Ann Oper Res (2007) 155: 361–390

DOI 10.1007/s10479-007-0215-z

An integrated staff-sizing approach considering

feasibility of scheduling decision

Yongjian Li · Jian Chen · Xiaoqiang Cai

Published online: 13 July 2007

© Springer Science+Business Media, LLC 2007

Abstract This paper presents an integrated staff-sizing system for analyzing and determin-
ing workforce management policies with consideration of staff flexibility in service organi-
zations, which addresses and captures the integrated requirements between long-term man-
power planning and short-term staff scheduling in the service sector. Multiple Objective Lin-
ear Programming (MOLP) is applied to optimize several diversified goals. Solution methods
to the MOLP models for the staff planning and staff scheduling are developed respectively,
then a solution approach is proposed to iteratively revise the unacceptable staff-sizing plan
or scheduling plan. Finally, an example of nurse sizing is analyzed and computational stud-
ies are carried out to investigate managerial insights.

Keywords Staff-sizing · Planning · Scheduling · Staff flexibility · Multiple objective linear

programming (MOLP)

Y. Li ()
Business School, Nankai University, Tianjin 300071, People’s Republic of China
e-mail: [email protected]

J. Chen
College of Economics and Management, Tsinghua University, Beijing 100084, People’s Republic
of China
e-mail: [email protected]

X. Cai
Department of System Engineering & Engineering Management, The Chinese University of Hong
Kong, Shatin N.T., Hong Kong
e-mail: [email protected]

X. Cai
College of Information Technology Science, Nankai University, Tianjin 300071, People’s Republic of
China
362 Ann Oper Res (2007) 155: 361–390

1 Introduction

We propose an integrated approach to analyze and determine the manpower planning and
scheduling decisions for service organizations. As it is well known, it is both important
and difficult to allocate available staff to meet the service demands efficiently, particularly
when the demands from customers (e.g., in hospitals, bank, post office, et al.) cannot be
backlogged and must be satisfied promptly. This staffing problem is different from the work-
force and production-planning problem found in production organizations (see, for example,
Taubert 1968; Samuel and Reutzel 1981; Gen et al. 1992; Cheng et al. 2001). Production
models rely upon the inventory and backorder capabilities. Furthermore, they typically ex-
clude other considerations that must be included in an effective staff planning and scheduling
models for service-sector applications. For example, they normally do not consider explic-
itly matching specific labor skills with job requirements at individual work centers. Yet this
careful matching is important in organizations like hospitals, schools and banks, where the
‘quality’ of the service is crucial, personnel are highly specialized, and post hoc quality
adjustment is not possible.
Two levels of staff planning decisions usually occur in the process of personnel manage-
ment. One is manpower planning, which determines the number of employees needed for the
planning period within budgetary constraints. The other is staff assignment and scheduling,
which is conducted weekly or monthly within the boundary of the number of staff defined
at the planning level. It has been argued that significant overstaffing often occurs in service
organizations as a direct result of separate and independent decisions between long-term
manpower planning and short-term staff assignment and scheduling. For example, in the
nurse workforce problem, the planned staff capacity for each ward in some hospitals is held
constant for the entire year at a level sufficient to meet the peak demand expected during the
year, which may result in substantial under-utilization of personnel.
There are some published works describing the use of optimization techniques to tackle
the manpower planning (staff-sizing) problem. Some of the previous results have been re-
viewed in (Bowey 1977; Price et al. 1980; Purkiss 1981; Edwards 1983; Ernst et al. 2004).
The models for manpower planning can be classified into two categories (Purkiss 1981). The
first is exploratory models, which can give the managers insights into the way his/her man-
power system works and the way it would respond to different stimuli. These exploratory
models range from the very simple ones that are applicable to almost every organization
(though often dealing with one special feature such as career progression) to very compre-
hensive stochastic simulation models for examining individual movement in the manpower
system (see for example, Zanakis and Maret 1981; McClean 1991; Georgiou and Vassiliou
1997). The second is the very powerful normative models, which can compute an optimal
set of personnel decisions (on recruitment, promotion, training, etc.) against goals stated in
some forms of objective function. A number of early attempts using linear programming
have been described in Smith (1971). Within this general schema, practices vary. For in-
stance, one important distinction is the way that movement between jobs is modelled (see
for example, Price et al. 1980; Silverman et al. 1988). One class of problems is recruitment
(see for example, Bres et al. 1980; Rao 1990; Gans et al. 2003; Cai et al. 2004; Li et al. 2005)
with the selection of a recruitment schedule over multiple time periods, which best meets the
goal pertaining to promotion opportunity, salary expenditure, desired levels of experience in
the workforce and requirements for manpower in each planning period.
The staff scheduling problem has received more attention than staff-sizing; The latest
annotated bibliography of personnel scheduling and rostering can be found in (Alfares 2004;
Ernst et al. 2004). The studies in the staff scheduling problem have generally assumed that
Ann Oper Res (2007) 155: 361–390 363

the workforce size and composition are predetermined and the staff-scheduling problem is
defined in a very limited planning horizon (e.g., one week or one month). These assumptions
ignore, however, the need to consider and address the changes in workforce size to respond
to the dynamic fluctuations of the staff demands (Henderson et al. 1982).
While a significant amount of research has focused upon the staff-sizing and staff
scheduling decisions individually, very little has considered the interaction between the two
decisions. Perhaps this is due to the fact that staff-sizing decisions are often made by the
manager at higher levels in an organization than scheduling decisions, and in a much longer
planning time horizon than staff scheduling. Despite the difference in time horizon and de-
cision level, staff-sizing and scheduling are clearly interdependent and should be examined
as an integrated system.
A common problem associated with the integration of staff-sizing and scheduling deci-
sions is that strict enforcement of aggregate staff-sizing decisions may lead to infeasibility
of the scheduling policies (Venkataraman and Brusco 1996). This suggests the need for a
recursive approach that enables the impact of the staff planning decisions made at the staff-
sizing phase to be rapidly evaluated in a scheduling context. Abernathy et al. (1973) have
described a recursive approach that begins with a simulation model for generating staffing
parameters for nursing units. A chance-constrained staffing model was subsequently used
to actually set the staffing levels. Venkataraman and Brusco (1996) have presented an inte-
grated nurse staffing and scheduling system, where the staffing and scheduling models were
described as a mixed-integer programs, respectively.
In the studies reviewed above, only a single objective concerning the cost incurred dur-
ing the entire planning horizon was dealt with. However, service organizations are concerned
generally about two types of objectives: service quality and cost. Service quality is repre-
sented through many indices, e.g., minimizing the quantity of employees augmented, min-
imizing the overtime incurred, ensuring the professional development of every employees,
etc. So multi-objective planning problem needs to be considered to reflect the goals of ser-
vice organizations. On the other hand, only one type of staff has been considered in the above
references. Generally, there are, however, several types of staff existing in an organization,
among which substitution may arise so as to reduce the manpower-related expenditure. So
staff flexibility needs also to be considered.
The main purpose of this article is to develop an integrated approach for analyzing
and determining workforce management policies in service organizations with consider-
ation of staff flexibility and feasibility of the corresponding scheduling decisions. It ad-
dresses and captures the integrated requirements between long-term staff-sizing and short-
term staff scheduling in the service sector. Multiple Objective Linear Programming (MOLP)
approaches are applied to optimize several diversified goals in the staff-planning and the
staff-scheduling problems respectively, which handle the conflicting objectives of costs and
service levels. Our MOLP approaches examine the effects of staff flexibility on staff-sizing
decisions at the aggregate and disaggregate levels respectively. An iterative procedure is
proposed to iteratively improve the solutions generated by the MOLP approaches for the
planning and scheduling problems.
The organization of this paper is as follows. In Sect. 2, we describe the general staff
planning and scheduling system. In Sects. 3 and 4, we develop MOLP models to formulate
the staff planning and scheduling problems, respectively. In Sect. 5, we propose our solu-
tion approach to compute the solutions for the planning and scheduling models. A recursive
procedure is also developed to iteratively improve the planning and scheduling solutions.
Numerical analysis is provided in Sect. 6. Section 7 gives a summary and some future re-
search topics.
364 Ann Oper Res (2007) 155: 361–390

Fig. 1 An integrated staff decision model

2 Integrated staffing decision model

The general framework of an integrated staffing decision system is illustrated in Fig.1. We

divide it into three stages (modules): demand forecasting, staff planning and staff schedul-
ing.
In stage I, there are, in general, two kinds of manpower demands encountered by a ser-
vice organization. One is the known demand—customers who make appointments with the
service provider in advance. The other is the unknown demand—customers who walk in
without appointments. In a real situation, demand uncertainty is inevitable due to causes
such as customers who come without appointments or customers who do not show up af-
ter making appointments. The outputs in stage I are usually aggregate forecasts of monthly
demand, which are the input to the staff planning model in stage II. These forecasts can
subsequently be disaggregated to generate shorter-term forecasts for the staff scheduling
problem in stage III.
Stage II concerns staff planning at the macro level, which generates the desirable size
of the workforce for each skill class across the entire planning horizon. In addition to the
forecast demands, the inputs to the staff planning model also include the staff planning
requirements. Planning requirements are specified in terms of labor hours required in each
period, as well as overtime policies, substitution policies, and training policies, etc., which
will be presented in details in Sect. 3. The outputs of the staff planning model are the number
of employees, the number of those being recruited or dismissed, the substitution number of
higher-level staff to lower-level one and the amount of overtime hours during each period.
The outputs from the staff planning model comprise the inputs to staff scheduling in stage
III. Specifically, the information extracted from the staff plan is the number of employees
who are assigned to every demand type. In addition, the inputs to staff scheduling also
include staff scheduling requirements, e.g., overtime policy, temporary employment policy
and day-off policy, etc.
The existence of an effective coordination between the different subproblems (stages) is
essential to the success of the integrated system. In this paper, we will mainly address the
planning and scheduling problems, under the assumption that the manpower demands have
been given by the forecast module. We will formulate the planning and scheduling models
as multiple objective linear programs in the following two sections, respectively.
Ann Oper Res (2007) 155: 361–390 365

3 Staff planning model

The typical questions that will be asked in a service organization include: Do we have
enough qualified staff to serve the customers? Is the system flexible enough to meet de-
mand fluctuation? We will study these problems using a multiple objectives staff planning
model, which considers a number of linear objectives and linear constraints, including staff
task flexibility, staff development requirements, staff level change constraints, etc.

3.1 Notation

The variables and parameters used in the staff planning model are listed in the following.

Parameters

T The number of planning periods under consideration. The length of each period may
be a month or a year, depending on the application.
n The number of job types and categories of full-time staff.
Dit1 The projected regular demand in staff hours for type-i job from customers who have
made appointments beforehand during the period t .
Dit2 The projected irregular demand in staff hours for type-i job from customers who
have made no appointments beforehand during the period t .
Hi The number of regular hours available of a category-i employee each period, ex-
cluding the personnel time and nonbillable duties.
fi The maximal proportion of type-i jobs that can be performed by the employees at
higher categories.
oi The maximum allowable overtime for a category-i employee as a proportion of
regular hours.
αi The base salary per category-i employee.
ei The overtime cost per hour per category-i employee.
βi+ /βi− The recruitment/dismissal costs per category-i employee.
sj i = 1, if category-i staff can be substituted by category-j staff (i ≤ j ); = 0, otherwise.
Especially, sii = 1 means that the category-i staff can be assigned to do their own
job.
Qi The target professional development hours over the planning horizon for category-i
staff.
Si− The shortfall in the achievement of the professional development target.

Variables

Xit The number of category-i employees to be available during period t . Xi0 is the initial
value.
uit /vit The number of category-i employees to be recruited/dismissed at the end of period
t.
Yj it The number of category-j employees being assigned to do type-i job (category-j
employees substitute category-i ones) during period t (i ≤ j ). Especially, Yiit is the
number of category-i employees being assigned to do their own job.
Oit The overtime hours of category-i employees during period t .
Eit The actual number of category-i employees for professional development during
period t .
366 Ann Oper Res (2007) 155: 361–390

3.2 Constraints

There are n types of jobs: type-i, i = 1, 2, . . . , n, requiring pre-specified manpower re-

sources respectively during period t , t = 1, 2, . . . , T . Correspondingly, there are n categories
of full-time staff: category-i, i = 1, 2, . . . , n, available in the organization, where category-i
employees can first be assigned to a type-i job (i = 1, 2, . . . , n). Moreover, different cate-
gories of employees hold different skills, which can be assigned to a lower-level of job when
needed. For example, category-1 employees have skill 1 and can be assigned to a type-1 job,
while category-2 employees have better skills, then they can be assigned not only to a type-2
job, but also to a type-1 job (to substitute category-1 employees). In general, we assume that
category-i employees have better skills than category-j employees for i > j .
The costs for using different categories of employees are different, and usually the salary
for category-i employee is more expensive than that for category-j employee when i > j .
Constraints that must be satisfied in service staff planning include:
(1) Staff supply-demand constraints
An initial number of employees at each category is assumed. Hiring/firing decisions
should be made to ensure that the manpower supply is always greater than or equal to the
manpower demand. The manpower supply-demand constraints are given by the following
dynamic equations:

Xi(t+1) = Xit + uit − vit , for t = 0, 1, 2, . . . , T − 1, ∀i. (1)

(2) Staff utilization constraints

Each staff member should be given a certain amount of regular work time each period,
which should include allowances for personnel times (e.g., illness) and for nonbillable duties
(e.g., administration, recruiting, practice development). In our problem, we have supposed
that jobs can be done not only by staff of the scheduled category but also by some higher
categories; other substitutions are not allowed. Over time may also be required to meet the
demands.

sj i Hj Yj it + Oit ≥ Dit for ∀i; t, (2)
j ≥i

where Dit = +
Dit1 Dit2
is the total demand of type-i job during period t that is the sum
of the projected regular-demands and irregular-demands, where the irregular-demands are
to be used to model the irregular manpower requirements due to public holidays, off-days,
extra manpower requirements that may be introduced by multiple shifts, etc.
(3) Staff gross constraints
The number of category-i employees that can be assigned to every type-j job (j ≤ i) is
limited to the total available number of employees at category-i, excluding those employees
attending professional development.

Yij t + Eit = Xit for ∀i; t. (3)
j ≤i

(4) Staff task flexibility constraints

There is a limit on task flexibility and a limit on the amount of type-i job that can be
performed by substitution.

Hi Yiit ≥ (1 − fi )Dit , for ∀i; t. (4)

Ann Oper Res (2007) 155: 361–390 367

If the task substitution level were too high, the management would recruit some number
of category-i employees for the type-i job instead of using substitution.
(5) Overtime constraints
The amount of overtime hours an employee can take has an upper limit.

0 ≤ Oit ≤ oi Hi Xit , for ∀i; t. (5)

(6) Professional development constraints

Some employment time must be dedicated to professional development and similar ac-
tivities, e.g., on-the-job training, seminars, conferences, etc. These requirements may vary
by staff category.
 
Eit Hi + Si− = Qi Xit , ∀i. (6)
t t

Usually, an organization may have some targets for professional development. Meeting
professional development targets may be formulated as an objective to be optimized.
If the professional development for category-i staff is constrained in one or several given
periods, then constraints are similar to (6) above.

3.3 Objectives

Various objectives may be established to reflect the needs and goals of the organization.
Generally, service organizations are concerned about two types of objectives: service quality
and cost. For example, in the health service sector, the quality of service largely relies on
proper assignment of the professional staff with the designated knowledge and skills, while
the cost of service is heavily depending on the staff-sizing solution because 60 to 80%
of budget of health service is occupied by staff cost (McConnell 2000). We consider the
following five objectives:
(1) Staffing cost
n 
 T T −1
 
+ −
Min z1 = (αi Xit + ei Oit ) + (βi uit + βi vit ) . (7)
i=1 t=1 t=0

This goal seeks to minimize the total cost arising from acquiring /maintaining the neces-
sary level of staff supply to meet the demand.
(2) Staff augmentation
 T −1
n 
Min z2 = uit . (8)
i=1 t=0

The objective seeks to minimize the number of new employees recruited during the plan-
ning horizon. This criterion is particularly important for service organizations, because new
employees are generally perceived as lack of the needed experience to provide high-level
service.
(3) Staff task substitution

 T 
n  
Min z3 = Yj it . (9)
i=1 t=1 i<j
368 Ann Oper Res (2007) 155: 361–390

The objective seeks to minimize the number of employees doing lower-level jobs. While
substitution is an effective means to tackle the problem of temporary and unpredictable
staff shortage in certain categories, it is generally regarded as an expensive solution because
higher-level employees inevitably carry higher costs.
(4) Overtime incurred

n 
T
Min z4 = Oit . (10)
i=1 t=1

Overtime is undesirable because prolonged working time could cause fatigue and de-
ficiency in service quality. In addition, there are also regulatory limits on the amount of
overtime that an employee can be assigned to take.
(5) Shortfall of professional development


n
Min z5 = Si− . (11)
i=1

The objective seeks to minimize the unachieved professional development time during
the planning horizon.

4 Staff scheduling model

Staff scheduling is usually conducted for a short period, e.g., one week or one month. An
employee works on a shift on a workday, and can receive F off-days during the scheduling
period. Labor demands are to be satisfied by regular time, overtime and temporary staff. In
general, overtime is priced more than the wage rate for a same staff category and temporary
staff (e.g., supplemental staff contracted through an outside agency) is assumed to be a last
resort and is generally priced more than the corresponding overtime cost.

4.1 Notation

The variables and parameters used in the staff scheduling model are listed below.

Parameters

S The number of days during a scheduling period under consideration.

dit The number of working hours of type-i job required on day t .
hi The number of regular hours per category-i employee per day.
wi The number of category-i employees who can be scheduled during the scheduling pe-
riod. For example, assume the scheduling plan is made in a period s, then wi = Xis −Eis
is the difference between states Xis and Eis obtained from the staff-planning model.
dit+ The number of surplus category-i employees, who are unassigned to do any job, and
not on leave, on day t .
δi The maximum allowable rate of overtime for category-i staff as a proportion of regular
hours during a period.
F The number of off-days that an employee can receive during a scheduling period.
cio Hourly cost of overtime of category-i staff.
cia Salary of a category-i temporary employee per day.
Ann Oper Res (2007) 155: 361–390 369

Variables

x̄j it The number of category-j employees who are assigned to do type-i job on day t .
Tit The amount of overtime hours provided on day t by the category-i staff.
Ait The number of temporary employees to be assigned to the type-i job on day t .
yit The number of category-i employees who take day t off.

4.2 Constraints

The constraints in the scheduling stage can be classified into two categories (Miller et al.
1976): (i) Feasibility set constraints, or hard constraints (e.g., demand constraint, overtime
constraint, day-off constraint, etc.), which define the sets of feasible staff schedules that must
be satisfied; (ii) Nonbinding constraints, or soft constraints (e.g. staff preferences, preferen-
tial scheduling patterns, etc.), whose violation incurs certain degree of un-satisfaction of the
staff concerned, but the solution is still usable by the organization. In our system, we only
take into account the feasibility set constraints, and we suggest that the soft constraints be
handled by managers using their experience and judgment, based on the actual situations
and conditions.
(1) Demand constraints

sj i x̄j it hj + Tit + Ait hi ≥ dit , ∀i, ∀t. (12)
j ≥i

The requirements of type-i job at any time must be guaranteed and can be satisfied by
category-i employees, category-j employees that satisfies sj i = 1, overtime of category-i
employees and temporary staff, with consideration of day-off for the full-time staff.
The number of full-time employees who are assigned each day is constrained by the
number of employees during this scheduling period obtained from the staff planning model.
This is the following constraint.

sj i x̄j it + yj t + dj+t = wj , ∀j, ∀t. (13)
i≤j

(2) Overtime constraints

0 ≤ Tit ≤ δi hi wi , ∀i, ∀t. (14)

It limits the use of overtime on each day during the scheduling period. This limit could
create infeasibility if only regular time and overtime hours are permitted. To avoid potential
infeasibility, temporary staff may have to be used to satisfy any excess demands. The num-
ber, Ait , of the temporary staff included in constraint (12) actually reflects this requirement.
(3) Off-day constraints

yit ≥ F wi , ∀i. (15)
t

Each full-time employee will receive at least F off-days during a scheduling period. We
assume that overtime comes from extending the regular working time. Thus, overtime does
not affect the day-off requirement. However, if overtime comes from working on a day-off,
this constraint should be modified accordingly.
370 Ann Oper Res (2007) 155: 361–390

Remarks (1) Note that the scheduling model formulated here will determine the number of
regular and temporary staff members assigned to each type of job on each day, the amount
of overtime for each category of staff on each day given the workforce constraint imposed
by the planning model, and also the day-off for each category of staff. It has not, however,
addressed the issue of staff assignment down to the level of shifts on each day. After the
daily assignment is determined that can meet the daily manpower demand, one may further
divide the staff into shifts, based on the manpower demands and other practical constraints
on the shifts. Note also that, when multiple shifts exist, the required number of employees
is often increased. To address this issue, certain tolerance of manpower should be added to
the manpower demand. In our model, this can be incorporated into the manpower demand
Dit2 ; see the definition below constraint (2).
(2) The solution that satisfies the hard constraints will be regarded as a feasible one.
As we indicated above, there are also nonbinding constraints that exist in practice, which in-
clude certain desirable work patterns, and personnel preferences (e.g., weekend-off requests,
minimum workstretch, etc., cf. Billionnet 1999; Uebe et al. 1990; Topaloglu and Ozkarahan
2004). It is not compulsory for an organization to satisfy every soft constraint. Instead, soft
constraints are generally handled by management at the relevant levels, according to the
situations and urgency/importance of the requests.

4.3 Objectives

We consider the following scheduling objectives: The objective of minimizing overtime and
temporary staff costs subject to all scheduling constraints; and the objective of maximizing
the surplus staff, in order to reduce the risk of underestimation of actual demands that may
occur due to inaccurate forecast; see Cai and Li (2000).
(1) Minimization of overtime and temporary staff cost

Min J1 = (cio Tit + cia Ait ). (16)
i t

(2) Maximization of surplus staff


Max J2 = dit+ . (17)
i t

5 Solution techniques

5.1 Solving the staff planning model

We first convert the MOLP model as given in Sect. 3 into one with a single objective, and
then determine the solution for the single-objective problem. The solution for the single-
objective problem can be found by either a standard linear integer programming algorithm,
or a specific algorithm (cf. Cai et al. 2004; Li et al. 2005). See Appendix 2 for more details.
Overall, our solution procedure is described as follows.
Step 1. Set weights λi (i = 1, 2, . . . , 5) for each objective according to their relative im-
portance, by an Analytic Hierarchy Process (AHP) (see Appendix 5).
Ann Oper Res (2007) 155: 361–390 371

Step 2. Normalize the objectives as follows:

zi − zimin
zi = , (18)
zi − zimin
max

where zimax and zimin are, respectively, the maximum and minimum values of the i-th ob-
jective zi when we consider only this objective in the staff planning model. In Appendix 1
we give simple results to compute zimax and zimin , which avoid the requirement to solve an
optimization problem in order to obtain zimax or zimin .
Step 3. Solve the following single objective problem
5
IntP1: min J= i=1 λi zi
(19)
s.t. Constraints (1–6).

The solution approach to the model IntP1 is given in Appendix 2.

5.2 Solving the staff scheduling model

We also use a three-step approach to solve the staff scheduling problem. The procedure can
be described as follows.
Step 1. Set weights λi (i = 1, 2, . . . , 5) for each objective according to their relative im-
portance, by an AHP approach.
Step 2. Normalize the objectives as follows:

J1 − J1min
J1 = , (20)
J1max − J1min
J2max − J2
J2 = , (21)
J2max − J2min

where Jimax and Jimin are, respectively, the maximum and minimum values when we consider
only the single objective Ji in the staff scheduling model. In particular,
 when we consider
only the second objective J2 , we can show easily that J2max = ni=1 wi and J2min = 0. The
computations for J1max and J1min are given in Appendix 3.
Step 3. Solve the following single-objective problem:

IntS1 : min J = λ1 J 1 + λ2 J 2
n  
n  
= λ1 cito Tit + cita hi Ait ) − λ2 yik − C (22)
i=1 t i=1 k∈i

s.t. Constraints (12–14),

λ1 J1min λ2 J2max
where λi = λi
Jimax −Jimin
, i = 1, 2 and C = max
J1 −J1min
− max
J2 −J2min
is a constant.
The solution approach to the model IntS1 can be found by either a standard linear mixed
integer programming algorithm, or a specific algorithm; See Appendix 4 for more details.
372 Ann Oper Res (2007) 155: 361–390

Fig. 2 The solution procedure of

the integrated system

5.3 Solution procedure for the integrated model

5.3.1 The general solution framework

The general framework of the solution procedure we propose is illustrated in Fig. 2.

Basically, the solution procedure contains two stages, corresponding to the planning so-
lution and the scheduling solution respectively. The two stages are, however, interrelated in
our overall solution procedure; i.e., the solution of one stage may be fed back to revise the so-
lution of the other stage. We have described, in the two subsections above, how to determine
the planning and scheduling solutions. The remaining task now is how to revise (improve)
the solutions based on the feedbacks. We propose to adjust those key weights/factors, within
their practically allowable ranges, in the planning/scheduling models so as to achieve the
desirable effects of solution improvements.
Table 1 lists the weights/factors that are to be adjusted in the solution procedure. Ta-
ble 2 lists the revision actions we propose to address the relevant unacceptable plan-
ning/scheduling solutions. On the one hand, the planning solution (the number of staff at
each category) will be direct input of the scheduling model to affect the scheduling solution.
On the other hand, the scheduling solution will be checked and unacceptable solution will
be fed back to the planning model so that the planning solution will be adjusted accordingly.
Ann Oper Res (2007) 155: 361–390 373

Table 1 Adjustable factors in the iterative solution procedure

Factor Description

oi Overtime coefficient in the planning stage

δi Overtime coefficient in the scheduling stage
F Off-days factor
fi Substitution factor
Qi Professional development factor
OPi The weight of the i-th objective in the staff planning model
OSi The weight of the i-th objective in the staff scheduling model

Table 2 Adjustment schemes to revise/improve unsatisfactory solutions

Information feedback Adjustment

action

The number of overtime hours is too large. Specifically, there ex- OP4 ↑
 
ists an i such that Tt=1 Oit /(hi Tt=1 Xit ) ≥ IF 1 , where IF 1 is the oi ↓
maximal ratio of total overtime allowed for category-i staff with
respective to the total normal working time.
The number of new employees being recruited is too large. Specif- OP2 ↑

ically, there exists an i such that Tt=1 uit ≥ IF 2 , where IF 2 is the oi ↑, fi ↑
maximal number of new category-i employees to be recruited.
Staff There are too many staff task substitutions. Specifically, there ex- OP3 ↑
  
planning ists an i such that Tt=1 i<j Yj it/ Tt=1 Xit ≥ IF 3 , where IF 3 is the fi ↓
stage maximal ratio of category-i employees who are assigned to take
other jobs with respective to the total number of category-i em-
ployees.
Unachieved professional development time is too high. Specifi- OP5 ↑

cally, there exists an i such that Si−/(Qi Tt=1 Xit ) ≥ IF 4 , where Qi ↑
IF 4 is the maximal ratio of the unachieved professional develop-
ment time with respective to the expected development time for
category-i staff.
There are too many temporary staff. Specifically, there exists an i OS1 ↑, OP2 ↓,
 
such that Tt=1 Ait / Tt=1 Xit ≥ IF 5 , where IF 5 is the maximal ra- OP4 ↑, δi ↑, F ↓,
tio of temporary category-i employees with respective to the total oi ↑, fi ↓
number of regular category-i employees.
Staff There are too many overtime hours. Specifically, there exists an i OS1 ↑, OP2 ↓,

schedul- such that St=1 Tit /(Shi wi ) ≥ IF 6 , where IF 6 is the maximal ratio of OP4 ↑, δi ↓, F ↓,
ing stage overtime of category-i staff with respective to their normal working oi ↓, fi ↓
time.

a ↑ (a ↓) means that the value of a should be increased (decreased); IF i (i = 1, 2, . . . , 6) is to be determined

according to the situation under consideration

For example, if the scheduling stage generates too many temporary staff, then this im-
plies that the size of the workforce determined by the planning stage is not sufficient (and
consequently, temporary staff must be sought to meet the manpower demands). The follow-
ing adjustments can be made. First, we may decrease, if possible, the weight OP2 for the
objective regarding staff augmentation. This is because if the objective of minimizing the
staff augmentation is over emphasized, the quantity of new employees being recruited will
374 Ann Oper Res (2007) 155: 361–390

be reduced and temporary staff may have to be sought in the scheduling stage to meet the
unexpected fluctuation of the demands.
Secondly, we may increase the weight OP4 for the overtime objective or decrease the
value oi of the corresponding category-i staff. This will discourage the use of overtime at
the planning level and consequently increase the number of staff to be recruited.
Thirdly, we may decrease the substitution factor fi of the corresponding category-i staff
to restrict the flexibility of staff substitution. This will also increase the number of staff to
be recruited at the planning level.
We propose to adopt the following methods to implement Table 2 for the adjustments of
weights: OSi and OPi , and the parameters: oi , δi , F , fi , and Qi .

5.3.2 Adjustment of the weights OSi and OPi

Recall that we use the AHP approach to set the weights of the objectives in the planning
and the scheduling phases. Now, when these weights are to be adjusted, we will modify,
accordingly, the corresponding components aij in the pairwise comparison matrix of the
AHP approach (see Appendix 5). Specifically, if the adjustment action is OPi ↑, then we
will reset aij as follows:
For i < j ≤ 5,

min{aij + 1, 9}, if aij ≥ 1,
aij = aij (23)
1−aij
, if aij < 1,

and for 1 ≤ j < i,


aj i − 1, if aj i > 1,
aj i = a (24)
max{ 1+aj ij i , 1
9
}, if aj i ≤ 1.

If the adjustment action is OPi ↓, then we will reset aij as follows:

For i < j ≤ 5,

aij − 1, if aij > 1,
aij = aij (25)
max{ 1+aij , 9 }, if aij ≤ 1,
1

and for 1 ≤ j < i,


max{aj i + 1, 1}, if aj i ≥ 1,
aj i = a (26)
max{ 1−aj ij i , 19 }, if aj i < 1.

If the adjustment action is OS1 ↑, then there is only one component ai2 to modify (see
Appendix 5: AHP):

min{ai2 + 1, 9}, if ai2 ≥ 1,
ai2 = ai2 (27)
1−ai2
, if ai2 < 1.

5.3.3 Adjustment of the parameters oi , δi , F , fi , and Qi

To adjust the parameters oi , δi , F , fi , and Qi , we adopt the following heuristic idea.

Suppose p is the parameter to be adjusted and suppose its current value is p k . Let p min
and p max be its lower and upper bounds. Further, we let p = (p max − p min )/γ , where γ > 0
Ann Oper Res (2007) 155: 361–390 375

is a scalar (e.g., γ = 10). Then, the new value for p will be computed as follows:
⎧
⎪
⎨p ,
min
if p k ± p ≤ p min ,
p k+1
= p , max
if p k ± p ≥ p max , (28)
⎪
⎩ k
p ± p, otherwise,

where + or − will be used when the parameter is to be increased or decreased according to

Table 2, respectively.

5.3.4 The algorithm

To summarize, our procedure is now described as follows.

Step 1. Compute the staff planning solution according to the procedure in Sect. 5.1.
Step 2. If the staffing plan is acceptable, then go to Step 3 directly. Otherwise,
Step 2.1 If the amount of overtime is too large, then return to Step 1, and increase
the weight OP4 , or decrease the factor oi of the corresponding staff cate-
gory.
Step 2.2 If the number of new employees being recruited is too large, then return to
Step 1, and increase the weight OP2 , or increase the factor oi , or increase
the factor fi of the corresponding staff category.
Step 2.3 If there are too many staff task substitutions, then return to Step 1, and in-
crease the weight OP3 , or decrease fi of the corresponding staff category.
Step 2.4 If the unachieved professional development is unacceptable, then return to
Step 1, and increase the weight OP5 , or increase Qi of the corresponding
staff category.
Step 3. Do loop from Step 4 to 5 with respect to each time period t = 1, 2, . . . , T .
Step 4. Compute the staff scheduling solution for the period t according to the procedure in
Sect. 5.2.
Step 5. If the staff scheduling plan is acceptable, then go to Step 3 with t = t + 1. Other-
wise,
Step 5.1 If there are too many temporary staff, then return to Step 1, and increase the
weight OP4 , or decrease the weight OP2 , or increase the factor oi and/or
decrease fi of the corresponding staff category;
Step 5.2 If there are too much overtime hours, then return to Step 1, and increase the
weight OP4 , or decrease the weight OP2 , or decrease the factor oi and/or
fi of the corresponding staff category.
Step 6. An acceptable, integrated planning and scheduling solution is obtained.

6 Numerical results

6.1 An example

We consider the case with two categories of full-time (FT) nurses (category-1 and category-
2) and two types of jobs (type-1 and type-2). The category-1 nurses have skill 1 (patient care
assistants) and can be assigned to a type-1 job (common care), while the category-2 nurses
have a higher level of skill set, and can be assigned not only to type-2 job (critical care),
but also to type-1 job. The sample data are partly extracted from the example of Venkatara-
man and Brusco (1996). The planning horizon comprises six planning periods, each with
376 Ann Oper Res (2007) 155: 361–390

Table 3 Data for the nursing problem

Four-week period (t) 1 2 3 4 5 6

Demands for Job-1 labor hours (D1t ) 4266 4221 4972 4310 4507 4442
Demands for Job-2 labor hours (D2t ) 2145 2536 2141 2865 2206 2252
Maximum ratio of overtime to regular time oi = 0.2, i = 1, 2
Maximum ratio of task flexibility f1 = 0.3, f2 = 0
Hours of professional development Q1 = 5, Q2 = 10
Initial number of employees X1 (0) = 20, X2 (0)=10
Cost coefficients α1 = $10/hour × 8hours/day × 28days = $2240
α2 = $12/hour × 8hours/day × 28days = $2688
β1+ = α1 /2 = $1120, β1− = α1 = $2240
β2+ = α2 /2 = $1344, β1 = α2 = $2688
e1 = $10/hour × 3/2 = $15, e2 = $12/hour × 3/2 = $18
H1 = H2 = 8hours/day × 20days = 160hours

4 weeks. The FT nurses are assumed to work 160 regular-time hours and have 8 off-days
per 4-week period. It is assumed that nurses work an 8-hour shift on each workday of their
schedule. Staff demands are to be satisfied from regular-time, overtime and temporary staff
(supplemental staff contracted through an outside nursing agency). The hourly wage rates
are set at $12 and $10 for FT category-2 and category-1 nurses, respectively. Overtime is
priced at one-and-a-half times the corresponding FT wage rate. Temporary staff is assumed
to be a last resort and is priced at twice the corresponding FT wage rate. Table 3 lists the
nurse staffing requirements for a 6-month planning horizon.
Each planning period (4 weeks) gives a scheduling horizon of 28 days. The mean daily
demand (in terms of working hours required) of category-i nurses is ai = Dit /28. The
sample demands dis on each day s in the scheduling period t are generated through a
Gaussian Distribution N (ai , 12). The maximum ratio of overtime to regular time on any
day is δ1 = δ2 = 0.25 respectively.
Our iterative procedure in Sect. 5.3 was implemented in Matlab 6.5 on a PC with 1.8 G
CPU. The staff planning model and staff scheduling model were solved by the Lingo 8.0
software package, respectively. The computation time required by the Lingo Tools to solve
the planning model was less than 1 second, and that to solve the scheduling model was less
than 20 seconds, and the computation time consumed in Matlab 6.5 was less than 1 second.
For the staff planning model, the relative importance of the objectives were set by a
pairwise comparison matrix as follows:
⎡1 5 4 4 5⎤
⎢1 1 3 2 2⎥
⎢5 ⎥
⎢ 2⎥
A = ⎢ 14 1
1 2 ⎥,
⎢1 3
⎥
⎣ 1 1
1 2⎦
4 2 2
1 1 1 1
5 2 2 2
1

where the element aij indicates the relative importance of objective i as compared to objec-
tive j ; cf. Table 22, Chapt. 14, Winston (1994).
Applying the Analytic Hierarchy Process, we generated a weight vector w =
[0.4981, 0.1904, 0.1331, 0.1054, 0.0730] that satisfies the consistency requirement.
Ann Oper Res (2007) 155: 361–390 377

Table 4 Solution obtained

Four-week period (t) 0 1 2 3 4 5 6

No. of Category-1 staffs 20 24 24 24 24 24 24

No. of Category-1 staffs to be 0 24 24 24 24 24 24
assigned to take Type-1 job
No. of Category-1 staffs to be hired 4 0 0 0 0 0 0
No. of Category-1 staffs to be fired 0 0 0 0 0 0 0
No. of Category-2 staffs 10 15 15 15 15 15 15
No. of Category-2 staffs to be 0 15 15 13 15 15 15
assigned to take Type-2 job
No. of Category-2 staffs to be hired 5 0 0 0 0 0 0
No. of Category-2 staffs to be fired 0 0 0 0 0 0 0
Overtime hours of Category-1 staffs 0 426 381 812 470 667 602
Overtime hours of Category-2 staffs 0 462 453 378 182 423 469
Professional development 0 0 0 0 0 0 0
hours of Category-1 staffs
Professional development 0 0 0 0 0 0 0
hours of Category-2 staffs
No. of Category-2 staffs to be 0 0 0 2 0 0 0
assigned to Type-1 job
Shortfall of professional development hours of Category-1 employees 2160
Shortfall of professional development hours of Category-2 employees 1800
Cost ($) Category-1 nurses Category-2 nurses
Recruitment cost: 4928 7280
Salary 354816 262080
Overtime 55407 46156
Totalcost: 730667

Table 5 Parameters for the iterative solution procedure

IF 1 IF 2 IF 3 IF 4 IF 5 IF 6
0.15 10 0.1 0.1 0.1 0.2

[oimin , oimax ] [δimin , δimax ] [f2min , f2max ] [Qmin max

1 , Q1 ] [Qmin max
2 , Q2 ] [F min , F max ]
[0.15, 0.25] [0.2, 0.3] [0.25, 0.35] [4, 6] [8, 12] [6, 10]
r1 r2 r3 r4 r5 r6
10 10 5 8 8 4
ri is the scalar corresponding to the parameter in the second row

With the input data in Table 3 and the weight vector w, we solved the staff planning
model by the procedure of Sect. 5.1. Table 4 reports the solution obtained. The solution in-
dicated that the professional development requirement was far from satisfaction. Moreover,
too much overtime was adopted. These implied that the solution for the staff plan should be
revised.
The iterative approach of Sect. 5.3 was then activated. Table 5 gives the values we used
for the parameters in the iterative solution approach. The progress of this procedure in solv-
378 Ann Oper Res (2007) 155: 361–390

Table 6 Iterative improvement process

Itera- Planning Scheduling Is the solution Adjustment

tions acceptable?
z1 z2 z3 z4 z5 Day J1 J2 OT TS

1 730667 9 2 5725 3920 ___ No. z5 is too Increase the

large: both weight OP5
the unachieved according to
professional Sect. 5.3.2
development of
two categories of
staff are too high
2 805803 12 1 7465 70 1 13479 0 816 16 No. There are Increase the
2 11010 0 684 13 too many weight OP4 and
3 22104 0 513 82 temporary decrease the
4 37544 0 415 77 employees in weight OP2
5 3991 0 247 11 most of according to
6 13710 0 805 18 periods Sect. 5.3.2
3 816639 15 3 2718 30 1 11483 0 730 2 No. There are Decrease the fac-
2 8803 0 865 1 too many tor o1 from 0.2
3 13589 0 877 2 overtime hours gradually, each
4 23394 1 1189 24 for category-1 time by a step-
5 2356 0 78 0 staff in most of size 0.01, to
0.16.
6 10638 0 693 0 periods
4 821271 17 3 1265 70 1 9457 0 341 17 No. There are Increase the fac-
2 9398 0 403 15 too many tor δ 2 from 0.25
3 9213 0 457 6 temporary gradually, each
4 9356 0 432 5 category-1 staff time by a step-
5 2129 1 75 2 in periods 1 and size 0.01, to 0.29
6 8925 0 237 1 2 in periods 1 and
2
5 821271 17 3 1265 70 1 9452 0 429 6 Improvement
can
2 9343 0 491 4 no longer be
3 9213 0 457 6 made by the iter-
ative solution ap-
4 9356 0 432 5 proach, so the so-
lution procedure
5 2129 1 75 2 stops and an ap-
proximate
6 8925 0 237 1 solution is ob-
tained.

ing the planning and scheduling problems, together with the relevant solutions, is reported in
Table 6, where the column ‘OT’ represents the amount of overtime required in each schedul-
ing period, and the column ‘TS’ represents the number of temporary employees being em-
Ann Oper Res (2007) 155: 361–390 379

Fig. 3 The influence of oi and

fi on the total cost

ployed in each scheduling period. After 5 iterations, a solution that is acceptable in terms of
both planning and scheduling requirements was obtained.

6.2 Managerial insigts/observations

We have observed, in our computations above, that the substitution factor fi and overtime
coefficient oi seem to have significant impacts on the planning and scheduling solutions,
in particular the overtime and the number of temporary employees. We have consequently
further examined these issues, in order to derive the necessary managerial insights, by con-
ducting computational studies as follows.
We set o1 = o2 , f1 = f2 and δ1 = δ2 = o1 + 0.05, and tested values for o1 and f1 in the
range from 0 to 0.3, respectively. Other parameters are same as those in the above example.
The iterative approach of Sect. 5.3 was applied for each combination of o1 and f1 . We
examined the effects of fi and oi on: (1) the total manpower-related cost; (2) the number of
new employees recruited; (3) the quantity of shortfall of professional development; (4) the
total overtime used; and (5) the number of temporary employees employed. The detailed
results are reported in Figs. 3–8.
We have the following observations:
(1) Increasing the staff flexibility or the maximum allowable proportion of overtime has
a positive influence on the reduction of the total cost. Moreover, the effect of changing o1
on the total cost is getting larger when f1 becomes larger, since the slope of the line for o1
increases when moving the value of f1 from 0 to 0.3. However, the influence seems to have
an upper limit, since even if one of f1 and o1 continues to increase, the effect line becomes
horizontal gradually after f1 = 0.2, as illustrated in Fig. 3.
(2) Increasing the staff flexibility or the maximum allowable proportion of overtime can
decrease the number of new employees to be recruited; see Fig. 4. In particular, values of o1
and f1 in the ranges of [0.2, 0.3] seem to have very significant effects on the reduction of
the number of new employees.
(3) Increasing the maximum allowable proportion of overtime may increase the number
of staff to be substituted, as illustrated in Fig. 5.
(4) Changing o1 or f1 has effects on the shortfall of professional development; see Fig. 6.
However, there does not seem to be a pattern detected.
380 Ann Oper Res (2007) 155: 361–390

Fig. 4 The influence of oi and

fi on the number of employees
being recruited

Fig. 5 The influence of oi and

fi on the number of employees
being substituted

(5) Figure 7 shows that increasing the value of f1 has a negative influence on the reduc-
tion of overtime. The effect appears to become more obvious when o1 becomes larger and
larger.
(6) Figure 8 shows the effects of f1 and o1 on the number of temporary employees to
be employed. First we found the effect is positive. It appears that increasing either f1 or o1
would increase the number of temporary employees that need to be hired in the scheduling
stage.
We further examined the effects of staff flexibility, using 5 sets of demand data for each
of the two scenarios: (i) Considering substitution; and (ii) Not considering substitution. Ta-
ble 7 shows the results. We can see that: (a) Total cost can be reduced when substitution is
considered; and (b) Fewer new employees need to be recruited when substitution is allowed.
However, the amount of overtime and the number of temporary staff may be increased since
the number of full time staff being employed is reduced.
Ann Oper Res (2007) 155: 361–390 381

Fig. 6 The influence of oi and

fi on the unachieved
professional development

Fig. 7 The influence of oi and

fi on the total overtime

7 Concluding remarks

In this study we have developed an integrated staffing decision system where impacts of
planning and scheduling on each other are taken into account, and features and characteris-
tics of service organizations are specifically addressed. Multiple-objective linear programs
have been adopted to model the staff planning and scheduling requirements, respectively.
Actions to iteratively improve the planning and scheduling solutions have also been sug-
gested. Extensive computational studies have also been conducted, to evaluate the effec-
tiveness of our proposed approach. Some managerial insights/observations have also been
revealed from the numerical results.
Further research includes evaluation of the system in large applications. Various decision
rules (adjustment schemes) need to be investigated further to provide the potential manage-
rial implications. Further researches may also include sensitivity analysis on the effects of
different pay scales as well as recruitment/dismissal costs on the overall cost reduction when
staff flexibility is introduced.
382 Ann Oper Res (2007) 155: 361–390

Fig. 8 The influence of oi and

fi on the number of temporary
employees

Table 7 Comparison of effects of substitution

Source Substitution No substitution t-value

Total cost $5.2148e+005 $5.4399e+005 −1251

Number of category-1 staff recruited 4 6.5 −2.0833
Number of category-2 staff recruited 4 5 −1.8333
Overtime of category-1 staff (hrs) 3100 2607 545.3
Overtime of category-2 staff (hrs) 2246 1690 753.1
Shortfall of professional development for category-1 staff (hrs) 0 0 No
Shortfall of professional development for category-2 staff (hrs) 0 0 No
Temporary staff of category-1 42 12 219.5
Temporary staff of category-2 21 12 121.5
Substitutions(hrs) 688 0 649

P = 0.05

Acknowledgements We wish to express our sincere appreciation to the two guest editors and the two ref-
erees for their constructive comments and suggestions to help improve our paper. This work was partly sup-
ported by NSFC Research Fund No. 70329001, 70501014, 60074018, Hong Kong Research Grants Council
under Earmarked Grant No. CUHK 442/05, and 985 Project of National Center on Business Management
and Institutional Innovation of China in Nankai University.

Appendix 1 Computing zimax and zimin

Generally, the values zimax and zimin (i = 1, 2, . . . , 5) could be computed respectively through
solving the problem as follows

min zi / max zi (29)

s.t. Constraints (1–6).
Ann Oper Res (2007) 155: 361–390 383

In (18), we don’t pay much attention to the exact values of zimax and zimin , but to the value
of the difference zimax − zimin . So it is enough in our computations, to get the near optimal
values of zimax and zimin through the following methods for i = 1, 2, . . . , 5.

Computing z1min and z1max

We can easily show that when no substitutions are considered, the total cost to be derived
is the upper bound of the objective z1 . Hence we set z1max the total cost of all n staff cat-
egoriesover the entire planning horizon without consideration of any substitutions. So,
z1max = 5i=1 z1i
S S
. The values z1i (i = 1, 2, . . . , 5) are computed through the following models
SMPi (i = 1, 2, . . . , 5):


T T −1

S
z1i = min (αi Xit + ei Oit ) + (βi+ uit + βi− vit ) (30)
t=1 t=0

s.t.

Xi(t+1) = Xit + uit − vit , t = 0, 1, . . . , T − 1, (31)

Hi Xit + Oit ≥ Dit , ∀t, (32)
0 ≤ Oit ≤ oi Hi Xit , ∀t. (33)

Since only the objective z1 is considered, the constraints (4) and (6) are omitted. The con-
straint (3) is reduced into Yiit = Xit , ∀i, t , so the constraint (2) is changed into the constraint
(32).
Cai et al. (2004) and Li et al. (2005) have addressed the optimization approach to the
models that are similar to SMPi .
The value z1min is computed through the following model.
SMPS:
n 
 T T −1
 
z1min = min (αi Xit + ei Oit ) + (βi+ uit + βi− vit ) (34)
i=1 t=1 t=0

s.t.
Xi(t+1) = Xit + uit − vit , ∀i, t = 0, 1, . . . , T − 1, (35)


sj i Hj Yj it + Oit ≥ Dit , ∀i, t, (36)
j ≥i

0 ≤ Oit ≤ oi Hi Xit , ∀i, t. (37)

Since only the objective z1 is considered, the constraints (3), (4) and (6) are omitted. The
model can be solved by commercial optimization software, e.g., Lingo tools.

Computing z2min and z2max

  D
We set z2max = ni=1 Tt=1 ( DHiti −  i(t−1) Hi
)+ , since we can show that the value
n T D
i=1
Dit
t=1 ( Hi −  i(t−1)
Hi
)+ is the upper bound of the objective z2 , where x is the
smallest integer that is greater than or equal to x.
384 Ann Oper Res (2007) 155: 361–390

 
We set z2min = ni=1,∃k>i,ski =1 { maxHiDi + max D̄iH−max
∗
Di
}, where max D̄i = ij =1:sij =1 Dj t
i
and Hi∗ = max{Hj , j = 1, 2, . . . , i − 1 that satisfies sij = 1}. It comes from the following
two reasons.
(1) If no substitutions are considered, then obviously when there is only a recruitment
activity and the recruitment activity occurs at the end of the initial period, the quantity of the
staff augmentation of category-i staff is minimal and the value is  maxHiDi , where max Di =
max{Dit , ∀t}.
(2) It is true that max Dij = max{Dit + Dj t , ∀t} ≤ max{Dit , ∀t} + max{Dj t , ∀t} =
max Di + max Dj . So max Dj ≥ max Dij − max Di . Then max Dij − max Dj category-i
employees being recruited during some periods are only assigned to the type-j job. Thus,
max D max Dij −max Di
we get  maxHiDi +  Hj j ≥  maxHiDi +  H∗
, where Hi∗ = max{Hi , Hj }.
i

Computing z3min and z3max

We set z3min = 0. The reason is that a feasible solution can be got even when no substitutions
are considered, and z3 ≥ 0.n
 i=1:∃j >i,sj i =1 Dit
We set z3max = Tt=1  Hj
. The reason is as follows: since we only pay at-
tention to the objective of maximizing the substitution quantity, it is optimal when all type-i
jobs are assigned to a category-j staff over the entire planning horizon for i < j and sj i = 1.

Then the quantity of jobs to be assigned to category-j is ni=1:∃j >i,sj i =1 Dit . Therefore, we
get the value of z3max .

Computing z4min and z4max

We set z4min = 0, since the lower bound of the overtime is 0.

 
We set z4max = ni=1 Tt=1 oi Hi  (1+o
Dit
i )Hi
. Since we can show that the value of z4max can be
reached only when no substitutions are considered. So, for a given type-i job, if we assume
that there are xit category-i employees during period t , to meet the demand of type-i job,
then Hi xit + oi Hi xit ≥ Dit . Thus, we get xit =  (1+o
Dit
i )Hi
and the maximum of overtime
being used for category-i staff during period t is oi Hi  (1+o
Dit
i )Hi
.

Computing z5min and z5max

Obviously, z5min = 0, since the best instance takes place when the needs of professional
development for all staff levelsT
are satisfied.
n t=1 Dit
We set z5 = i=1 Qji  Hj
max
, where ji = argmaxj {Qj |sj i = 1, j = 1, 2, . . . , n} for
i 
i = 1, 2, . . . , n. From the constraint (6), we get Si− ≤ Qi t xit , ∀i. With consideration of
substitution, it is possible that all type-i jobs, with the quantity of  HDjit , are assigned to
i
category-ji (ji > i) staff. Then these category-j  i
employees should receive the professional
T
t=1 Dit
development hours with the quantity of Qji  H ji
.
Ann Oper Res (2007) 155: 361–390 385

Appendix 2 Solving the model IntP1

zimin
Let zi = zimax − zimin , Ci = zi
and ā = a
1
, then

n 
 T T −1
 
z̄1 = (ᾱi Xit + ēi Oit ) + (β̄i+ uit + β̄i− vit ) − C1 (38)
i=1 t=1 t=0

n T −1
1
z̄2 = uit − C2 . (39)
i=1 t=0
z2
 
The objective z3 is equivalent to z3 = ni=1 Tt=1 ( DHiti − Xit )+ , since the optimal solu-
tion must meet the following two conditions,
(1) Eit > 0 occurs, only if Xit > Dit , ∀i, t ;
(2) When Dit > xit , there must be Yij t = 0, ∀1 ≤ j < i, i = 1, 2, . . . , n.
So, we get
n  T  + 
1 Dit
z̄3 = − Xit − C3 , (40)
i=1 t=1
z3 Hi

n  T
1
z̄4 = Oit − C4 , (41)
i=1 t=1
z4
T
From the constraint (6), we get Si− = t=1 (Qi Xit − Hi Eit ), ∀i. Then,

n  T
1
z̄5 = (Qi Xit − Hi Eit ) − C5 . (42)
i=1 t=1
z5

5
By replacing z̄i in min J = i=1 λi z̄i with the formulas (38–42), we get the following
objective
IntP2:
 T      + 

n  λ 5 Qi λ4 λ3 Dit
J= λ1 ᾱi + Xit + λ1 ēi + Oit + − Xit
i=1 t=1
z5 z4 z3 Hi
  T −1  
λ5 Hi + λ2 −
− Eit + λ1 β̄i + uit + λ1 β̄i vit −C (43)
z5 t=0
z2

s.t. Constraints (1–5).


In (43), C = 5i=1 λi Ci is a constant since every value Ci is a constant and λi is pre-
determined. Constraint (6) is omitted in IntP2 since it has been included in the objective
(43).
Further, let α̃i = λ1 ᾱi + λ5 Q
z5
i
− λ5 zH5i , ẽi = λ1 ēi + λz44 , f˜i = λz33 , p̃i = λ5 zH5i , Êit = Xit −
Eit , β̃i+ = λ1 β̄i+ + λ2
z2
, and β̃i− = λ1 β̄i− , then the model IntP2 is redescribed as the following
formulation.
386 Ann Oper Res (2007) 155: 361–390

IntP3:
 T   +   T −1

n  D
α̃i Xit + ẽi Oit + f˜i + −
it
J= − Xit + p̃i Êit + [β̃i uit + β̃i vit ] (44)
i=1 t=1
Hi t=0

s.t Constraints (1–5),

j −1
where the constraint (3) is changed into the formulation i=1 sj i Yj it = Êj t .
The model IntP3 can be solved by commercial optimization software packages, e.g.,
Lingo Tools.

Appendix 3 Computing Jimax and Jimin

Computing J1max and J1min

We can show that the objective J1 reaches the upper bound J1max when no substitu-
tions are considered. So, to meet the demand of type-i job on day t , it satisfies that
{dit − (wi − yit )hi }+ ≤ Tit + Ait hi ≤ (dit − wi hi )+ . Further, from the constraint (14),
Tit ≤ min{δi hi wi , (dit − wi Hi )+ }. If δi hi wi < (dit − wi Hi )+ , then the demand of type-i
job can not be fulfilled only by both regular work hours and overtime of category-i staff. So
temporary staff needs to be employed, and the number, Ait , of temporary staff is equal to
+
 (dit −wi hih−δ
i
i hi wi )
approximately. Thus,

 S 
n   
(dit − wi hi − δi hi wi )+
J1max = cio min{δi hi wi , (dit − wi Hi )+ } + cia .
i=1 t=1
hi

The value of J1min needs to be computed through solving the following model.
MinJ1:

J1min = Min (cio Tit + cia Ait ) (45)
i t

s.t.

n
sj i x̄j it hj + Tit + Ait hi ≥ dit , ∀i, ∀t, (46)
j =i


j
sj i x̄j it + yj t ≥ wj , ∀j, ∀t, (47)
i=1

0 ≤ Tit ≤ δi hi wi , ∀i, t, (48)


S
yj t ≥ F wj , ∀j. (49)
t=1

Since we do not care for the variables dit+ and dit+ ≥ 0 for all i and t , the constraint (13)
is changed into the constraint (47).
The model MinJ1 can be solved by a similar method discussed in the Appendix 4, since
the model MinJ1 is a special case of the model IntS1.
Ann Oper Res (2007) 155: 361–390 387

Computing J2max and J2min

Since we do not care for the objective J1 , and there is no constraint for the quantity Ait
of temporary staff, a special case occurs when all jobs will be assigned to temporary staff.
Then, under this case, all Yj it are equal to 0. From the constraint (13), we get yj t + dj+t = wj .
      
So nj=1 St=1 yj t = nj=1 St=1 (wj − dj+t ) = S nj=1 wj − nj=1 St=1 dj+t . Further, from
n S n
the constraint (15), we get j =1 t=1 yj t ≤ j =1 (S − F )wj . Therefore, we have J2max =
n
j =1 (S − F )wj .
On the other hand, for any feasible solution of the model IntS1, if we let yj t := yj t + dj+t

and d + := 0, then it is still a feasible solution because the constraint t yj t ≥ F wj . So
 j t + +
t dj t = 0 is feasible. Further, since dj t ≥ 0 for all j and t , we get J2 = 0.
min
j

Appendix 4 Solving the model IntS1

The constraint matrix of IntS1 exhibits a block angular structure. The rows obtained from
the constraint (15) contain all the variables yj t for every index j , so they couple all variables
yj t for a same category staff. However, the remaining rows can be decomposed concerning
the indices t into S sets with a same time index. Thus, if the constraint (15) is relaxed and
incorporated into the objective function (22) with nonnegative weights κi (i = 1, 2, . . . , n),
the resulted model is as follows.
MIntS1:

v ∗ = max min (λ1 cito Tit + λ1 cita Ait − λ2 dit+ − κi yit ) − Cd (50)
κ Tit ,Ait ,dit+ t
i

s.t.

n
sj i x̄j it hj + Tit + Ait hi ≥ dit , ∀i, ∀t, (51)
j =i


j
sj i x̄j it + yj t + dit+ = wj , ∀j, ∀t, (52)
i=1

0 ≤ Tit ≤ δi hi wi , ∀i, t, (53)

n
where Cd = j =1 κj F wj is a constant and thus can be neglected in the following computa-
tion. The model MIntS1 can be decomposed into S time-level subproblems as follows.
SubIntSt (t = 1, 2, . . . , S):


n
vt = max min (λ1 cito Tit + λ1 cita Ait − λ2 dit+ − κi yit ) (54)
κ Tit ,Ait ,dit+
i=1

s.t.

n
sj i x̄j it hj + Tit + Ait hi ≥ dit , ∀i, (55)
j =i
388 Ann Oper Res (2007) 155: 361–390


j
sj i x̄j it + yj t + dit+ = wj , ∀j, (56)
i=1

0 ≤ Tit ≤ δi hi wi , ∀i, (57)

The stepsizes κt can be computed iteratively by a subgradient optimization (see, Geoffrion

1974).
Every subproblem SubIntSt can be solved independently. First we assign only category-i
staff to the type-i job. If overtime or temporary staff of category-i are needed, then we con-
sider how to use category-j staff to substitute category-i, to reduce the amount of overtime
or temporary staff of category-i, provided that there are redundant category-j employees
for j > i and sj i = 1. In the objective function (54), both the coefficients of dit+ and yit are
non-positive, which means that it can also reduce the value of objective function (54) when
the redundant category-j employees take day t off (that is, yit becomes larger) or do nothing
on day t (that is, dit+ becomes larger). So, there is a balance among the values λ1 cito , λ1 cita ,
λ2 and κi .
To summarize, the approach is described as follows.

Step 1. Compute d it = dit − wi hi for all i.

(Compute the values Tit , Ait for every category of staff independently)
If d it > 0, then yit = dit+ = 0, Tit = min{δi hi wi , d it } and Ait =  d it h−T
i
it
.
If d it ≤ 0, then let wi = wi −  dhiti , Tit = Ait = 0.
Step 2. Do loop for i = 1, 2, . . . , n that satisfies d it > 0.
(Compute substitutions that can reduce the objective value of (54))
Step 2.1 If (Ait ≥ 0 and λ1 cia ≤ λ2 ) or (λ1 cio ≤ λ2 < λ1 cia and Ait = 0), then no
substitutions occur for the category-i staff and do the next loop with i =
i + 1. Otherwise, do the next step with k = 1.
Step 2.2 Compute j (k) = arg minj {κj |sj i = 1 and wj > 0, j = i + 1, . . . , n}.
If (Ait ≥ 0 and λ1 cia ≤ κj (k) ) or (λ1 cio ≤ κj (k) < λ1 cia and Ait = 0), then
no substitutions occur for the category-i staff and do the next loop with
i = i + 1.
Step 2.3 If Ait > 0 and λ1 cia > max{λ2 , κj (k) } ≥ λ1 cio , then let wj (k) = max{wj (k) −
Ait , 0} and Ait = max{Ait − wj (k) , 0}.
If Ait > 0, then return to step 2.2 with k = k + 1.
Step 2.4 If Tit > 0 and λ1 cio > max{λ2 , κj (k) }, then let w j (k) = max{wj (k) −  Thiti −
Ait , 0}, Ait = max{Ait − w j (k) , 0}, and Tit = max{Tit + Ait hi − w j (k) hi , 0}.
If Tit > 0, then return to step 2.2 with k = k + 1.
Step 3. Do loop for i = 1, 2, . . . , n
(Compute values dit+ and yit )
If w i > 0 and λ2 > κt , then dit+ = wi and yit = 0.
If w i > 0 and λ2 ≤ κt , then dit+ = 0 and yit = wi .
Otherwise, dit+ = 0 and yit = 0.

Obviously, the time requirements of both Steps 1 and 3 are O(n). In the Step 2, the worst
case occurs when at most n − i computations are needed for some i (i = 1, 2, . . . , n), so
the time requirement in Step 2 is at most (n2 ). Thus, the computational complexity of each
subproblem is O(n2 ) in the worst case.
Ann Oper Res (2007) 155: 361–390 389

Appendix 5 Introduction of the AHP approach

For details of the Analytical Hierarchy Process (AHP) approach, please refer to Sect. 14.3,
(Winston 1994). The approach is described as follows briefly.
Suppose there are n objectives. We begin by writing down an n × n matrix (the pairwise
comparison matrix) A = (aij )n×n , where the entry aij indicates how much more important
objective i is than objective j . “Importance" is to be measured on an integer-valued 1–9
scales. For all i, it is necessary that aii = 1, and for consistency, it is necessary that aj i = 1/k
when aij = k.
Then, for each of A’s columns, divide each entry in column i of A by the sum of the
entries in column i and yield a new matrix (call it Anorm = (ȧ ij )n×n , for normalized). So we
get the weight vector w = (wi , i = 1, 2, . . . , n), where wi = nj=1 ȧij /n.
Finally, we should check the consistency of the decision maker’s comparisons, which
uses the following two-step procedure.
n ith entry in AwT
( n1 i=1 T )−n
Step 1. Compute CI = ith entry in w
n−1
;
Step 2. Compare CI to the random index (RI) for the approximate value of n, shown in
Table 23 of Sect. 14.3 (Winston 1994). If CI RI
< 0.1, the degree of consistency is satisfactory;
Otherwise, inconsistency may exist and the AHP may not yield meaningful results.

References

Abernathy, W. J., Baloff, N., Hershey, J. C., & Wandel, S. (1973). A three-stage manpower planning and
scheduling model—a service-sector example. Operations Research, 21(3), 693–711.
Alfares, H. K. (2004). Survey, categorization and comparison of recent tour scheduling literature. Annals of
Operations Research, 127(1–4), 145–175.
Billionnet, A. (1999). Integer programming to schedule a hierarchical workforce with variable demands.
European Journal of Operational Research, 114(1), 105–114.
Bowey, A. M. (1977). Coorporate manpower planning. Management Decision, 15, 421–469.
Bres, E. S., Burns, D., Charnes, A., & Cooper, W. W. (1980). A goal programming model for planning officer
accessions. Management Science, 26(8), 773–783.
Cai, X., & Li, K. N. (2000). A genetic algorithm for scheduling staff of mixed skills under multi-criteria.
European Journal of Operational Research, 125(2), 359–369.
Cai, X., Li, Y. J., & Tu, F. S. (2004.) An improved dynamic programming algorithm for optimal manpower
planning. Lecture series on computer and computational sciences (Vol. 1, pp. 102–105). VSP Interna-
tional Science Publishers.
Cheng, C. H., Madan, M. S., Gupta, Y., & So, S. (2001). Solving the capacitated lot-sizing problem with
backorder consideration. Journal of the Operational Research Society, 52(8), 952–959.
Edwards, J. S. (1983). A survey of manpower planning models and their applications. Journal of the Opera-
tional Research Society, 34(11), 1031–1040.
Ernst, A. T., Jiang, H., Krishnamoorthy, M., Owens, B., & Sier, D. (2004). An annotated bibliography of
personnel scheduling and rostering. Annals of Operations Research, 127(1/4), 21–144.
Gans, N., Koole, G., & Mandelbaum, A. (2003). Telephone call centers: tutorial, review, and research
prospects. Manufacturing & Service Operations Management, 5(2), 79–141.
Gen, M., Tsujimura, Y., & Ida, K. (1992). Method for solving multiobjective aggregate production planning
problem with fuzzy parameters. Computers & Industrial Engineering, 23(1–4), 117–121.
Geoffrion, A. M. (1974). Lagrangian relaxation for integer programming. Mathematics Programming Study,
2, 82–114.
Georgiou, A. C., & Vassiliou, P. -C. G. (1997). Cost models in nonhomogeneous Markov systems. European
Journal of Operational Research, 100(1), 81–96.
Henderson, J. C., Krajewski, L. J., & Showalter, M. J. (1982). An integrated approach for manpower planning
in the service sector. Omega, 10(1), 61–73.
Li, Y. J., Chen, J., Cai, X., & Tu, F. S. (2005). Optimal manpower planning with temporary labor and contract
period constraint. Lecture notes in computer science (Vol. 3521, pp. 350–359). Springer.
390 Ann Oper Res (2007) 155: 361–390

McClean, S. (1991). Manpower planning models and their estimation. European Journal of Operational
Research, 51(2), 179–187.
McConnell, E. A. (2000). Staffing and scheduling at your fingertips. Nursing Management, 31(3), 52–54.
Miller, H. E., Pierskalla, W. P., & Rath, G. J. (1976). Nurse scheduling using mathematical programming.
Operations Research, 24(5), 857–887.
Price, W. L., Martel, A., & Lewis, K. A. (1980). A review of mathematical models in human resource plan-
ning. Omega, 8, 639–645.
Purkiss, C. (1981). Corporate manpower planning: a review of models. European Journal of Operational
Research, 8(4), 315–323.
Rao, P. P. (1990). A dynamic programming approach to determine optimal manpower recruitment policies.
Journal of the Operational Research Society, 41(10), 983–988.
Samuel, G. D., & Reutzel, E. T. (1981). A dynamic programming approach to work force scheduling with
time-dependent performance measures. Journal of Operations Management, 1(3), 165–172.
Silverman, J., Steuer, R. E., & Whisman, A. W. (1988). A multi-period, multiple criteria optimization system
for manpower planning. European Journal of Operational Research, 34(2), 160–170.
Smith, A. R. (1971). Models of manpower systems. New York: Elsevier.
Taubert, W. H. (1968). A search decision rule for the aggregate scheduling problem. Management Science,
14, B343–B360.
Topaloglu, S., & Ozkarahan, I. (2004). An implicit goal programming model for the tour scheduling problem
considering the employee work preferences. Annals of Operations Research, 128(1-4), 135–158.
Uebe, G., Schäffer, M., & Kitta, R. (1990). A cycle integer covering problem. International conference on
mathematical and computer modelling (pp. 1079–1084). Oxford: Pergamon.
Venkataraman, R., & Brusco, M. J. (1996). An integrated analysis of nurse staffing and scheduling policies.
Omega, 24(1), 57–71.
Winston, W. L. (1994). Operation research: applications and algorithms. Belmont: Duxbury.
Zanakis, S. H. & Maret, M.W. (1981). A Markovian goal programming to aggregate manpower planning.
Journal of the Operational Research Society, 32(1), 55–63.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.