 74
References 76
CHAPTER 6
CONDITIONAL STOCHASTIC DOMINANCE IN PROJECT
PORTFOLIO SELECTION
6.1 Introduction 77
6.2 The Model 78
6.3 Summary and Conclusions 89
References 93
CHAPTER 7
MEAN-GINI ANALYSIS IN PROJECT SELECTION
7.1 Introduction 95
7.2 The Model 101
7.3 Conclusions 114
References 117
CHAPTER 8
A SAMPLING-BASED METHOD FOR GENERATING
NONDOMINATED SOLUTIONS IN STOCHASTIC MOMP
PROBLEMS
8.1 Introduction 119
8.2 Stochastic, Nondominated Solutions 123
8.3 Sampling Approaches to Solving MOMP Problems 125
8.4 Computational Issues 126
8.5 Summary and Conclusions 134
Appendix 8.1
Example SAS Code 135
References 144
CHAPTER 9
AN INTERACTIVE MULTIOBJECTIVE COMPLEX SEARCH FOR
STOCHASTIC PROBLEMS
9.1 Introduction 147
9.2 Direct Search Methods 149
9.3 Applying Complex Search to Multiobjective Mathema-
tical Programming.Problems 152
9.4 An Example of Multiobjective Complex Search 155
9.5 Conclusions 158
References 160
V111
CHAPTER 10
AN EVOLUTIONARY ALGORITHM FOR PROJECT SELECTION
PROBLEMS BASED ON STOCHASTIC MULTIOBJECTIVE
LINEARLY CONSTRAINED OPTIMIZATION
10.1 Introduction 163
10.2 Stochastic Multiobjective Linearly Constrained
Programs 164
10.3 Multiobjective Evolutionary-Based Algorithm 166
10.4 Computational Examples 174
10.5 Summary and Conclusions 183
Appendix 10.1
Input File for the Algorithm Parameters for the
SMOLCP Example 185
Appendix 10.2
Java Program that Defines the First Objective Function in the
SMOLCP Example 187
References 188
INDEX 191
IX
PREFACE
The project selection problem is one that has been given much atten-
tion in the literature. In the project selection problem the decision maker is
required to allocate limited resources across an available set of projects, for
example, research and development (R&D) projects, information technology
(IT) projects, or other capital spending projects. In choosing which projects
to fund, the decision maker must have some concrete objective in mind, e.g.,
maximization of profit or market share or perhaps minimization of time to
market. And in some cases the decision maker may wish to simultaneously
satisfy more than one of these objectives. Most often, these multiple objec-
tives will be in conflict, resulting in a more complicated decision making task.
The decision maker may be able to partially fund some projects, or
conversely some projects may involve a binary decision of fully funding or
not funding at all. The decision maker may also have to resolve issues of in-
terdependency-that is that the value of funding an additional individual proj-
ect may vary depending upon the success or failure of projects that are already
in the portfolio. The decision maker then must take all these factors into ac-
count in seeking an appropriate project selection model, choosing a method-
ology which evaluates the appropriate objective(s), subject to relevant re-
source constraints as well as constraints relating to projects with binary (full
or none at all) funding restrictions.
There is a considerable body of literature describing an abundant va-
riety of models designed for the project selection problem. For our purposes
here, the literature can be broken down into two main streams: that which we
will label the traditional management science stream and that which we will
call the financial modeling stream. The first stream, the management science
literature, derives largely from mathematical programming treatments along
with some use of classical decision theory. In order to use these approaches it
is usually assumed that the existing decision alternatives (projects) are rea-
sonably well-known and that the necessary information for modeling these
alternatives is at hand at the initiation of the planning process. The majority
of the management science models treat the decision process of choosing a set
of new projects to form a wholly new portfolio. But some of the models we
will present also address the problem of adding one or more new projects to
an already existing project portfolio. Most of the research in this body of lit-
erature is confined to decisions which are made at one point in time, that is,
the models are static in the sense that they represent a one-time decision to
assemble or analyze a given portfolio.
An important junction in the decision making process occurs when
the decision maker chooses the appropriate objective(s). If a single objective
(e.g., market share) is chosen, then the problem may be handled with ordinary
mathematical programming techniques that have been used in project plan-
ning models for some time now. If, however, the decision maker wishes to
pursue several objectives simultaneously (e.g., maximization of revenue in
each of several future time periods), some form of multiobjective program-
ming will be needed. It is our belief that this multiobjective case is the more
realistic one, thus, in this book, we will show several applications of multiob-
jective programming to the project portfolio problem.
A key assumption of the mathematical programming models above is
that all relevant information about the projects is known. However, this may
not always be true. Some allocation decisions must be made in the presence
of uncertainty. Uncertainty may exist concerning the ultimate result of a proj-
ect (e.g., the amount of revenue) or the success or failure likelihood may be
known only as a probability distribution. Uncertainty may be represented by
probability distributions around the coefficients in the objective function or in
the constraints. In this book we will illustrate treatments for each of the above
forms of uncertainty. We will, however, assume that adequate information is
available to represent these projects in the model. The required information
may be in the form, for example, of a probability of project success or a prob-
ability distribution around a coefficient in the objective function (e.g. project
return).
When we are dealing with uncertainty and multiple objectives, we
may need to resort to the use of stochastic dominance criteria to screen a set
of solutions. Stochastic dominance is appropriate for all probability distribu-
tions and is minimally restrictive with respect to thedecision maker's utility
function. In this monograph there are several forms of stochastic dominance,
which are of interest. First order stochastic dominance simply compares the
cumulative distribution functions for two projects and makes the choice on
this basis alone. The first order criterion is applicable to all decision makers
with monotone utility functions; that is, decision makers who prefer higher
returns to lower ones and/or those who prefer less risk to more risk. In some
instances, the first order criterion does not yield an unambiguous choice. In
these cases it may be necessary to resort to second order stochastic
dominance. The decision calculus here is based on the area between the two
cumulative distribution curves. This second-order criterion is appropriate for
a narrower class of decision makers, those who are risk-averse. We will also
in some cases apply a conditional stochastic dominance criterion. Conditional
stochastic dominance analysis identifies dominant and nondominant projects
conditioned on the projects, which make up the current portfolio. This
criterion requires no explicit knowledge of the decision maker's utility
function and is applicable to all risk averse decision makers. Finally, in some
cases we will apply a stochastic dominance criteria which compares
xii
alternatives based on the expected value and the probability of achieving
desired levels of one or more measures.
The second main stream of literature that we wish to consider here is
that which derives from financial portfolio research, some of which can be
applied directly to the project selection model. The earliest such methodology
is the mean-variance model, which compares two projects according to their
mean and variance. Any project which has a higher mean return for a given
variance or a lower variance for a given mean will be preferred. The mean-
variance approach is limited in its practical applicability because it involves
rather strict assumptions about the decision maker's risk orientation, and be-
cause it may require a large number of pairwise comparisons if there is a large
number of projects under consideration.
Another source of financial literature is the traditional financial opti-
mization models in which the variance in the portfolio returns is minimized
subject to a constraint on expected return. This approach, however, requires
solution of a non-linear optimization problem (The portfolio variance is non-
linear.) which may be impractical when there are large numbers of projects to
consider. A more recent treatment of the project selection problem deriving
from the financial literature is the mean-Gini approach. This, like the mean-
variance criterion, is a two-parameter method. That is, only two parameters
must be estimated for each R&D project. (Mean and variance for the mean-
variance approach; mean and Gini coefficient for the mean-Gini approach).
The Gini coefficient, like the variance, is a measure of dispersion in outcomes
or investment risk. However, when the Gini is used, as opposed to the vari-
ance, preferred portfolios may be designed based on a simple heuristic. In the
mean-Gini analysis--as will be shown later in this work--there is one impor-
tant difference between the financial application and the project portfolio ap-
plication. In the financial application the necessary probability distributions
for each security are unknown and are estimated from sample (i.e., market)
data. For project portfolio applications the probability distributions (describ-
ing various levels of success) tend to be simple discrete distributions, permit-
ting (in principle) complete enumeration of all possible outcomes in the port-
folio.
Largely in parallel to (and distinct from) the R&D project portfolio
selection literature is a body of work describing Information Technology (IT)
portfolio selection. The objectives of this body of work have much in com-
mon with the R&D portfolio modeling work and we will assume in most of
this work that our models apply equally well to R&D or IT project selection
problems. Essentially the problem here is to find the optimal set (portfolio) of
IT projects when resources are constrained. The greatest difference between
the IT models and the R&D models is the heightened importance of project
interdependencies in the IT models. In IT project applications, as opposed to
R&D project applications, there is, due to the very nature of the projects, an
Xlll
increased incidence of interdependency. For example, two IT projects may
share some identical sections of computer code. They may share as well
hardware such as workstations and networks. And whereas R&D interaction
modeling is typically pairwise, realistic IT modeling requires that higher-or-
der interdependencies (among three or more projects) be represented.
Project selection models that capture the various characteristics de-
scribed above (interdependencies, uncertainty and the ability to partially fund
or the requirement to either not fund or fully fund projects) may be quite
complicated mathematically. These characteristics can lead to complicated
mathematical programs that include one or more objectives that may be linear
or nonlinear, deterministic or stochastic and with variables that are real, inte-
ger, or binary. Appropriate solution procedures for these complex mathe-
matical programs are also needed. We will present these as well.
This monograph is intended to pull together in a single publication the
latest work in this field. It is not intended as a survey, but rather as a vehicle
for establishing some unity in the field of project selection modeling. The
models presented here rely heavily on mathematical programming but also
draw from decision theory and finance. Our intention is to present models
that are broadly applicable in the project selection context, to describe the as-
sumptions and limitations of these models, and to provide solution method-
ologies appropriate for solving these models. The chapter outline below traces
out the main themes of the book.
XVI
Chapter 9 shows another approach to the stochastic non-linear prob-
lem, this one having non-linearity in both the objective function and the con-
straints. The method presented in Chapter 8 uses random sampling to identify
a set of feasible solutions for the project selection problem. It is shown that
this random sampling can be computationally burdensome so that the method
presented in Chapter 8 is limited to a small number of real valued variables.
In Chapter 9 complex search is applied to the project selection problem. In
complex search only the initial feasible solution is generated randomly. Then
subsequent solutions are found using a systematic search. In this way the
computational burden is greatly reduced. The example problem shown in
Chapter 9 illustrates a model with a non-linear objective function and several
non-linear constraints, which are solved by relying on the progressive defini-
tion of the decision maker's preferences.
Chapter 10 concludes our treatment of non-linearity and
interdependence in the project selection problem. This chapter presents a new
algorithm that treats the project selection problem in cases of uncertain
objectives, partial funding, and interdependencies in the objectives. The
method shown here is based on a multiobjective evolutionary algorithm and
on concepts from linear programming and presents the decision maker with a
very good approximation of the true efficient frontier. The
algorithm is able to solve project selection problems modeled as
multiobjective linear programs and multiobjective non-linear programs with
linear constraints.
We would like to acknowledge here our indebtedness to individuals
who have contributed to the research in this volume. In particular we want to
acknowledge Randy Case, with whom we have co-authored work in this area
and whose data is used in several places here. We are also indebted to the
many editors and referees who have helped to sharpen and clarify the research
we have performed over the past ten years. Finally, we want to thank
Suzanne Proulx for technical assistance in producing this manuscript.
xvii