Optimization Analysis For Design and Planning of Multi-Project Programs

European Journal of Operational Research 107 (1998) 492±506
Optimization analysis for design and planning of multi-project

programs
Victor D. Wiley a, Richard F. Deckro b,*
, Jack A. Jackson Jr. b
a
Oce of Aerospace Studies, Directorate of Requirements, 3550 Aberdeen Ave. SE, Kirtland AFB, NM 87117-5776, USA
b
Department of Operational Sciences, Air Force Institute of Technology, 2950 P. Street, Wright-Patterson AFB, OH 45433-7765, USA
Received 1 March 1996; received in revised form 1 April 1997
Abstract
Program management concerns the long term planning, coordination, and control of major technological, engineer-
ing, scienti®c, and/or developmental activities. In general, programs tend to be exceptionally large, consisting of several
parallel or sequential projects or groups of projects. While a large number of modeling eorts have focused at the pro-
ject level, this paper looks at the application of optimization techniques to the initial design and development of multi-
project programs. The classic work breakdown structure (WBS) is used as a framework to provide an aggregate model
to investigate the eects of funding levels, resource allocation, and program, project, and component durations. Decom-
position, sensitivity analysis, and parametric programming are utilized to provide the decision maker detailed informa-
tion for establishing program parameters, conditions, and bounds. Ó 1998 Elsevier Science B.V. All rights reserved.
Keywords: Program/project management; Scheduling; Planning; Decomposition; Sensitivity analysis; Parametric

programming
1. Introduction
The Project Management Institute suggests that program management involves the planning, organiz-
ing, stang, directing, and controlling of a program to achieve its objectives [19], p. 142. The responsibility
for the program's planning, resource allocation, direction, and control are the duties of the program man-
ager, who must determine the proper information and control system for coordinating all project activities
[19], p. 118. In production and operational settings, aggregate level planning is a hierarchical structure that
parallels the management decision process, reduces the amount of detailed information needed for planning
purposes, and increases the accuracy of forecasts [3], p. 22, [20]. The Work Breakdown Structure (WBS)
utilized in program and project management is also a hierarchical planning structure. The WBS is de®ned
*
Corresponding author. E-mail: rdeckro@a®t.af.mil.
0377-2217/98/$19.00 Ó 1998 Elsevier Science B.V. All rights reserved.

PII S 0 3 7 7 - 2 2 1 7 ( 9 7 ) 0 0 3 3 4 - 2
V.D. Wiley et al. / European Journal of Operational Research 107 (1998) 492±506 493
as the framework relating statements of work, contract line items, con®guration items, technical and man-
agement reports, and the hardware, software, and data elements of the system [12], p. 88. The WBS trans-
lates the results of the systems engineering analysis of the system requirements into a structure of the
products and services which comprise the entire work eort [12], p. 88, just as the aggregate production plan
translates aggregated requirements into material, workforce, and capacity requirements in a production en-
vironment [2,3,20], among others.
In general, programs tend to be exceptionally large, consisting of several parallel or sequential projects
or groups of projects. While a large number of modeling eorts have focused on the project level, [6±
10,14,18,22], among others, little work examines the application of optimization techniques to the initial
design and development of multi-project programs.
The actual initiation of a program is fundamentally dierent from the management and control of a pro-
gram or projects. Decisions made when a program is established dictate how the program will be managed
throughout its life. By considering questions of resource and funding levels before timelines are set, the pro-
gram manager is in a better position to eectively develop, control, and manage the program. Carter and
Baker [4], p. 18 have suggested that a change that would cost x dollars at the conception stage will cost 10x
at the design stage, 100x at the prototype stage, and 1000x at the production stage. By investigating the
eects on program costs and duration due to funding and other resource constraints before a program
has been established is essential for an organization to remain competitive. Such information would be
valuable in setting contract due dates, milestones, and funding levels for individual projects and activities
that make up the program. Kerzner describes 14 typical reasons why program plans fail [15], p. 605. These
points underscore the need for an analysis tool that provides the program decision makers with the ability
to generate feasible program schedules, determine appropriate funding levels for each project, and to con-
duct sensitivity analysis on these schedules and funding levels [11].
In this paper, we discuss a program aggregate planning model, based upon the WBS, to establish project
parameters. We assume the analysis is conducted in the early portion of the program's initiation stage and
is to be used to aid the program manager and his or her sta to establish program and project requirements.
As in conventional production hierarchical planning, the variables will be assumed to be continuous at this
level of aggregation.
At the aggregate program manager/planner level, the classic Dantzig±Wolfe block angular structure may
be derived directly from the WBS [5,7,16]. Program level requirements and resources serve as the overall
binding constraints, while the individual project requirements provide the distinct blocks within the struc-
ture. With the solution of a particular funding/duration scenario, sensitivity analysis is conducted to pro-
vide the program designer with insight into the inter-relationships in the system [13].
Section 2 outlines the basic multi-project programming model. Section 3 brie¯y reviews the Dantzig±
Wolfe approach used in this study. Section 4 presents an illustrative three project R&D program example
and summarizes the analysis of this example. Section 5 develops an extension to the basic model which al-
lows the consideration of either crashing or extending activities. The example developed and solved in Sec-
tion 4 is then resolved, allowing crashing or extending of activities in Section 6. The paper concludes in
Section 7 with a review of the work and some suggestions for future research.
2. Multi-project program planning model
The structure of the program planning model presented in this section has been developed from the hi-
erarchical project structure. While the model will be presented from the program level, it could be utilized
for planning at any level of the WBS hierarchy.
A. Model assumptions. The following initial assumptions are made:
1. The ``normal'' activity duration times are known and constant,
494 V.D. Wiley et al. / European Journal of Operational Research 107 (1998) 492±506
2. activity crashing is permitted to an upper bound by period for each activity,

3. the time/cost trade-o function for activity ``crashing'' may be modeled as a linear relation, and
4. all variables will be non-negative.
Normal is considered to be whatever the program sponsor or ®rm considers normal for the type of ac-
tivity under consideration [6,8]. Thus, the de®nition of normal duration may vary from organization to or-
ganization.
B. Model formulation. In this section, the basic model utilized in this study is presented. An adaptation of
the notation presented by Wiest and Levy [22], p. 80, has been used to set up the model. The variables and
parameters are de®ned as follows:
Variables
xi realization time of node i
yn number of periods activity (i,j) is crashed
e number of periods the program ®nishes early where i represents the node number of a speci®c event
in a speci®c project
Parameters
Tij normal duration of activity (i,j)
Dij the cost of activity (i,j) at normal duration
Kij cost per period activity (i,j) is crashed
Mij upper bound on yij
Bk the budget for project k
B the budget for the program
OH overhead cost per time unit
T total program time allowed
TCD target completion date (TCD 6 T)
n the terminal node for the program
R reward bonus per period for ®nishing early
Ik the set of connected arcs (i, j) in project k
I the set of connected arcs (i, j) in the program
P` the program upper limit on additional periods for type ` personnel
Pk` the project k upper limit on additional periods per type ` personnel
Sij` the type ` personnel cost associated with ``crashing'' activity (i, j)
Model PP
min z OHP R Pxn i j KijP yijP Objective (1)
s.t. OHx
PPP i j n K ij yij 6 B ÿ i j Dij for i; j 2 I Program budget (2)
i j ` S ij` yij 6 P `l
; 8` and i; j 2I Program personnel (3)
ÿxi xj yij P Tij ; for i; j 2 I Program precedence (4)
xn 6 TCD Program due date (5)
ÿxiP
P xj yij Tij ;Pfor P i; j 2 Ik 8k. Project precedence (6)
Pi Pj P K ij y ij 6 B k ÿ i j Dij 8k and i; j 2 Ik Project k Budget (7)
i j ` Sij` yij 6 Pk` ; 8k; ` and i; j 2 Ik Project personnel (8)
yij 6 Mij for i; j 2 I Max Crash for (i, j) (9)
xi ; yij ; P 0 Non-negativity (10)
While not a necessary requirement, the model assumes every activity may be crashed. Resource con-
straints on both program and project budgets and on available classes of personnel are assumed. However,
any limitations which are required by a given program and which can be modeled as linear expressions may
be incorporated into the model. While most of the model relations are of a standard format, a brief discus-
sion of key variations is given below. The reader is directed to [6±10,14,21,22], among others, for a more
detailed discussion of the standard relations.
Objective function. The objective function may take any of a number of standard forms, as required by
the modeling environment. A least cost program schedule objective, with ®xed costs excluded, is used in this
presentation.
X X
min z OH Rxn i
K y :
j ij ij
This objective minimizes the net of overhead and any early bonus when the program is completed by xn
plus the sum of the cost of crashing. It is assumed that the program is completed on or before its target
completion date when e is not explicitly included in the objective function.
Personnel-month constraints. The personnel-month constraints for the program and the projects limit the
amount of additional personnel-months available for crashing activities. The number of personnel-months
required by each activity for various classi®cations of personnel are determined as a function of the ``crash''
cost associated with each activity.
Precedence constraints. Standard ®nish-to-start precedence constraints for crashing models have been
utilized for each project and the program. In a multi-project program, particularly programs driven by cut-
ting edge technological innovations, some inter-relations between projects may exist. While these inter-re-
lations may be modeled as a large, loosely connected project, this work views such situations as two projects
with the inter-relation model as a ``dummy'' or linking precedence constraint connecting the appropriate
events in each project.
3. Dantzig±Wolfe decomposition model for multi-project programs
A review of the complete model given by expressions (1)±(10) reveals an interesting structure. There is a
set of constraints common to the overall program. These include the program's budget constraint (2), pro-
gram personnel constraints (3), linking constraints (4), and the program due date constraint (5). In addi-
tion, there are certain project speci®c constraints; project precedence (6), project budget (7), project
personnel (8) and crashing limits (9). Given that we have assumed a level of aggregation which supports
the use of non-negative continuous variables in our model, the mathematical program represented by
the aggregate multi-project program planning model possess the classic Dantzig±Wolfe block angular struc-
ture. The common requirements apply throughout the program and will be incorporated into the Dantzig±
Wolfe master problem. The project speci®c constraints will be incorporated into the Dantzig±Wolfe sub-
problems, with one subproblem for each project in the program. It should be noted that while the proce-
dure developed here has been applied at the top most levels of the WBS, it could be applied at any speci®c
level of the WBS of interest to the program manager.
The Dantzig±Wolfe master problem and subproblem are expressed below, based on the presentation
style used by Lasdon [16].
(A) Master problem form

X Xÿ
min z k
c xi k OH Rxn M dev
i k k ik
X Xÿ
i
s:t: a k xk kik ÿ dev B;
Xk Xi ÿ
a xi k 6 P bm for m 1; . . . ; p;
i k k ik
Xk
k 1; 8k;
i ik
kik P 0;
where p is the number of coupling constraints, ck the objective function coecients associated with project
k, ak the coupling constraint coecient matrix associated with project k, xik the ith proposal associated with
project k, kik the weight associated with xik , dev the ``deviational'' variable associated with the mth coupling
constraint as needed, bm the RHS associated with the mth coupling constraint, and M is the cost per unit of
dev (a very large number).
The ®rst constraint in our master problem is not in the traditional Lasdon [16] presentation. The imple-
mentation of Dantzig±Wolfe decomposition requires a starting basic feasible solution in the master prob-
lem. In a large program, a combination of feasible project schedules that provides a feasible program
schedule may not be readily available. To account for this, a ``deviational'' variable, dev, has been added
at an appropriately high objective function cost to the overall budget constraint to drive the subproblem
solutions generated at each Dantzig±Wolfe iteration toward feasibility. In eect, dev has relaxed the bud-
getary constraint. While big M was suggested to weight dev in this formulation, other weights, such as the
cost of capital, are possible, as dictated by the operational situation. As long as the precedence relations and
activity durations will allow the program to complete by the due date, a solution (which may not be bud-
getarily feasible) will be attained.
The addition of the dev variable avoids the necessity of searching for feasible solutions to the subprob-
lems to initiate the decomposition process. As will be shown, by adding the deviation to the budget, it also
provides an opportunity to examine potential schedules which are not feasible for the current proposed
funding level. Once a basic feasible solution is identi®ed, the deviational variable is forced out of the basis
by the master problem.
(B) Subproblem form

The subproblem formation follows the traditional approach without additions.
min z ck ÿ pak xk

X X X X
s:t: i j
Kij yij 6 B k ÿ i
D 8k and i; j 2 Ik ;
j ij
X X
i
S y 6 PHkl 8k; l and i; j 2 Ik ;
j ijl ij
ÿxi xj yij P Tij for i; j 2 Ik ;
0 6 yij 6 Mij for i; j 2 Ik ;
where ck is the objective function coecients associated with project k, ak the coupling constraint coecient
matrix associated with project k, xk the variables associated with project k, and p is the shadow prices as-
sociated with the common constraints received from the master problem.
All the approaches developed over the years to exploit the block angular structure of the Dantzig±Wolfe
model may be applied to the solution and analysis of this model. Once the dev variable has been forced to
zero, feasible schedules and resource allocations are available to the program manager. Duality also allows
the decomposition approach to be utilized as a heuristic. If the possible improvements do not warrant fur-
ther iterations, the solution procedure may be halted with a feasible schedule and allocation, although not
an optimal schedule.
If the program manager is developing his or her program in conjunction with the speci®c project
managers, he or she might also consider utilizing the planning information provided at each feasible iter-
ation. Bamoul and Fabian [1] develop an analysis method for allocating corporate resources among decen-
tralized divisions based on a Dantzig±Wolfe decomposition of the corporate planning model. If we replace
``corporate'' with ``program'' and ``division'' with ``project'', the approach developed by Bamoul and Fa-
bian, and built on by others, can be directly applied. The individual project managers would compete with
each other through the program's master problem for shared resources. Such an analysis would allow a
program manger to more accurately evaluate the eects of a strong champion's requests for resources.
While the squeaky wheel may get the oil, we now have a mechanism to evaluate whether the squeak's eect
is worth the oil.
4. Illustrative example
To illustrate the program design planning model, a simple three project R&D program will be used. The
program's project networks are given in Fig. 1. At the high level of aggregation represented by this plan-
ning stage, the project ``activities'' may be considered to be major components in the WBS. Project A has 12
activities (or components), Project B has 19 and Project C has 16. Nodes have been labeled with an alpha-
numeric designation where the letter represents the project and the number represents the particular node
within the project.
Each project will have a set of precedence constraints dictated by the structure of the network in Fig. 1.
In addition, the program will have the overall precedence constraints linking the individual project comple-
tions, the realizations of nodes A11, B15, and C12, to the program completion node xn . The dotted arrow
represents an inter-project relationship between Project A and Project B. While any number of inter-rela-
tions could have been modeled, we have included only one to highlight the approach.
The R&D program has a budgetary limit, and considers two classes of personnel ± managerial and en-
gineering/scienti®c. These resource limits also extend to each of the projects. The program's overall com-
pletion date, TCD, is 84 months, with a planning horizon, T, of 90 months. Table 1 summarizes the
program and project bounds for budget and personnel.
The full mathematical programming model for this illustrative three project program requires 87 vari-
ables, of which 48 are bounded variables, and 65 technological constraints. The decomposition's master
and subproblems were solved using CPLEX. A MATHCAD based routine was utilized to update the
Fig. 1. Illustrative example.

Table 1
Project and program parameters
Project Budget (Bk ) Overhead cost Upper bound on Upper bound on
(in $10,000) per month (OH) additional engineer additional manager
in $10,000) months (Pk1 , P1 ) months (Pk2 , P2 )
A 1180 325 75
B 1635 125 25
C 1370 200 50
Overall 4185 0.305 650 150
weightings. Inputs were developed on an Excel spreadsheet. A visual basic program was then used to con-
vert the model into MPS format to input into CPLEX. The individual activity information is summarized
in Table 2.
The basic model represented above was solved via Dantzig±Wolfe decomposition in seven iterations of
the decomposition approach. Fig. 2 presents the program cost and duration associated with the solution at
each iteration of the master problem. The program durations are in parenthesis above the bar graphs that
represent the total program cost, with ®xed costs included. Table 3 summarizes the actual objective func-
tion values (without ®xed elements) with the durations for each iteration. The z values for iterations 1 and 2
are omitted, as they included the weighted dev term.
In the ®rst two iterations the dev relaxation allowed mathematically feasible solutions (due to the relax-
ation of the budget constraint) that are not feasible for the R&D program's speci®cations. The variable dev
speci®cally identi®es the budget overage.
From iteration 3 onwards, the solutions represent feasible program solutions to the original model. Note
that at iteration 4, the decision maker could, based on weak duality, stop the procedure as being close en-
ough to the optimal program schedule (especially at this aggregate level). While no improvement is made in
the program duration after iteration 4, an additional $123,915 was saved from iterations 4±7 by varying the
crashing of speci®c activities.
Once the solutions generated by the iterations are feasible for the relaxed requirement(s), the underlying
schedules could be implemented. The decomposition procedure gives the program manager a portfolio of
possible schedules. While not shown here, a complete schedule is produced at each iteration, with the appro-
priate funding and personnel levels. Traditional sensitivity analysis is also available to the program manager.
Fig. 3 represents the personnel-month requirements associated with the solution at each iteration of the
master problem. Initially, personnel levels rise as the model attempts to satisfy the relaxed program budget.
After the R&D programs budget requirement is met in iteration 3, more personnel (and more crashing) are
utilized to shorten the program in iteration 4. While the program duration does not change after iteration 4,
we do see variations in workforce levels which reduce overall costs until the optimal solution is reached in
iteration 7.
Additional analysis is possible. Beyond the project ``bidding'' suggested by Bamoul and Fabian's ap-
proach [1] and other single variation sensitivity analyses, a parametric programming analysis was carried
out. One reason for utilizing the decomposition approach is to solve very large problems; however, the
parametric programming codes available to these researchers were based on standard linear programming,
not decomposed models. To implement the parametric programming analysis, the optimal solution to the
Dantzig±Wolfe decomposition was used as an advanced start basis to solve the original model in LINDO.
The advanced start basis, which is optimal, allows the parametric analysis to be conducted even on a rel-
atively large program.
Fig. 4 summarizes a parametric analysis of work force availability for crashing. Starting with the levels
of personnel-months for managers and engineer/scientists allocated in the optimal solution of the decom-
position, crashing work force levels in each class of employee resulted in a reduction in work force in each
Table 2
Activity information
Activity Normal duration (Tij ) Normal cost (Dij ) Upper bound crashing (Mij ) Cost of crashing (Kij )
A1,A2 3.0 71.8 1.7 15.0
A1,A4 4.4 15.1 2.5 2.0
A1,A3 20.2 73.3 5.7 17.4
A2,A5 2.8 63.1 1.1 11.1
A4,A7 11.1 19.5 3.5 1.8
A3,A10 26.0 17.0 5.0 1.5
A5,A6 32.8 99.4 7.8 23.7
A6,A8 27.8 73.8 4.2 16.9
A7,A10 8.1 71.0 1.8 12.6
A8,A9 13.8 42.2 ) )
A9,A11 5.9 83.6 2.2 15.8
A10,A11 1.5 45.6 1.0 10.3
B1,B2 27.1 52.2 9.7 1.7
B1,B8 35.4 60.1 22.8 0.7
B1,B5 2.5 42.5 1.7 8.1
B2,B3 25.3 4.6 17.7 0.3
B2,B6 9.0 33.7 8.9 0.6
B3,B4 9.1 82.4 8.7 7.3
B3,B7 25.5 21.7 3.6 5.1
B4,B14 32.7 40.5 8.3 7.5
B5,B9 6.2 49.4 3.4 11.1
B5,B10 24.8 68.5 8.5 3.4
B6,B7 3.3 20.8 0.8 3.2
B6,B8 15.6 85.6 3.3 9.8
B7,B11 5.3 27.7 4.8 2.3
B8,B14 27.4 35.2 16.6 4.5
B9,B12 14.0 95.0 6.0 2.4
B10,B13 18.1 17.0 6.8 1.3
B11,B14 32.7 62.7 18.1 6.3
B12,B15 21.0 69.5 15.0 1.3
B13,B15 15.4 37.6 14.5 0.4
B14,B15 3.4 25.0 2.4 0.6
C1,C2 20.3 75.7 1.9 5.2
C1,C3 32.5 94.4 14.6 11.3
C1,C6 24.2 36.9 11.1 4.6
C2,C3 2.7 53.8 2.6 1.5
C2,C5 19.2 95.8 10.2 23.1
C3,C4 25.8 34.5 4.0 1.1
C4,C5 16.2 39.0 13.9 1.8
C4,C7 18.9 87.0 10.5 6.0
C5,C8 9.6 8.1 9.5 1.5
C6,C8 21.2 64.3 4.2 7.5
C6,C10 28.7 61.6 11.7 5.9
C7,C12 2.4 23.4 1.5 3.2
C8,C9 9.5 8.1 3.6 1.0
C9,C11 6.8 9.2 3.5 1.3
C10,C12 2.4 73.3 1.7 12.1
C11,C12 28.8 17.2 15.2 1.5
class at a rate consistent with the ratio of managers to engineer/scientists expected in the program (3.75 : 1
for the number of engineers to managers, with managers receiving 125% of the pay of engineers). Worker
months by class were reduced within the range of feasible program budgets.
Fig. 2. Dantzig±Wolfe master problem iteration results.
Table 3
Dantzig±Wolfe master iteration results
Iteration Duration z
1 129.2 )
2 119.3 )
3 86.1 $59,648,472
4 79.5 56,504,200
5 79.5 56,414,740
6 79.5 56,390,249
7 79.5 56,380,285
As can be seen in Fig. 4, as the personnel available for crashing is reduced, the cost initially remains ¯at
as the program reallocates crashing activities, but eventually rises as personnel limitations prevent the mod-
el from shortening the program. This creates additional costs in overhead and loss of early completion bo-
nuses. In this analysis, the program target completion date and horizon limits were eliminated. After the
reductions in personnel represented by the fourth iteration (bar 4) of the parametric analysis, the solutions
generated no longer satis®ed the programs target completion date, 84 months. After the eighth iteration,
the program was unable to satisfy its total horizon due date of 90 months. The budget constraint limited
any further trade-o between crashing personnel availability reductions and the increase in program dura-
tion.
A parametric analysis of the program budget was also completed. Starting with the initial budget and
the decomposition solution, the budget was reduced to zero. This analysis identi®ed multiple optimal so-
lutions. It also identi®ed a solution that had slightly greater schedule duration than the optimal, but had
the same cost vector. As overall funding was reduced, contingency funds were driven from the program.
This had the eect of increasing the programs duration and reducing the schedules ¯exibility.
Fig. 3. Dantzig±Wolfe master problem minimum personnel-months.
Fig. 4. Parametric analysis of personnel-months.

The ability to execute traditional sensitivity analysis and parametric analysis allows the program man-
ager powerful insight into program trade-os. One obvious question would be the analysis of a trade-o
between project budgets. In a large organization, where a project may impact several programs, it may
be desirable to accelerate the completion of a particular project, even if such an acceleration is suboptimal
to the speci®c program, if it advances overall organizational goals. The sensitivity and parametric analysis
opportunities provided by the model developed here enables such analysis.
5. Extended model
The initial example has provided the program manager with valuable data on constructing an initial pro-
gram plan where crashing is allowed. In many operational settings, particularly where contracts have not
yet been let or timetables set, it may also be possible to extend some non-critical activities. The assumption
is that extended activities would be completed in a time frame longer than their ``normal'' duration, but at a
cost below their normal cost. The savings from extending an activity need not be the same as the cost per
period for crashing the activity. Like crashing, it may not be possible to extend every activity. A model
which considers both activity extending and activity crashing is developed below.
A. Extended model assumptions. The following assumptions are added to the base model assumptions of
Section 2.
(5) Activity ``extending'' is allowed to an upper bound of Nij periods for each activity (i, j)
(6) The time/cost trade-o function for activity extending will be modeled as a linear relation at this ag-
gregate level
B. Extended model formulation. To incorporate activity extending, the following variables are intro-
duced. Let zij be the number of time units activity (i, j) extended, jij cost per time unit activity (i, j) extended
(zij ), #ij` the personnel-months per time unit activity (i, j) extended (zij ), and Nij the upper bound on zij
Objective function. The extended model's new variables change only the form of the minimum cost ob-
jective function in the base model. Introducing the extended model variables into the base model objective
function yields the following objective:
X X X X
min z OH Rxn i
K y ÿ
j ij ij i
j z :
j ij ij
The negative coecient in front of the activity extending variables represents the program increasing
available funds for using less than the normal amount of resources for an activity. This reward is not a jus-
ti®cation for the extension of a program, rather it allows a program decision maker to take resources away
from non-critical path activities and apply those resources to critical path activities. In this model, a critical
path activity would be extended only if the cost savings gained by extending the activity outweighed the
additional overhead and bonus opportunity loss incurred by lengthening the overall program duration.
Of course, extending an activity or activities will lower the slack (and ¯exibility) available in the program.
Budget constraints. The budget constraints, at both project and program level, for the extended model
include the same summation term introduced into the objective function. The budget constraints have the
following forms for project and program, respectively:
X X X X X X
i
K y ÿ
j ij ij i
j z 6 Bk ÿ
j ij ij i
D
j ij
8k and i; j 2 Ik ;
X X X X X X
i
K y ÿ
j ij ij i
j z 6B ÿ
j ij ij i j
Dij 8i; j 2 I:
Personnel-month constraints. The changes in the personnel-month constraints re¯ect the additional
months made available by not applying the normal personnel-month requirements to extended activities.
The equations for the personnel-month constraints for the project and the program, respectively, are given
as follows:
X X X X X X
i j `
S ij` yij ÿ i j
# z 6 Pk` 8k; ` and i; j 2 Ik ;
` ij` ij
X X X X X X
i j
S y
` ij` ij
ÿ i j
# z
` ij` ij
6 P` 8` and i; j 2 I:
Due date constraints. The due date constraint of the base model is changed to represents the absolute
date that the program cannot be extended beyond, T. If the activity extension variables, zij , are not all
bounded above by the Nij , then this constraint acts as the limiting factor on how long an activity can be
extended.
Precedence constraints. The precedence constraints now re¯ect the additional activity extension
variables. The precedence relation constraints for the project and the program, respectively, are as
follows:
ÿ xi xj yij ÿ zij P Tij for i; j 2 Ik 8k;
ÿ xi xj yij ÿ zij P Tij for i; j 2 I:

Since the term yij ÿ zij creates two linearly dependent columns for any (i, j), the dependence prevents an
activity from being simultaneously crashed and extended in the optimal solution.
6. Extended example
The illustrative example presented in Section 4 serves as the base model for the extended ÿ model. The
coecients for the activity extending costs are ÿhalf of the coecients for activity crashing jij 0:5Kij .
The personnel-months coecients are similar #ijl 0:5ijl . ÿThe upper bound on the activity extending
variables is twice the upper bound of the crashing variables Nij 2Mij :
The results from the decomposition technique are similar to those presented in Section 4. In the solving
of the subproblems, the activities are either crashed, extended, or left alone. However, the initial interme-
diate master problem iterations can result in convex combinations of activities that may have activities be-
ing both crashed and extended.
Fig. 5 summarizes the results of the iterations of the R&D program example where extend and crashing
are allowed, while Fig. 6 displays the personnel usage at each iteration of the Dantzig±Wolfe approach. The
®rst two iteration bases contain the dev and do not satisfy the program budget. At iterations 3±5, the pro-
gram schedules generated by the master program contain convex combinations of dependent vectors; there-
fore, these program schedules are meaningless in the extended model solution space. However, for
iterations 6±9, the master problem solutions correspond to feasible program schedules in the extended mod-
el solution space. These schedules could be utilized in planning a program.
A comparison of the crashing-only model's solution to the crashing and extending model's solution re-
veals that the crashing and extending model requires a shorter duration (71.1 months vs. 79.53 months) and
has a cost which is approximately $900,000 lower. By judiciously utilizing the funds gained from extending
some activities to crash other critical activities, a program plan with a shorter horizon and a superior cost/
bonus pro®le can often be developed.
It should be noted that allowing the possibility of extending some or all the activities could result in an
increase in the number of critical activities. Activities whose resources are ``bled away'' to ``feed'' critical
activities may eventually be extended to the point where they become critical or near critical. This may
be controlled, in part, by a careful analysis of upper bounds on extending some or all the activities. In ad-
Fig. 5. Dantzig±Wolfe master problem iteration results ± extended model.
Fig. 6. Dantzig±Wolfe master problem minimum personnel-months ± extended model.
dition, the inclusion of the option to extend activities increases the solution space. The decomposition ap-
proach took nine iterations to ®nd an optimal solution, rather than the seven iterations taken by the crash-
only model for this example.
7. Conclusions
In this paper we have developed a method utilizing the WBS and Dantzig±Wolfe decomposition to gen-
erate feasible aggregate level multi-project program plans and schedules. The Dantzig±Wolfe procedure
provides a means of generating interim solutions and their appropriate funding pro®les. The decision maker
may then choose any one of these solutions besides the optimal solution based upon his/her own experience
and risk tolerance. To implement the Dantzig±Wolfe decomposition technique, deviational variables were
used to relax some of the constraints in the Dantzig±Wolfe decomposition master problem until a starting
basic feasible solution could be attained. This precluded the need to search for convex combinations of the
subproblem solutions that would be feasible to the master problem
Upon generating the optimal solution in the Dantzig±Wolfe decomposition procedure, a traditional sen-
sitivity analysis was conducted on this solution. The traditional sensitivity analysis provided information on
how variable the optimal assignment of resources would be to changes in a single factor.
If the information from the Dantzig±Wolfe decomposition technique (schedules, funding levels, and sen-
sitivity analysis) point to the need for more information about certain parameters within the model, a para-
metric analysis may be conducted. This analysis was undertaken.
There are several areas available for future research. A cursory review of the subproblems' structure re-
veals an underlying network structure. It may be possible to exploit this structure utilizing network ap-
proaches which accommodate side constraints.
While not exhibiting a purely parallel structure, the Dantzig±Wolfe approach appears to be an obvious
area to apply parallel programming [17]. The subproblems can be solved in parallel, waiting only for the
master program to be formed and solved at each iteration.
The model presented here, coupled with the post-optimality analysis options outlined, provide the pro-
gram manager with a vast array of alternatives to investigate options in designing and initiating multi-pro-
ject programs.
Acknowledgements
The authors wish to thank the anonymous reviewers and the editors for their helpful comments and sug-
gestions. This work was supported in part by the Air Force Space Warfare Center.
References
[1] W.J. Baumol, T. Fabian, Decomposition, pricing for decentralization and external economies, Management Science 11 (1964) 1±
32.
[2] G.R. Bitran, D. Tirupati, Hierarchical production planning, in: S.C. Graves, A.H.G. Rinnooy Kan, P.H. Zipkin (Eds.),
Handbooks in Operations Research and Management Science: Logistics of Production and Inventory, vol. 4, Elsevier,
Amsterdam, The Netherlands, 1993.
[3] D. Candea, Issues of hierarchical planning in multi-stage production systems, Technical Report No. 134. Contract N00014-75-C-
0556, Operations Research Center, Cambridge, MA, 1977 (AD-A042559).
[4] D.E. Carter, B.S. Baker, CE Concurrent Engineering: The Product Development Environment for the 1990s, Addison-Wesley,
Reading, MA, 1992.
[5] G.B. Dantzig, P. Wolfe, Decomposition principle for linear programs, Operations Research 8 (1960) 101±111.
[6] R.F. Deckro, J.E. Hebert, W.A. Verdini, Project scheduling with work packages, OMEGA International Journal of Management
Science 20 (1992) 169±182.
[7] R.F. Deckro, E.P. Winkofsky, J.E. Hebert, R. Gagnon, A decomposition approach to multi-project scheduling, European
Journal of Operational Research 51 (1991) 110±118.
[8] R.F. Deckro, J.E. Hebert, Resource constrained project crashing, OMEGA International Journal of Management Science 17
(1989) 69±79.
[9] E. Demeulemeester, W. Herroelen, W.P. Simpson, S. Baroum, J.H. Patterson, K.-K. Yang, On a paper by Christo®des et al. for
solving the multiple-resource constrained, single project scheduling problem, European Journal of Operational Research 76
(1994) 218±228.
[10] E. Demeulemeester, W. Herroelen, A branch-and-bound procedure for the resource-constrained project scheduling problem,
Management Science 38 (1992) 1803±1818.
[11] Department of Defense, Defense Acquisition, DOD Directive 5000.1, USD(A), Washington, 23 February 1991.
[12] A. Dincerler, Project scheduling in project-oriented production systems: ORC-85-11, Contract N00014-76-C-0134, Operations
Research Center, Berkeley, CA, 1985 (AD-A159980).
[13] A.M. Georion, The purpose of mathematical programming is insight, not numbers, Interfaces 7 (1976) 81±92.
[14] J.E. Kelley, Critical-path planning and scheduling: mathematical basis, Operations Research 9 (1961) 296±321.
[15] H. Kerzner, Project Management: A Systems Approach to Planning, Scheduling, and Controlling, 5th ed., Van Nostrand
Reinhold, New York, 1995.
[16] L.S. Lasdon, Optimization Theory for Large Systems, MacMillan, New York, 1970.
[17] A. Nagurney, D.-S. Kim, Parallel computation of large-scale dynamic market network equilibria via time period decomposition,
Mathematical Computer Modelling 15 (1991) 55±67.
[18] J.H. Patterson, A comparison of exact approaches for solving the multiple constrained resource, project scheduling problem,
Management Science 30 (1984) 854±867.
[19] L.C. Stuckenbruck (Ed.), The Implementation of Project Management: The ProfessionalÕs Handbook, Project Management
Institute, Addison-Wesley, Reading, MA, 1981.
[20] G.A. Steiner, The systems approach in planning, in: D.W. Ewing (Ed.), Long-Range Planning for Management, 3rd ed., Harper
and Row, New York, 1972.
[21] C. Vercellis, Constrained multi-Project planning problems: a Lagrangean decomposition approach, European Journal of
Operational Research 78 (1994) 267±275.
[22] J.D. Weist, F.K. Levy, A Management Guide to PERT/CPM with GERT/PDM/DCPM and other Networks, 2nd ed., Prentice-
Hall, Englewood Clis, NJ, 1977.

Optimization Analysis For Design and Planning of Multi-Project Programs

Transféré par

Informations du document

Description originale:

Titre original

Copyright

Formats disponibles

Partager ce document

Partager ou intégrer le document

Options de partage

Avez-vous trouvé ce document utile ?

Ce contenu est-il inapproprié ?

Droits d'auteur :

Formats disponibles

Optimization Analysis For Design and Planning of Multi-Project Programs

Transféré par

Droits d'auteur :

Formats disponibles

European Journal of Operational Research 107 (1998) 492±506