See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/304573109

Strategic Health Workforce Planning

Article in IIE Transactions · June 2016

DOI: 10.1080/0740817X.2016.1204488

CITATION
S READS

17 417

4 authors, including:

Weihong Hu
Mariel S. Lavieri
Georgia Institute of
Technology University of Michigan

5 PUBLICATIONS 64 CITATIONS 87 PUBLICATIONS 1,379 CITATIONS

SEE PROFILE SEE PROFILE

Xiang Liu
Tsinghua University
20 PUBLICATIONS 197 CITATIONS

SEE PROFILE

All content following this page was uploaded by Weihong Hu on 25 August 2020.

The user has requested enhancement of the downloaded file.

 Strategic Health Workforce Planning

Weihong Hu∗, Mariel S. Lavieri†, Alejandro Toriello∗, Xiang Liu†

∗
H. Milton Stewart School of Industrial and Systems Engineering

Georgia Institute of Technology

Atlanta, Georgia 30332

weihongh at gatech dot edu, atoriello at isye dot gatech dot edu

†
Department of Industrial and Operations Engineering

University of Michigan

Ann Arbor, Michigan 48109

{lavieri, liuxiang} at umich dot edu

April 27, 2016

Abstract

Analysts predict impending shortages in the health care workforce, yet wages for health care workers
already account for over half of U.S. health expenditures. It is thus increasingly important to
adequately plan to meet health workforce demand at reasonable cost. Using infinite linear
programming method- ology, we propose an infinite-horizon model for health workforce planning
in a large health system for a single worker class, e.g. nurses. We give a series of common-sense
conditions any system of this kind should satisfy, and use them to prove the optimality of a
natural lookahead policy. We then use real-world data to examine how such policies perform in
more complex systems; in particular, our experiments show that a natural extension of the
lookahead policy performs well when incorporating stochastic demand growth.

1
1 Introduction

Health workforce planning plays a key role in the United States and worldwide. Analysts project that by
2020 the U.S. will experience a shortage of up to 100,000 physicians, up to one million nurses and up
to 250,000 public health professionals [57]. Adequate staffing of medical units has been shown to have a
direct impact in the quality of patient care [43], and also accounts for a considerable fraction of health
care costs, with wages for health care workers representing 56% of the $2.6 trillion spent on health care
in the United States in 2010 [36].
As the U.S. population continues to age [56] and demand for health care continues to grow, different
sectors of the population will compete for constrained and costly health care resources. It thus
becomes increasingly important to understand how the health care needs of the population are linked to
long-term workforce management plans of doctors, nurses and other medical personnel. The challenge
is to ensure that sufficient resources are available in the future to meet the growing health care needs of the
population, while accounting for the costs associated with meeting these needs. These workforce levels
should meet the demand for resources in the present and be positioned to meet demand for the
foreseeable future [50], an essentially infinite horizon. Furthermore, workforce plans should account
for lags implied by training new members of the workforce, attrition stemming from retirements, firings
and resignations, and also the adequate supervision of workers at different levels of the workforce
hierarchy by their superiors.
Current practice has mostly focused on monitoring and evaluating health human resource systems [20],
yet a systematic framework is needed to understand the long term implications of the sequential
decisions made in those systems. Given the significant costs and the impact on health care outcomes
associated with workforce decisions, it is essential for stakeholders in large health systems to
understand the role of the planning horizon and the long-term consequences of health workforce plans.
We therefore propose to study the planning of workforce training, promotion and hiring within
such systems, with the main goal of designing a natural policy for decision makers to implement, and
concurrently determining common-sense conditions under which this policy is in fact optimal.
Governments, regulatory bodies, professional associations, representatives from the private sector and
senior health system executives may use the results presented in this paper to gain a deeper understanding
on where incentives should be placed to best meet the health workforce needs of the population. Our
2
focus is on decisions at a health care

3
policy or public policy level (i.e. not on individual hiring and firing decisions), and thus our model
includes several stylized simplifications. The problem scale we are interested in has workforces
numbering in the thousands or the tens of thousands, e.g. state or provincial health systems, large
hospital conglomerates, or the U.S. Veterans Administration. We therefore model the workforce as a
continuous flow and allow fractional quantities in our solutions.
We also assume centralized control of the system, which may only be realistic in some cases.
Neverthe- less, even for those systems in which this is not entirely the case, the conditions we list can
help decision makers with limited control in monitoring the system’s behavior and deriving policy
recommendations; this is precisely the approach [51] take to study the U.S. pediatric nurse practitioner
workforce.
Although uncertainty is present in any health system’s dynamics, the model we propose is deterministic,
and represents a first step in understanding how hiring, training and promotion interact. The deterministic
model allows for some preparation against uncertainty through sensitivity analysis. In addition, the structure
of solutions suggested by our analysis can be successfully extended to models with uncertainty; we
include computational experiments on a model with stochastic demand growth to demonstrate this.

1.1 Our Contribution

We propose a discounted, infinite linear programming model for strategic workforce planning, which
in- cludes training, promotion and hiring decision for a class of health workers within a hierarchical system.
The model takes as input a demand forecast, workforce payroll, training and hiring costs, workforce
hierarchy parameters and a discount factor. Though similar finite models have appeared in earlier work
[37, 38, 39], our focus here is to derive structural results and study workforce management policies
that are provably optimal under reasonable assumptions. Specifically, we consider the following to be our
main contributions:

i) We give a series of common-sense conditions any system of this kind should satisfy under our
as- sumptions, demonstrate the pathological behavior that can occur when they are not satisfied
and de- rive useful structural properties of the optimal solutions from the conditions. Though
based on our assumptions, these conditions may help guide decision making in more complex
systems.
4
ii) We prove that a natural lookahead policy is optimal for our model. In addition to optimizing
this model in particular, the result is useful because lookahead policies mimic how more complex
models

5
 may be managed in practice.

iii) We provide a two-part computational study based on real-world nursing workforce data. The
first component demonstrates the effectiveness of the lookahead policy in a more complex
deterministic system with additional detail, such as worker age. The second component shows that
lookahead poli- cies perform extremely well in a setting with stochastic demand growth, arguably the
most important source of uncertainty in our model.

The remainder of the paper is organized as follows: This section closes with a literature review.
Section 2 formulates our model and states the conditions we assume. Section 3 uses the conditions to
show some structural properties of optimal solutions, proves the optimality of our proposed policy and
discusses duality and sensitivity. Section 4 discusses experiments that test our policy on more complex
models, and Section 5 concludes outlining future research avenues. The Appendix contains technical
proofs and some additional modeling information.

1.2 Literature Review

Workforce planning models are not new to the industrial engineering and operations research literature,
with work stretching back several decades, such as [1, 9, 34, 41, 46, 52]. Workforce management
models have been developed to manage workforce in call centers [27], military personnel [29],
medical school budgets [16, 40], as well as to address cross-training and flexibility of the workforce
[44, 58]. [11, 25, 55] provide overviews of workforce/manpower planning models, while [15] discuss
the need of a greater interface between operations and human resource management models and the
complexities associated with those models. Recent work continues to address workforce issues in
operational or tactical time frames,
e.g. [10]; this focus on shorter horizons extends also to health care and emergency workforce
planning [14, 23, 26, 62]. The long-term workforce capacity planning models [6, 28, 54] are related to our
work, yet they concentrate only on the recruitment and retention of personnel without incorporating some of
the other decisions required to manage health care personnel. On the other hand, models such as [8, 61]
concentrate on skill acquisition and on-the job learning, focusing on a shorter time scale. The results in
[59, 60] and the recent survey [55] particularly highlight the need to research long-term health workforce

6
planning, among other areas.

7
 The work in [37, 38, 39] develops a workforce planning model of the registered nursing workforce
of British Columbia. The model ranges over a 20-year planning horizon, and provides policy
recommendations on the number of nurses to train, promote and recruit to achieve specified workforce
levels. Our proposed model includes similar decisions, but is formulated over an infinite horizon.
Furthermore, whereas this past work was only numerical, we include both a theoretical analysis on the
structure of optimal policies as well as numerical experiments.
Infinite-horizon optimization has been widely applied to various operational problems, mostly via dy-
namic programming [45, 63]. However, the last two or three decades have also seen the direct study
of infinite mathematical programming models and specifically infinite linear programs for operations manage-
ment applications. Some problems studied in the literature include inventory routing [2, 3], joint replenish-
ment [4, 5], production planning [19, 48], and equipment replacement [35]. To our knowledge,
although dynamic programming has been applied to model some workforce management issues, e.g. [28, 46],
infinite linear programming has not yet been considered in the literature to address this topic. Furthermore,
work- force management possesses differences with other resource management problems that deal
mostly with products [28], which impedes the direct application of existing results.
A general reference for infinite linear programming is [7]. Our models operate in countable dimensions,
and follow the general structure of models such as [19, 31, 32, 47, 48, 49, 53]. For a recent overview
of optimization in health care, we refer the reader to [12].

2 Model Formulation and Assumptions

We consider an infinite-horizon, discounted workforce planning model with the following characteristics.
There is a deterministic demand forecast for each period, and the population of workers at the lowest level of
the system, e.g. junior nurses, must be at least equal to that period’s demand. The system has a fixed number
of levels above this first level; worker population at each higher level must be at least a fixed fraction
of the same period’s population one level below, to ensure adequate supervision. Between one period
and the next, a fixed fraction of each level’s population leaves the system, accounting for retirements,
firings and resignations. New workers may be added to any level directly via hiring, or indirectly
through student admission and training at the first level, and promotion at higher levels; there is no

8
down-sizing, i.e. mass

9
firing to reduce workforce levels. Student populations take one period to train before entering at the
first workforce level; similarly, only workers who have been in a level for at least one period may be
promoted. We discuss how to extend our results to models with longer training in the subsequent
sections.
The model is defined by the following parameters.

• n ≥ 2: Number of workforce hierarchy levels.

• hk > 0: Per-period variable payroll costs for level k = 1, . . . , n.

• ck > 0: Variable training (k = 0) or hiring (k ≥ 1) costs for level k = 0, . . . , n.

• ck,k+1 ≥ 0: Variable promotion cost from level k = 1, . . . , n −1 to k +1. Workers may only be
promoted once they have worked at a particular level for at least one period.

• γ ∈ (0, 1): Discount factor, adjusted to account for cost increases. That is, if γ¯ is the nominal
discount rate and α > 1 is the cost growth rate, then γ = αγ¯; this is the reciprocal of the “health
care inflation.”

• dt > 0: Forecasted level-1 workforce demand for period t = 1, . . . .

• qk,k+1 ∈ (0, 1): Minimum fraction of level-k workers needed at level k + 1, for k = 1, . . . , n − 1.

• pk ∈ (0, 1): Per-period retention rate of workers that stay in the system at level k = 0, . . . , n from
one period to the next. The attrition rate 1 − pk is the fraction of workers at level k expected to
leave the system from one period to the next; this includes firing, retirement and quitting.

• s0k: Students (k = 0) or workers in level k = 1, . . . , n at the start of the current period, before attrition.

The model’s decision variables are:

• stk: Students (k = 0) or workers in level k = 1, . . . , n at end of period t = 1, . . . .

• xtk: Students admitted (k = 0) or workers hired at level k = 1, . . . , n in period t = 1, . . . .

t
k,k+1 : Workers promoted from level k = 1, . . . , n − 1 to k + 1 in period t = 1, . . . .
•x

Our strategic workforce planning problem then has the following formulation.

n−1 n
inf C(s x ) = ∞ n
ck xkt + ∑ ck,k+1
, ∑ k,k+1 +∑h s
t
(1a)
γt−1
∑ k=
k=
k=1
k k

0
1 xt

10
 s.t. st1 ≥ dt, ∀ t = 1, . . . (1b)

stk+ — qk,k+1st ≥ 0, ∀ k = 1, . . . , n − 1, ∀ t = 1, . . . (1c)

1
k
p1st−1 − st + p0xt−1 − xt + xt = 0, ∀ t = 1, . . . (1d)
1 1 0 12 1

pkst−1 − st + xt − xt + xt = 0, ∀ k = 2, . . . , n − 1, ∀ t = 1, . . . (1e)
k k k−1,k k,k+1 k
pnst−1 − st + xt + xt = 0, ∀ t = 1, . . . (1f)
n n n−1,n n

pkst−1 − xt ≥ 0, ∀ k = 1, . . . , n − 1, ∀ t = 1, . . . (1g)
k k,k+1

st, xt ≥ 0, ∀ t = 1, . . . , (1h)

where st = xt for t = 0,.....We take the feasible region to be the subset of solutions for which the objective
0 0

is well defined and finite [48]. In the model, the objective (1a) minimizes discounted cost over the infinite
horizon. The demand satisfaction constraint (1b) ensures enough level-1 workers are present to satisfy
projected demand each period, while (1c) ensures the minimum required fraction of level-(k + 1)
workers are present to supervise level-k workers. The flow balance constraints (1d–1f) track workers present
at each level from one period to the next, and (1g) limits the promoted workers from level k to k + 1 to those
present in level k for at least one period. The domain constraints (1h) ensure non-negativity of worker
levels, hires, promotions and student admissions.

Whereas most of the model’s parameters are stationary and can thus be explicitly given or recorded, the
demand forecast is an infinite sequence that cannot be explicitly given. In practical terms this forecast can
only be modeled implicitly, for instance by giving a first-period demand and a per-period growth rate. While
our results hold for an arbitrary sequence satisfying our assumptions, our policy requires explicit knowledge
of only the first few values of the sequence (two in the model as currently stated, but see Corollary 3.4 below
for an extension). For a discussion of related issues with non-stationary data in infinite-horizon optimization,
see e.g. [30].
We next list several conditions the model should satisfy. These conditions are common in many real
world settings or are reasonable approximations, and are necessary for most of our subsequent results.
Many are also necessary to avoid pathological behavior. We begin with technical assumptions.

Assumption 2.1 (Technical assumptions).

11
 i) Finite total demand: Total discounted demand converges.

∑ γt−1dt < ∞ (2a)

t=1

ii) Linear costs: The objective function is linear with respect to the decision variables s and x.

The former assumption is necessary to have a finite objective and thus a feasible problem. The
latter is required to apply linear programming techniques. Though large changes in a system’s workforce
could render some costs non-linear (e.g. a large increase in hiring leading to an increase in hiring and
payroll costs because of the labor market’s supply), our results suggest that in the long run moderate
decisions predominate, and thus the assumption of linearity is reasonable.

Assumption 2.2 (Growing demand). The sequence (dt) is non-decreasing.

dt ≤ dt+1, ∀ t = 1, . . . (2b)

This assumption reflects most contemporary health care systems in which demand is expected to
grow for the foreseeable future, and ensures that training and promotion will be perpetually necessary within
the system. As discussed by [21], given changes in the demographics of the population, as well as
expanded coverage under the Affordable Care Act, demand for primary care services in the United States is
expected to grow by 14% by 2025. This expected growth in demand for health care workers is not unique
to the United States; it is estimated that an additional 1.9 billion people will seek access to health care
by 2035 [17]. In more general cases, even if demand is only expected to be eventually non-decreasing,
our conclusions can be applied starting at the period where non-decreasing growth begins, with a finite
model accounting for the system in preceding periods.
The first non-technical assumption concerns the relative costs of payroll, promotion and hiring.

Assumption 2.3 (Promotion is preferable). Even when factoring attrition, payroll costs and discounting,
promotion is cheaper than hiring.

c0 ck + hk
c k ≤ c 1, + ≤c , ∀ k = 1, . . . , n − 1 (2c)
γ p0 γ pk ,k+1 k+1

12
 If this assumption does not hold at some level in the hierarchy, there is no incentive to train and promote
from within beyond that point. This condition should be satisfied by many workforce systems, both in health
care and in other industries.
The next assumption is slightly more specific to the health care industry, but still common in
other industries.

Assumption 2.4 (Non-increasing retention). The hierarchy does not tend to become top-heavy:

pk ≥ pk+1, ∀ k = 1, . . . , n − 1 (2d)

This assumption is natural in health care hierarchies such as nursing, where higher-level workers are
usually older, since older workers tend to retire or leave the system for other reasons at a higher rate. The
assumption is more problematic, for example, in industries where tenure guarantees at an intermediate level
imply an unnaturally high attrition at lower levels.
For some of our results, it is necessary to further strengthen the previous assumption.

Assumption 2.4’ (Equal retention). Retention and attrition are equal at all hierarchy levels:

pk = pk+1, ∀ k = 1, . . . , n − 1 (2d’)

Though it appears restrictive, in many real-world systems the top and bottom retention rates in fact only
differ by a few percentage points [37, 38, 39].

Assumption 2.5 (Non-decreasing payroll). Salaries increase within the hierarchy, even when
accounting for attrition:

hk hk+1
≤ , ∀ k = 1, . . . , n − 1 (2e)
1 − γ pk 1 − γ pk+1

As the next example shows, this condition prevents undesirable behavior.

Example 1 (Down-sizing by promotion). Consider a two-level system which is drastically over-staffed. Let
dt = ε for all t, where ε > 0 is a small positive number, and let 1s0 ≫ ε. If (2e) is not satisfied, it may be

13
optimal because of (2d) to promote all but ε workers to level 2, effectively down-sizing the workforce
by promoting most of it, and achieving lower costs in the process. Such behavior could lead to detrimental side
effects, such as poor morale in the remaining workforce.

Assumption 2.6 (Moderate demand growth). Demand does not grow too quickly:

dt+1 pmin
≤ , ∀ t = 1, . . . , (2f)
dt qmax

where pmin = mink pk and qmax = maxk qk,k+1.

Intuitively, the assumption ensures enough worker population at each level to promote to the next level
as demand grows; it is easily satisfied in most systems. For example, if n = 2, p1 = p2 = 0.8 and q12 =
0.25, (2f) requires the demand growth to be no more than 320% per period, a condition met in
virtually any system. Furthermore, as the next example shows, when this assumption is not met, the
planning horizon necessary to compute an optimal solution may be arbitrarily long.

Example 2 (Excessive demand growth). We consider a two-level system that experiences excessive demand
growth for a given number of periods, and constant demand thereafter. To simplify the numbers in
the example, we set p1 = p2 = q12 = 1. For a fixed m ≥ 2 let

 t−1
2 − 1, t = 2, . . . ,
d1 = 1,
m m
dt ,
= 2 − 1, t = m + 1, . . .
m−1
m

and s0 = 0, s0 = s0 = 1; note that dt+1/dt > p1/q12 = 1 for t = 2, . . . , m − 1. Table 1 details the first demand
0 1 2

values in the sequence, for m = 3, 4, 5. The table also lists a solution that satisfies demand without
any hiring, which can be made optimal by choosing large enough hiring costs. Although projected
demand for
the first three periods is identical in all cases, the optimal number of students admitted in the first
period changes with m; for general m, we get x1 = (2m−2 − 1)/2m−3. In other words, the current period’s
0
decision may depend on a horizon of arbitrary length m.

As Example 2 suggests, the condition (2f) can be relaxed; we include the best possible condition of this
kind we could derive in the Appendix (see the proof of Claim A.3). However, (2f) is much simpler to
state
14
 t 1 2 3 4 5 ·· ·
dt3 1 1 3 3 3 ·· ·
x0t 1 3 0 0 0 ·· ·
1 3/2 3 3 3 ·· ·
s1t
1 3/2 3 3 3 ·· ·
s2t
dt4 1 1 3 7 7 ·· ·
x0t 3/2 7/2 7 0 0 ·· ·
1 3/2 7/2 7 7 ·· ·
s1t
1 3/2 7/2 7 7 ·· ·
s2t
dt5 1 1 3 7 15 ·· ·
x0t 7/4 15/4 15/2 15 0 ·· ·
1 15/8 15/4 15/2 15 ·· ·
s1t
1 15/8 15/4 15/2 15 ·· ·
s2t
Table 1: Sample demand sequences and solutions with no hiring for m = 3, 4, 5 in Example 2.

and suffices for any practical situation.

3 Optimal System Behavior

We begin our characterization of optimal solutions of (1) by outlining structural properties satisfied in mod-
els that meet our assumptions. We include only simple proofs here and relegate any complex proof to the
Appendix.

Lemma 3.1 (No unnecessary hiring). Suppose the model parameters satisfy Assumptions 2.1 through
2.3. There is an optimal solution of (1) in which no hiring takes place when promotion is possible:

xt1 = 0, ∀ t = 2, . . . (3a)
(pks
t−1
−x t
)x t
= 0, ∀ k = 1, . . . , n − 1, ∀ t = 1, . . . (3b)
k k,k+1 k+1

Proof. If a solution does not satisfy either condition, a simple substitution produces another solution with
equal or lesser objective that does satisfy the conditions.

Lemma 3.2 (No excess training or promotion). Suppose Assumptions 2.1, 2.2, 2.3, 2.5 and 2.6 hold.
Fur- thermore, suppose either Assumption 2.4 holds and n = 2, or Assumption 2.4’ holds. Then there
is an

15
optimal solution of (1) in which no excess promotion or student admittance occurs:

(st − dt)xt−1 = 0, ∀ t = 2, . . . (4a)

1 0
(st − qk,k+1st )xt = 0, ∀ k = 1, . . . , n − 1, ∀ t = 1, . . . (4b)
k+1 k k,k+1

Like the preceding lemma, Lemma 3.2 follows from applying a substitution or perturbation to
any solution that does not satisfy it. However, unlike in the hiring case, a perturbation in promotion has
ripple effects in higher levels of the hierarchy and in later periods that render it much more complex.
With these two structural properties in place, we are able to characterize optimal solutions of (1). Con-
sider the two-period restriction of (1) given by

n n−1 n
min ∑ ck(x1 + γx2) + ∑ ck,k+1(x1 k+1 + γx2 k+1) + ∑ hk(s1 + γs2) (5a)
k k k, k k
k=0 k=1
k,
k=1

s.t. s ≥ dt,
t
∀ t = 1, 2 (5b)
1
stk+ — qk,k+1st ≥ 0, ∀ k = 1, . . . , n − 1, ∀ t = 1, 2 (5c)
1
k
p1st−1 − st + p0xt−1 − xt + xt = 0, ∀ t = 1, 2 (5d)
1 1 0 12 1

pkst−1 − st + xt − xt + xt = 0, ∀ k = 2, . . . , n − 1, ∀ t = 1, 2 (5e)
k k k−1,k k,k+1 k
pnst−1 − st + xt + xt = 0, ∀ t = 1, 2 (5f)
n n n−1,n n

pkst−1 − xt ≥ 0, ∀ k = 1, . . . , n − 1, ∀ t = 1, 2 (5g)
k k,k+1

st, xt ≥ 0, ∀ t = 1, 2, (5h)

A one-period lookahead policy constructs a solution to (1) by iteratively solving (5), fixing the variables
for t = 1, stepping one period forward by relabeling t ← t + 1 for all variables and parameters, and repeating
the process. In practice, this corresponds to a decision maker planning the current period’s promotion,
training and hiring based on current demand and the next period’s forecasted demand, while ignoring
demand for subsequent periods.

Theorem 3.3 (Optimality of one-period lookahead policy). Suppose Assumptions 2.1, 2.2, 2.3, 2.5 and
2.6 are satisfied. Suppose either Assumption 2.4 holds and n = 2, or Assumption 2.4’ holds. Then one-

16
period

17
lookahead policies are optimal.

Corollary 3.4 (Increased training time). Suppose students require L ≥ 1 periods to train instead of one,
with all other system characteristics remaining the same. Under the conditions of Theorem 3.3, L-period
looka- head policies are optimal, where an L-period lookahead is defined analogously to a one-period
lookahead but with L additional periods instead of one.

Proof. The proof of Theorem 3.3 still applies; we are simply relabeling level-0 variables.

These results indicate that good workforce planning decisions can be made using a minimal amount of
forecasted information, which strengthens the robustness of the resulting solution since forecasts of more
distant demand naturally tend to be less reliable. This also places our result within the context of
solution and forecast horizons; see, e.g., [18] for formal definitions and discussion.
Moreover, lookahead policies mimic how such large workforce systems can be managed in practice.
The optimization of the lookahead model (5) is split into two separate sets of decisions: First, the model
decides how to meet the current period’s demand (period 1) at minimum cost given the current
workforce levels and the new entering workforce; this is a more tactical, short-term decision in which the only
recourse is hiring and promotion. Then, based on this decision, the model strategically projects ahead one
period to decide the number of students to admit into training, so that the next period’s demand can be
met also at minimum cost. In (5), the second-period variables are only used to determine this admission
quantity, and are not in fact implemented.
Although the lookahead policy given by (5) optimally solves the original infinite-horizon problem, it is
worth noting that this policy is quite simple to implement. At every period, it involves only the solution of
a small, two-period LP; for example, in a system with four hierarchy levels, the number of variables in
the model is 24, a model size that can be solved even in spreadsheet optimization packages in a few
seconds or less. This suggests that these policies could be useful in more complex settings; we
explore this idea experimentally in Section 4.
Another important question related to (1) is duality. A dual satisfying the typical complementary rela-
tionships can shed additional light on the structure of optimal solutions to (1). Furthermore, optimal dual
prices may also be useful as indicators of the model’s sensitivity to parameters such as demand.
However,

18
 the infinite horizon implies significant technical complications and gives rise to pathologies not encountered
in the finite case.
Extending the typical LP dual construction to (1) yields

∞ n−1
sup
η1
D(µ, λ, η) = ∑ dt µt − p0s0λ 1 − ∑ pks0(λ 1 + ) − pns0λ 1 (6a)
1 0 1 k k k,k+1 n n
t=1 k=1
µt − qk,k+1 µt − λ t + pkλt+1 +
s.t. t+1 ≤ γt−1hk, ∀ k = 1, . . . , n − 1, ∀ t = 1, . . . (6b)
pkη
k k+1 k k k,k+1

n µt − λnt + pnλnt+1 ≤ γt−1hn, ∀ t = 1, . . . (6c)

p0λt1 ≤ γt−2c0, ∀ t = 2, . . . (6d)

λ tk ≤ γt−1ck, ∀ k = 1, . . . , n, ∀ t = 1, . . . (6e)

— λkt + λk+1
t
− ηtk,k+1 ≤ γt−1ck,k+1, ∀ k = 1, . . . , n − 1, ∀ t = 1, . . . (6f)
µ t , ηt ≥ 0, λ t unrestricted, ∀ t = 1, . . . , (6g)

where we similarly define the feasible region as a subset of the points for which the objective is well defined
and finite. However, this model does not satisfy strong or even weak duality with (1).

Example 3 (No weak duality; adapted from [53]). Suppose s0 > 0 and let M > 0. Define λˆ t = −M/pt ,
n n n

∀ t = 1, . . . , and set all other variables to zero. The solution is feasible for (6), and its objective
function value is positive and goes to infinity as M → ∞. However, (1) is clearly feasible and bounded below
by zero.

The following result addresses this problem.

Theorem 3.5. Suppose we can change the equality constraints (1d–1f) to greater-than-or-equal
constraints (and thus impose λ t ≥ 0) for all but a finite number of indices t without affecting optimality in
(1). Let (sˆ, xˆ) and (µˆ , λˆ , ηˆ ) be feasible for (1) and (6) respectively.

i) Weak duality: D(µˆ , λˆ , ηˆ ) ≤ C(sˆ, xˆ).

ii) Strong duality: Both solutions are optimal and D(µˆ , λˆ , ηˆ ) = C(sˆ, xˆ) if and only if
complementary slackness holds (in the usual sense) and transversality [48, 49, 53] holds:

n−1
liminf p 0 λˆ t+1xˆt + ∑ p k (λˆ t+1 + ηˆ t+1)sˆt + p n λˆ t+1sˆt = 0. (7)
t→∞ 1 0 k k=1

19
k k n n

20
Proof. If all constraints eventually become greater-than-or-equal, then the off-diagonal constraint matrix of
(1) in inequality form is eventually non-negative, implying that [48, Assumption 3.1] holds, and thus
the results follow from [48, Theorems 3.3 and 3.7].

Corollary 3.6. The conditions of Theorem 3.5 apply, and therefore weak and strong duality hold, if
demand is eventually non-decreasing.

The results in [48] imply we can use optimal solutions of (6) as shadow prices to perform sensitivity
analysis on (1).

Example 4 (Sensitivity analysis). Consider a two-level system in which the incoming worker
populations in period 1 require some promotion from level 1 to level 2, with enough level-1 workers
remaining after promotion to meet demand in period 1 but not later. Based on these initial conditions
and Assumptions 2.1 through 2.6, Lemmas 3.1 and 3.2 imply the following structure to the optimal
solution:

xt = xt = 0, xt , xt > 0, xt < p1st−1, st = q12st , ∀ t = 1, . . .

1 2 0 12 12 1 2 1

s1 > d1; st = dt, ∀ t = 2, . . .

1 1

The solution for (6) that satisfies complementary slackness and transversality is:

µ11 = 0

µt =
γt−2c0
1
p0 1 − γ p1 + q12(1 − γ p2) + γ q12c12(1 − γ p2)
t−1

+ γt−1(h1 + q12h2), ∀ t = 2, . . .

2 1 c0
µ1 = 1 + q12 p0(p1 − p2) + c12(1 − γ p2) + (h2 − h1)

2 c0
µt = γt−2(1 − γ p2) p0 + γc12 + γt−1h2, ∀ t = 2, . . .
1 c0
λ 11 = 1 + q p0 (p1 + q12 p2) − q12c12(1 − γ p2) − (h1 +
12

t γt−2c0
qλ12
1h=2) , ∀ t = 2, . . .
p0
21
 1 c0
λ 12= 1 + q12 p0 (p1 + q12 p2) + c12(1 + γq12 p2) − (h1 + q12h2)

22
 c0
λ t = γt−2 + γc12 , ∀ t = 2, . . .
2
p0
η = 0, ∀ t = 1, . . .
t

12

It can be verified that this solution is dual feasible provided the assumptions hold. Suppose in particular
that demand grows based on a rate 1 < β < 1/γ, so that dt = βt−1d1. It follows that

∞
t
∑ dt µ1 1 c0 — 1 q12 1 — γ 2) + q12 c12 1 — γ 2) 1 12 2 βγ
t= = d 1 γ + p( p ( + h+ q h .
1 p
γ 1 − βγ
p0
This expression indicates how the optimal cost would change if either d1 or β vary slightly from their
given values.

4 Computational Study

To evaluate the efficacy of our proposed models and policies, we performed computational
experiments based on the British Columbia nursing workforce described in [37, 38, 39]. Health care
human resource data is more readily available from Canadian provinces because of their centralized
control of health care. However, similar data from U.S. systems can be used within a model such as ours
to derive policy recom- mendations, e.g. [51], even though U.S. health systems are usually de-
centralized.
We first discuss the performance of lookahead policies applied to more complex, albeit
deterministic, settings. We then develop a simulation model that considers uncertainty in demand growth and
evaluate the performance of our lookahead policies in this setting.

4.1 Deterministic experiments

While model 1 provides useful insights into the behavior of strategic workforce planning models, possible
extensions include the differentiation of workers by age (as it affects attrition rates), and the extension of the
length of student training (to four years). In order to evaluate the performance of our lookahead
policies, we began by solving the problem over a 25-year planning horizon (a full information model)

23
and used the solution to the first 20 years as our benchmark. We then compared the results to a
solution obtained by implementing a four-year lookahead policy of this extended model.

24
 Figure 1 outlines the structure of the extended model and its parameters (see further model description
in the Appendix). Students are admitted into the training program, where they take four years to train before
entering the workforce. The probabilities of students continuing their education depend on the school
year of the student (with greater attrition in the first year of the program). After graduation, students enter
the first workforce level as direct care nurses. In level 1, the number of workers has to meet current
demand. This demand is met by workers that have not retired or been promoted, graduates from the training
program and workers hired externally. Level 2 consists of nurse managers, a supervisory position to the
first level; nurse managers are either hired externally or promoted from the first workforce level. To
account for transition shock and adaptation to the profession [24], we assume that Level-1 workers must
have worked for at least one year before being promoted into the second workforce level. In both levels,
retention rates depend on the age of the workers. The average retention in level 2 is slightly higher than
the average retention in level 1, which would violate Assumption 2.4 if the averages applied to all age
groups. Furthermore, since the parameter is age-dependent in this model, the actual retention in each level
depends on the age distribution of the worker population. This difference in attrition rates did not impact
our results, further supporting the

robustness of our findings. We set the discount factor to γ = 0.95.

We tested the model in nine scenarios. Among the nine scenarios, the baseline scenario represents the
estimated demand in British Columbia, Canada, starting in 2007; we calculated demand by
extrapolating the population growth between 1996 and 2006 [13]. Scenarios 1 through 4 evaluate the
impact of different demand growth rates. Scenario 5 evaluates the impact of limiting the growth of
the training program. Scenarios 6, 7 and 8 evaluate the performance of the lookahead policy in extreme
conditions where demand has a peak, hiring growth is limited, and costs are varied. The parameter
characteristics and descriptions of the scenarios are summarized in the following list.

Baseline Scenario Fixed demand growth rate of 1.25% per year. Projected demand growth in
British Columbia, Canada.

Scenario 1 Fixed demand growth rate of 0.01% per year. Very low demand growth.

Scenario 2 Fixed demand growth rate of 2.5% per year. High demand growth.

25
Scenario 3 Linearly accelerating demand growth from 0% per year to 2.5% per year over 25 years.

26
 Attrition Rate
by School Year
Admissions Students
$57,991/student 4 years in school
10%, 2%, and 5%
in year 1, 2, and 3

Entering Probability
91%

Workforce Level 1
Hiring Level 1 Age-dependent Attrition
Meet growing demand
$113,880/worker
Payroll cost $87,600/worker/yr

$28,500/worker
Promotions
Decisions
Unit Cost

Workforce Level 2
Hiring Level 2 Supervise Level 1 workers (1 level 2 : 10.5 level 1) Age-dependent Attrition
$283,800/worker Payroll cost $95,104/worker/yr
Has worked in Level 1 for at least 1 yr

Figure 1: Flow chart of model used in computational examples.

Scenario 4 Linearly decelerating demand growth from 2.5% per year to 0% per year over 25 years.

Scenario 5 Fixed demand growth rate of 1.25% per year and student population growth limited to no
more than 1% per year. Major restrictions in training growth.

Scenario 6 Fixed demand growth rate of 1.25% per year in years 1 through 9 and 11 through 25,
demand doubled in year 10. Level-1 hiring growth limited to no more than 50% per year.
This scenario simulated a sudden jump of demand, which might be due to a drastic change in
roles and scope of practice of the workforce. We assumed that drastic changes in the number of
workers hired could not be made without incurring very large recruitment costs.

Scenario 7 Fixed demand growth rate of 1.25% per year in years 1 through 9 and 11 through 25,
demand doubled in year 10. Level-1 hiring growth limited to no more than 50% per year, and
zero student admission cost. In addition to the jump in demand and limited hiring growth, we
eliminated the admission cost to increase the incentive to admit students in advance and thus
potentially undermine
27
 the four-year lookahead model.

Scenario 8 Fixed demand growth rate of 1.25% per year in years 1 through 9 and 11 through 25,
demand doubled in year 10. Level-1 hiring growth limited to no more than 50% per year, and
zero level-1 payroll cost. In addition to the jump in demand and limited hiring growth, we
eliminated the level-1 payroll cost to increase the incentive to admit students in advance and thus
potentially undermine the four-year lookahead model.

We compared the solutions obtained using the full information model and the lookahead model.
Figure 2 shows results for the baseline scenario and scenarios 1 through 5. In these scenarios, we obtained the
same solutions using the full information and the lookahead models. The lookahead model was robust
in these scenarios, even if Assumption 2.4 was slightly violated by our system’s parameters. Even in
Scenario 5, where education growth was drastically limited, the full information model did not differ from the
lookahead policy because training students a year in advance incurred extra payroll costs, making early
training more expensive than hiring. Though education growth was limited, hires served as a back-up
action in Scenario 5 and made the lookahead and the full information methods operate identically.
Figure 3 shows results for scenarios 6, 7, and 8; in this case, the lookahead policy resulted in
slightly higher total costs. Compared to the full information solution, the percentage differences in total
cost were only 0.026%, 0.129%, and 0.014% respectively. The lookahead model resulted in more
admissions, more level-2 hirings, and fewer level-1 hirings than the full information model. Since level-1
hiring was limited, fewer level-1 workers were hired and more students were trained as an alternative.
Level-2 workers were hired when the model reached a point where promotions could not meet the level-2
workforce demand due to the insufficient number of level-1 workers. The lookahead model failed to
anticipate future changes in demand, not training sufficient students nor hiring sufficient level-1 workers
in advance.
Overall, the lookahead policy showed robustness in the nine scenarios modeled. In the most
extreme scenarios, where demand had a sudden jump and hirings or admissions were limited, the
lookahead policy and the full information policy still showed very little difference, particularly in total
cost.

28
 Admissions and Hirings Breakdown
Baseline scenario and scenarios 1 through 5
100000

90000

80000

70000

Number of Admissions/Hirings 60000

50000
# Lv.2 Worker
40000 Hirings

30000 # Lv.1 Worker

Hirings
20000

10000 # Student
Admissions
0

Scenarios/Policies

Figure 2: Breakdown of the total number of admissions and hirings in baseline scenario and scenarios
1 through 5 over the course of 20 years.

Admissions and Hirings Breakdown

scenarios 6 through 8
237000

236000

235000

234000
Number of Admissions/Hirings

# Lv.2 Worker

233000 Hirings

# Lv.1 Worker
232000
Hirings
231000
# Student

Admissions

230000

229000

228000

Scenario 6 Scenario 6 Scenario 7 Scenario 7 Scenario 8 Scenario 8

Full Info Lookahead Full Info Lookahead Full Info Lookahead

Scenarios/Policies

Figure 3: Breakdown of the total number of admissions and hirings in scenarios 6 through 8 over the course
of 20 years.
29
4.2 Experiments with stochastic demand growth

To further evaluate the lookahead policy, we examined the performance of our model in a stochastic setting
where the demand growth rate in each year (denoted ρ) is an i.i.d. random variable uniformly distributed
between 0% and 2.5% (the mean growth rate is thus kept at 1.25%, as in the deterministic baseline scenario
[39]).

We applied the lookahead policy sequentially. After the simulated demand dt is realized in year t, we

project year (t + 1)’s demand to be dˆt+1 = 1 +(1 + δ )E[ρ] dt, where δ is a forecast factor used to
represent the planner’s level of risk-aversion. When δ > 0, the planner assumes demand grows faster
than the mean;
for δ < 0, the planner assumes the demand grows slower than the mean; for δ = 0, the planner plans
for the expected growth. After solving the lookahead model for years t and t + 1, the process steps forward
one year, true demand in t + 1 is observed, and hiring decisions are made if the workforce is insufficient to
meet the demand. The algorithm proceeds to the next period and the look-ahead policy is sequentially
applied.
This procedure iterates until period 20. In our simulation, each policy was solved with 2000 replications.
We considered two benchmarks for the lookahead policies. The first is the full information model;
as in the deterministic experiments, the full information solution solves a single LP with full
(deterministic) access to the uncertain parameters. In the stochastic case, this implies solving one full
information LP for every simulated replication and averaging the resulting costs. Because this solution has
earlier access to the uncertain data, it provides a lower bound on any policy’s cost.
In addition, we included as a second benchmark a na¨ıve policy implemented without resorting to our
LP. This policy sets workforce level targets for the current period based on demand or incoming
level-1 workforce, whichever is greater, and meets these levels by promoting as much as possible
before hiring. The policy then determines student admissions using the following simple rule: If the
level-1 workforce exactly meets demand, admissions are scaled up from the previous year based on expected
demand growth. Conversely, if the level-1 workforce exceeds demand, admissions are scaled down by the
same percentage that the workforce exceeds demand by. For example, if workforce is 105% of demand,
admissions are set to 95% of the previous year’s number.

As shown in Figure 4, by varying the forecast factor δ over 1% increments between −100% and 100%,
the lookahead policy achieves lowest cost at δ ∗ = −33% (the best delta policy). All lookahead

30
 policies

31
 ∗
were within 1.3% of the full information total cost (within 0.8% at δ = −33%). These cost
differences appear small because payroll cost makes up more than 90% of the total. In particular,
there is a fixed
amount of unavoidable payroll cost needed to satisfy demand, regardless of any decisions. By subtracting
the unavoidable payroll cost from the total cost, we are left with the controllable costs, i.e. promotion cost,
hiring cost, admission cost, and payroll cost in excess of the unavoidable. The lookahead policies
are

all within 42% of the full information controllable cost, and the policy with δ ∗ = −33% is within 26%.
Conversely, the na¨ıve heuristic exceeds the full information benchmark by 1.4% in total cost, and by
46% in controllable cost. The extreme lookahead policies (with δ = ±100%) perform only slightly
better than the na¨ıve heuristic, 42% versus 46% above full information in controllable costs, while the
best lookahead
policy is only 26% above, indicating that an intelligent choice of forecast factor can significantly
impact the policy’s performance: Based on our cost estimates, the use of a lookahead policy with the best
forecast factor would represent annual savings of up to $14.8 million per year compared to the na¨ıve
heuristic, or up to $252 million over the 20-year planning horizon, assuming a discount factor of 0.95.

1.6% 60.0%
% difference in total costs with respect to fulll

% difference in controllable costs with respect

1.4%
50.0%
1.2%
to fulll information model

40.0%
information model

1.0%

0.8% Naive 30.0% Naive

heuristic heuristic
0.6% Lookahead Lookahead
20.0%
0.4%
10.0%
0.2%

0.0% 0.0%
-100% -50% 0% 50% 100% -100% -50% 0% 50% 100%
Delta Delta
(a) Total Cost (b) Controllable Cost

Figure 4: Gaps for forecast factor δ lookahead policy and na¨ıve heuristic compared to full information
cost, in terms of (a) total cost and (b) controllable cost.

∗
In our simulation, δ is less than 0. This implies that it is more favorable to adopt a policy that
plans for demand growth smaller than the mean. To explain the rationale behind this behavior, Figure
5 shows the breakdown of the total cost as a proportion of the full information cost, and Figure 6 shows
32
a similar breakdown in terms of controllable cost. The model assumes the workforce cannot be downsized,
and pay-

33
 No training

Naïve heuristic

Plan for highest growth (δ = 100%)

Plan for no growth (δ = -100%)

Plan for the mean (δ = 0)

Payroll cost
Admission Cost
Best delta (δ = -33%)
Hiring Cost
Promotion Cost
Full information
92% 94% 96% 98% 100% 102% 104% 106%

Figure 5: Breakdown of total cost under different policies: payroll costs dominate the other costs.

No training

Naïve heuristic

Plan for highest growth (δ = 100%)

Plan for no growth (δ = -100%)

Plan for the mean (δ = 0)

Payroll Cost in Excess
Admission Cost
Best delta (δ = -33%) Hiring Cost
Promotion Cost
Full information

0% 50% 100% 150% 200% 250% 300%

Figure 6: Breakdown of controllable cost under different policies: look-ahead policies stay within 40% from
full information model with minimal gap of 22% at δ ∗ = 33%.

roll cost makes up more than 90% of the total. Therefore, an oversized workforce will remain in the
system many years and will thus increase the cost dramatically. To further explore this idea, we also
simulated a policy with no training that directly hires 100% of its workforce. However, the no-training policy
performed far worse, with a gap of 4.7% in terms of total cost or 150% in controllable cost.
Furthermore, the full in- formation solution does not have any excess hiring costs, as these can be completely
avoided with complete access to information; therefore, it is likely that the true optimal cost is closer to our
best lookahead policy than to this lower bound.

34
5 Conclusions

This paper contributes a new modeling framework for strategic health workforce planning. Through infinite-
horizon optimization, we are able to model the long-term implications of training, hiring and promotion
decisions made within a health care system. Our results (cf. Theorem 3.3 and Corollary 3.4) indicate
that very short planning horizons suffice to determine optimal decisions when employing the model in a
rolling horizon framework. We obtained these results under a set of mild assumptions (Assumptions 2.1
through 2.6) that correspond to reasonable real-world conditions, such as the requirement that payroll
increase in the workforce hierarchy. Using real-world data from British Columbia, we further demonstrate
how lookahead policies perform well in a variety of situations that generalize our base model,
specifically in the case of uncertain demand growth. These results are particularly useful, as the lookahead
solution mirrors workforce management policies that could be implemented in practice. Models such as the
one we propose can be used to obtain qualitative checks on whether a particular health workforce system is
behaving optimally, or what conditions it must meet to do so. For example, in [51] the authors apply a
similar model to derive policy recommendations for the U.S. pediatric nurse practitioner workforce.
A next step in our work is to directly model and optimize the system’s uncertainty, specifically in demand
growth or retention rates [42]. It is important to understand whether the conditions we develop in this work
and their structural consequences (or appropriate modifications) still hold in more general settings. The more
nuanced analysis required in this case may give insight into the impact of uncertainty on health workforce
costs and management decisions; for example, [22, 33] investigate similar questions for short-term nurse
staffing. From a theoretical perspective, the infinite linear programming tools we use still apply in
the presence of uncertain demand growth or retention, provided these can be modeled as finite-support
random variables.
Because this work is applied to guide strategic health workforce decisions, we can formulate
more realistic models by incorporating other elements. For instance, (1) could be expanded to include a
variety of health care providers and changes in scopes of practice. Assuming that the interaction of all
worker types with demand is linear, multiple worker types can be incorporated in models similar to (1), by
differentiating across both type and level, where each worker type includes its own hierarchy with its
own supervision constraints (1c) and dynamics, but the multiple types serve patient demand jointly.

35
 By providing an initial understanding of this infinite-horizon model, our goal is to move a step forward
in the field of strategic health workforce planning, and to motivate others to continue doing research in
this important application.

Acknowledgments

The authors thank the associate editor and two anonymous referees for their valuable comments and sug-
gestions.

References

[1] W.J. Abernathy, N. Baloff, J.C. Hershey, and S. Wandel. A Three-Stage Manpower Planning and
Scheduling Model–A Service-Sector Example. Operations Research, 21:693–711, 1973.

[2] D. Adelman. Price-Directed Replenishment of Subsets: Methodology and its Application to Inventory
Routing. Manufacturing and Service Operations Management, 5:348–371, 2003.

[3] D. Adelman. A Price-Directed Approach to Stochastic Inventory/Routing. Operations Research,

52:499–514, 2004.

[4] D. Adelman and D. Klabjan. Duality and Existence of Optimal Policies in Generalized Joint
Replen- ishment. Mathematics of Operations Research, 30:28–50, 2005.

[5] D. Adelman and D. Klabjan. Computing Near-Optimal Policies in Generalized Joint Replenishment.
INFORMS Journal on Computing, 24:148–164, 2011.

[6] H.S. Ahn, R. Righter, and J.G. Shanthikumar. Staffing decisions for heterogeneous workers with
turnover. Mathematical Methods of Operations Research, 62:499–514, 2005.

[7] E.J. Anderson and P. Nash. Linear Programming in Infinite-Dimensional Spaces. John Wiley &
Sons, Inc., Chichester, 1987.

[8] A. Arlotto, S.E. Chick, and N. Gans. Optimal Hiring and Retention Policies for Heterogeneous
Work- ers who Learn. Management Science, 2013. Forthcoming.

36
 [9] W. Balinsky and A. Reisman. Some Manpower Planning Models Based on Levels of Educational
Attainment. Management Science, 18:691–705, 1972.

[10] J.F. Bard and L. Wan. Workforce design with movement restrictions between workstation groups.
Manufacturing and Service Operations Management, 10:24–42, 2008.

[11] D.J. Bartholomew, A.F. Forbes, and S.I. McClean. Statistical Techniques for Manpower Planning.
John Wiley & Sons, Inc., second edition, 1991.

[12] S. Batun and M.A. Begen. Optimization in healthcare delivery modeling: Methods and
applications. In Handbook of Healthcare Operations Management, pages 75–119. Springer, 2013.

[13] BC Statistics. Population Estimates, British Columbia. Available on-line at http://www.bcstats.

gov.bc.ca, April 2013.

[14] D. Bienstock and A.C. Zenteno. Models for managing the impact of an epidemic. Working pa-
per, Department of Industrial Engineering and Operations Research, Columbia University. Version

Sat.Mar.10.153610.2012. Available at http://www.columbia.edu/~dano/papers/bz.pdf, 2012.

[15] J. Boudreau, W. Hopp, J.O. McClain, and L.J. Thomas. On the interface between operations and human
resources management. Manufacturing and Service Operations Management, 5:179–202, 2003.

[16] M.L. Brandeau, D.S.P. Hopkins, and K.L. Melmon. An integrated budget model for medical
school financial planning. Operations Research, 35:684–703, 1987.

[17] J. Campbell, G. Dussault, J. Buchan, F. Pozo-Martin, M. Guerra Arias, C. Leone, A. Siyam, and
G. Cometto. A Universal Truth: No Health Without a Workforce. Technical report, Global
Health Workforce Alliance and World Health Organization, Geneva, 2013. Forum Report, Third
Global Fo- rum on Human Resources for Health, Recife, Brazil.

[18] T. Cheevaprawatdomrong, I.E. Schochetman, R.L. Smith, and A. Garcia. Solution and Forecast
Hori- zons for Infinite-Horizon Nonhomogeneous Markov Decision Processes. Mathematics of
Operations Research, 32:51–72, 2007.

37
[19] W.P. Cross, H.E. Romeijn, and R.L. Smith. Approximating Extreme Points of Infinite
Dimensional Convex Sets. Mathematics of Operations Research, 23:433–442, 1998.

[20] M.R Dal Poz, N. Gupta, E. Quain, and A.L.B. Soucat, editors. Handbook on Monitoring and Evalua-
tion of Human Resources for Health with Special Applications for Low and Middle-Income
Countries. World Health Organization, 2009.

[21] T.M. Dall, P.D. Gallo, R. Chakrabarti, T. West, A.P. Semilla, and M.V. Storm. An Aging Population
And Growing Disease Burden Will Require A Large And Specialized Health Care Workforce By 2025.
Health Affairs, 32:2013–2020, 2013.

[22] A. Davis, S. Mehrotra, J. Holl, and M.S. Daskin. Nurse Staffing Under Demand Uncertainty to
Reduce Costs and Enhance Patient Safety. Asia-Pacific Journal of Operational Research, 2013.
Forthcoming.

[23] F. de Ve´ricourt and O.B. Jennings. Nurse staffing in medical units: A queueing perspective.
Operations Research, 59(6):1320–1331, 2011.

[24] J.E.B. Duchscher. Transition shock: the initial stage of role adaptation for newly graduated registered
nurses. Journal of Advanced Nursing, 65(5):1103–1113, 2009.

[25] A.T. Ernst, H. Jiang, M. Krishnamoorthy, B. Owens, and D. Sier. An Annotated Bibliography of
Personnel Scheduling and Rostering. Annals of Operations Research, 127:21–144, 2004.

[26] M.J. Fry, M.J. Magazine, and U.S. Rao. Firefighter staffing including temporary absences and wastage.
Operations Research, 54(2):353–365, 2006.

[27] N. Gans, G. Koole, and A. Mandelbaum. Telephone call centers: Tutorial, review, and research
prospects. Manufacturing and Service Operations Management, 5:79–141, 2003.

[28] N. Gans and Y.P. Zhou. Managing learning and turnover in employee staffing. Operations
Research, pages 991–1006, 2002.

[29] S.I. Gass, R.W. Collins, C.W. Meinhardt, D.M. Lemon, and M.D. Gillette. The army manpower
long- range planning system. Operations Research, 36:5–17, 1988.

38
[30] A. Ghate. Infinite horizon problems. In J.J. Cochran, L.A. Cox, P. Keskinoc¸ak, J.P. Kharoufeh, and
J.C. Smith, editors, Wiley Encyclopedia of Operations Research and Management Science. John
Wiley & Sons, 2010.

[31] A. Ghate, D. Sharma, and R.L. Smith. A Shadow Simplex Method for Infinite Linear Programs.
Operations Research, 58:865–877, 2010.

[32] A. Ghate and R.L. Smith. Characterizing extreme points as basic feasible solutions in infinite linear
programs. Operations Research Letters, 37:7–10, 2009.

[33] L.V. Green, S. Savin, and N. Savva. Nursevendor problem: Personnel Staffing in the Presence of
Endogenous Absenteeism. Management Science, 2013. Forthcoming.

[34] R.C. Grinold. Manpower planning with uncertain requirements. Operations Research, 24:387–399,
1976.

[35] P.C. Jones, J.L. Zydiak, and W.J. Hopp. Stationary dual prices and depreciation. Mathematical Pro-
gramming, 41:357–366, 1988.

[36] R. Kocher and N.R. Sahni. Rethinking Health Care Labor. New England Journal of Medicine,
365:1370–1372, 2011.

[37] M.S. Lavieri. Nursing Workforce Planning and Radiation Therapy Treatment Decision Making: Two
Healthcare Operations Research Applications. PhD thesis, University of British Columbia, 2009.

[38] M.S. Lavieri and M.L. Puterman. Optimizing Nursing Human Resource Planning in British Columbia.
Health Care Management Science, 12:119–128, 2009.

[39] M.S. Lavieri, S. Regan, M.L. Puterman, and P.A. Ratner. Using Operations Research to Plan the
British Columbia Registered Nurses Workforce. Health Care Policy, 4:113–131, 2008.

[40] H.L. Lee, W.P. Pierskalla, W.L. Kissick, J.H. Levy, H.A. Glick, and B.S. Bloom. Policy decision
modeling of the costs and outputs of education in medical schools. Operations research, 35:667–683,
1987.

39
[41] A. Martel and W. Price. Stochastic Programming Applied to Human Resource Planning. Journal
of the Operational Research Society, 32:187–196, 1981.

[42] D. Masursky, F. Dexter, C.E. O’Leary, C. Applegeet, and N.A. Nussmeier. Long-term forecasting
of anesthesia workload in operating rooms from changes in a hospital’s local population can be
inaccurate. Anesthesia and Analgesia, 106:1223–1231, 2008.

[43] J. Needleman, P. Buerhaus, S. Mattke, M. Stewart, and K. Zelevinsky. Nurse-staffing levels and
the quality of care in hospitals. New England Journal of Medicine, 346(22):1715–1722, 2002.

[44] E.J. Pinker and R.A. Shumsky. The efficiency-quality trade-off of cross-trained workers.
Manufactur- ing and Service Operations Management, 2:32–48, 2000.

[45] M. Puterman. Markov Decision Processes: Discrete Stochastic Dynamic Programming. John
Wiley & Sons, Inc., 2005.

[46] P.P. Rao. A Dynamic Programming Approach to Determine Optimal Manpower Recruitment Policies.
Journal of the Operational Research Society, 41:983–988, 1990.

[47] H.E. Romeijn, D. Sharma, and R.L. Smith. Extreme point characterizations for infinite network
flow problems. Networks, 48:209–222, 2006.

[48] H.E. Romeijn and R.L. Smith. Shadow Prices in Infinite-Dimensional Linear Programming.
Mathe- matics of Operations Research, 23:239–256, 1998.

[49] H.E. Romeijn, R.L. Smith, and J.C. Bean. Duality in Infinite Dimensional Linear Programming.
Math- ematical Programming, 53:79–97, 1992.

[50] E. Salsberg and A. Grover. Physician workforce shortages: Implications and issues for academic health
centers and policymakers. Academic Medicine, 81(9):782–787, 2006.

[51] G.J. Schell, X. Li, M.S. Lavieri, A. Toriello, K.K. Martyn, and G.L. Freed. Strategic Modeling of the
Pediatric Nurse Practitioner Workforce: How Policy Changes Can Yield Self-Sufficiency.
Pediatrics, 135:298–306, 2015.

40
[52] D.P. Schneider and K.E. Kilpatrick. An optimum manpower utilization model for health maintenance
organizations. Operations Research, 23:869–889, 1975.

[53] T.C. Sharkey and H.E. Romeijn. A simplex algorithm for minimum-cost network-flow problems
in infinite networks. Networks, 52:14–31, 2008.

[54] H. Song and H.C. Huang. A successive convex approximation method for multistage workforce
ca- pacity planning problem with turnover. European Journal of Operational Research,
188(1):29–48, 2008.

[55] J. Turner, S. Mehrotra, and M.S. Daskin. Perspectives on Health-Care Resource Management
Prob- lems. In M.S. Sodhi and C.S. Tang, editors, A Long View of Research and Practice in
Operations Research and Management Science, volume 148 of International Series in Operations
Research & Management Science, pages 231–247. Springer, New York, NY, 2010.

[56] U.S. Census Bureau. U.S. Population Projections. Available on-line at http://www.census.gov/
population/www/projections/, January 2012.

[57] U.S. Department of Health and Human Services. National center for health workforce analysis. http:
//bhpr.hrsa.gov/healthworkforce/, January 2012.

[58] G. Vairaktarakis and J.K. Winch. Worker cross-training in paced assembly lines. Manufacturing
and Service Operations Management, 1:112–131, 1999.

[59] S.A. Vanderby, M.W. Carter, T. Latham, and C. Feindel. Modeling the Future of the Canadian Cardiac
Surgery Workforce using System Dynamics. Working paper, 2013.

[60] S.A. Vanderby, M.W. Carter, T. Latham, M. Ouzounian, A. Hassan, G.H. Tang, C.J. Teng, K. Kings-
bury, and C.M. Feindel. Modeling the Cardiac Surgery Workforce in Canada. The Annals of
Thoracic Surgery, 90:467–473, 2010.

[61] G.R. Werker and M.L. Puterman. Strategic Workforce Planning in Healthcare Under
Uncertainty. Working paper, 2013.

41
 [62] N. Yankovic and L.V. Green. Identifying good nursing levels: A queuing approach. Operations
Research, 59(4):942–955, 2011.

[63] P.H. Zipkin. Foundations of Inventory Management. McGraw-Hill, 2000.

A Proof of Lemma 3.2

All the arguments below apply to solutions that satisfy Lemma 3.1.

A.1 Proof when n = 2

A.1.1 Proof of (4b)

Assume a feasible solution is given for which (4b) is violated in some period. We start from the earliest such
period, relabeling it as period 1 without loss of generality, and make the following changes:

−ε, t=

t p2+p1q1,2
∆x ,2 = ε, t=2 ,
1
1 1+q1,2

 (p1+p2q1,2)t−
3 (p 1 — p2)2q 1,2 ε, t = 3, . .
 (1+q1,2)t−1 .
ε, t=1
∆st = 1 ,
(p(1+q
1
+p q
2
)t−2
t−1 (p1 − p2)ε, t = 2, . . .
 1,2 1,2)


−ε, t=1
∆s2t = .
(p 1+p2q )t−2
t−1 (p1 − p2)q1,2ε, t = 2, . . .
1,2 (1+q1,2)

The resulting solution is feasible for small positive ε. Furthermore, we achieve an objective
improve- ment

∞
!
∆C = p2 + (p1 + p2q1,2)t−3
1 γ q p 2 t−1
γ ε
p1q1,2
c
(p1 − p2)γ ε
1,2 — + )+
t−1p 1,2( 1 2 )
1 + q1,2 (1 + )t− −
∞ ∑ q1,2 1 !
(p1 + p2q1,2)t−2
+h 1+ ∑ t−1
t=
(1 +
1
2 q1,2
42
 ∞
!
(p1 + p2q1,2)t−2
+h −1 + q1,2 ∑ (1 + (p1 − p2)γ t−1 ε
) t−1
t=2
2 (1 + q1,2)(1 − p1qγ1,2
)(p2γ − 1)
= c1,2 ε
1 + q,21 − (p1 + 2p q1,2 )γ
(1 + q1,2)((1 − γ p2)h1 − (1 − γ p1)h2)
+ ε
1 + q1,2 — (p1 + 1,2)γ
p2q
< 0,

where the last inequality follows by Assumption 2.5.

The rationale behind the construction is to choose training and promotion perturbations so that staff at
the two levels increase proportionally in later periods, which implies feasibility; on the other hand, the
cost decrease exceeds the increase when discounts and monotonic payrolls are applied, which leads to the
lower objective.

A.1.2 Proof of (4a)

Since (4b) can be achieved without resorting to (4a), we consider the solutions where (4b) is satisfied while
0
(4a) is not. Again we rename the earliest such period to be period 1. We have
0 x > 0.

2
Case 1: x1,2 < p1s11.

Construct a new feasible solution with the formulas below:



− ε , t=0
p0

∆xt = 
0
p1
ε, t=1 ,
p0


0, t = 2, . . .



−ε, t = 1
∆s1t =  .
0, t = 2, . . .

The resulting objective improvement is

1 p1
∆C = c0 − p0γ + p0 ε − h1ε
43
 < 0.

2
Case 2: x1,2 = p1s11.

We first note that at most one of xt+1 = p1st and st+1 = dt+1 can be true provided (4b) for any t ≥ 1.
1,2 1 1
Assuming both equalities hold for some t, we then have st+1 ≥ p2st + xt+1 = p2st + p1st . Since xt+1 > 0
2 2 1,2 2 1 1,2

implies st+1 = 2 1
t +p1dt
p2q1,2qd1,2 2 q1,2 t
st+1 by (4b), we further have 1t+1 p2sqt 1,2
q 2 1,2 1 =s ≥ ≥ =
d +p1st + p1
d , but
p
t+
this contradicts Assumption 2.6. 1

Let i be the smallest possible period with xi < p1si−1. From the above observation we have st > dt,t ≤
1,2 1 1

i − 1. Thus we can perturb as follows to obtain a new feasible solution:


ε
− p0 , t=
 −p1(p1+p q )t−1
p0q2t 1,2 ε, t = 1, . . . , i
∆x0t = 0 1,2
i−2 ,
 p1(p1p+p 2q1,2)
i−2 ε, t=i−1
−2 0q1,2


0, t = i, . . .


−p1ε, t=2
 p2(p1+p2q1,2)t−3
—
1 ε, t = 3, . . . ,i −
qt−2
∆x t1,2
1,2
1 ,
=  p1 p2(p1+p2q1,2)
i−
3
ε, t=i
qi−3
1,2


0, t = i + 1, . . .

−ε, t=1

q1,2)t−2
∆st1 = − p1(p1+p2t−1 ε, t = 2, . . . , i −
q1,2

1, 
0, t = i, . . .

  p1(p1+p2q1,2)
t−2
 − t−
q2 1,2 ε, t = 2, . . . , i − 1
∆s2 =
t
.

0, t = i, . . .

44
 The corresponding objective improvement is

!
i−2
∆C = 1 p1(p1 + p1(p1 + p2q1,2) i−2 i−2
γ
t−1 t−1
p2q1,2) γ ε
! pi−2
t−2γt−1 +
c i−1 q
0 — p (p + p q
∑ 1 1 p20q1,2
− t=1 )
t1,2 0 1,2
p0γ
+h1
−1 − ∑ t−1
ε
t=2
q1,2
!
i−1
+h2 p1(p1 + p2q1,2)t−2γt−1
—∑ t−2
ε
t= q1,2
2
!
i−1 2
+c — p1γ — ∑ p (p1 + p1 p2(p1 + p2q1,2)i−3γi−1
ε
1,2 p q 1 )t−3γt−1
t=
3 2 1,2 +
t 2 i3
q−1,2 q−1,2 !
i−2 i−2
i−2 i−2 i−2 p12 t−1 pi−2−t
p1(p1 + p2q1,2) γ
γi−2 + ε
p1 p2p γ
∑ (p1 + p2qp1,20q)1,2
2
< c0 — 0 — t=1 t i—2
p0q 1,2
!
+c — p1
i−1 p12(p1 + p2q1,2)t−3 pi−t
2 γi−1
+ p1 p2(p1 + p23q1,2)i−3γi−1
2
1,2 i−2 i−1 q−i ε
p γ
—∑ 2
t
= 0. t= q−1,2 1,2
3

A.2 Problem structure when n ≥

3

Given a solution that violates Lemma 3.2, our goal is to construct a new solution that is both feasible

and incurs a lower total cost. While the big picture appears similar to the proof when n = 2, things are

much
more complicated here: the effect of may not end at level k + 1; instead it can force and thus
∆st k
xt+2
k+1,k+2
st+2 to change, which will propagate to higher levels; even worse, lower levels may also be influenced since
k+
there
2 may be multiple violated levels and the perturbation may not start from level 1. Therefore, it is unlikely
that we can rely on one-time substitutions as before.
Instead, our strategy is to construct a perturbation period by period. To develop such a dynamic approach

45
we first introduce three sets of new variables:
t
• rt = s
k , ∀ k = 0, . . . , n, ∀ t = 1, . . .,
k pt
t
• zt = x
k , ∀ k = 0, . . . , n, ∀ t = 1, . . .,
k pt

xt
t
• zk,k+1 = k,k+1
, ∀ k = 1, . . . , n − 1, ∀ t = 1, . . ..
p
t

46
 The original problem can be reformulated as follows.

n−1 n
inf W (r z ) = ∞ n
ck zkt + ∑c
, h rt (8a)
γt−1 pt
∑ k=
∑ 1 ztk,k+1 k,k+1 +∑ k k
k= k=1
0
s.t. rt1 ≥ dt/pt ∀ t = 1, . . . (8b)
t
rk+ ≥ qk,k+1rt , ∀ k = 1, . . . , n − 1, ∀ t = 1, . . . (8c)
1
k
rt−1 − rt + zt−1 − zt + zt = 0, ∀ t = 1, . . . (8d)
1 1 0 1,2 1

rt−1 − rt + zt —t + zt = 0, ∀ k = 2, . . . , n − 1, ∀ t = 1, . . . (8e)
z
k k k−1,k k,k+1 k

rt−1 − rt + zt + zt = 0, ∀ t = 1, . . . (8f)
n n n−1,n n

t
k,k+1
z ≤ rt−1, ∀ k = 1, . . . , n − 1, ∀ t = 1, . . . (8g)
k
rt, zt ≥ 0, ∀ t = 1, ..., (8h)

where rt = zt for t = 0,.....The constraints above can be divided into three sets: demand/ratio constraints
0 0

(8b–8c), promotion bounds (8g), and network flow constraints (including flow conservation (8d–8f)
and nonnegativity (8h)). Graphically, if we consider the r variables as flows between successive periods
and the z variables as flows between successive levels, a feasible solution can be represented by an
infinite time-space network. The equivalence of the reformulated problem and the original problem
stems from a one-to-one correspondence between their solutions. Therefore, any result obtained from one
version applies to the other as well.
Next we identify four structural characteristics of our problem(s). Claim A.1 describes the cost of certain
structures and will help justify the superiority of a perturbed solution; Claim A.2 is a dominance property
and will enable us to consider a relatively small set of solutions for perturbation; Claims A.3 and A.4 analyze
necessary conditions for feasibility and will shed light on how to perturb.

Claim A.1. For the reformulated problem, any flow circulating counterclockwise (either in a cycle, on a
doubly-infinite path, or on a one-way infinite path) incurs negative cost.

47
Proof. zt , ∀ t = 1, . . . can be reduced to a common super source node representing level 0 in the network.
0
Define a basic unit in the grid-like network as either case below:

48
 t t+ t t+
k+1 1
1
1
k
.
0

The corresponding total costs per unit counterclockwise flow are

γt−1 pt(hk − hk+1) + ptck,k+1(pγt − γt−1) < 0,

−γt−1h1 pt + (−γt−1c0 pt + γtc0 pt+1) < 0,

respectively. We will refer to the two types of basic units as basic square and basic triangle,
respectively. Any cycle can be decomposed into a finite number of basic squares and/or triangles; any
doubly-infinite path can be decomposed into a countable number of basic squares; and any one-way
infinite path can be decomposed into a countable number of basic squares and/or triangles. Since
counterclockwise flows around both basic units incur negative costs, the same is true for arbitrary
cycles/infinite paths.

Claim A.2. For t ≥ 1, let lt and lt+1 be levels such that st

> qtlt ,lt +1st , > 0, and st+1 >
x
lt +1 lt lt ,lt lt+1+1
+1

qlt+1 . Assuming lt and lt+1 exist, for any k with min{lt, lt+1} ≤ k ≤ max{lt, lt+1} there exists
t+1
,lt+1 +1s l
t+1
′
some
′ t ≤ t such that > 0. A solution cannot be optimal if both st+1 = = 0 hold
xt q st+1
t+1 and
k,k+1 l +1
x
lt ,lt +1 k,k+1
t
lt
for some k with min{lt, lt+1} ≤ k ≤ max{lt,
lt+1}.

Proof. Clearly lt /= lt+1. Consider level i, the largest such k if lt > lt+1, or the smallest such k if lt < lt+1.

Case 1: i = lt.

t+1 t t
Let lt ,lt = = ε. Since xt+1 = 0, we know xt+1 < and thus feasibility is not
lt ,lt lt
∆z +1 −∆z +1 ps +1
lt ,lt lt +1,lt
violated. By Claim A.1 this corresponds to a counterclockwise+2flow
+1 around a basic square and incurs
less total cost.

49
Case 2: i /= lt.

We first have

st+1 = —t+1 ≤ pst ,

pst x
i+1 i+ i+1,i+2 i+1
1

50
 st+1 ≥ pst + xt+1 ≥ ps .
t

i i i−1,i i

If lt > lt+1, then xt+1 > 0 by definition of i, and hence st+1 < pst , which together with st+1 ≥
i+1,i+2 i+1 i+ i+1
1
qi,i+1st+1 indicates that st > qi,i+1st. Similarly, if lt < lt+1, then xt+1 > 0 and again st+1 > pst indicates
i i+1 i i−1,i i i
that st t
> qti,i+1s . If = 0, then s t
> qi,i+1s further indicates that s
t t−1
> qi,i+1s . Recursively
t−1
x
i+1 i i+1 i i+1 i

i,i+1
utilizing this fact for t − 1,t − 2, . . ., finally we can find a period t0 ≥ t′ (t0 = t if xt > 0) where st0 >
t i,i+1 i+
qti,i+1
0
s 0 and > 0. Now construct a new solution by letting ∆z t+1
= −∆z t0
= ε; it is feasible
1 due
x
i i,i+1 i,i+1 i,i+1
to slack and zero promotions at level i in periods t0 + 1, . . . , t, and its lower cost is guaranteed by Claim
A.1.

Claim A.3. For t ≥ 1, For t ≥ 1, let g be a level such that xt+1 = 0, g ≤ n. If st+1 = qg−1,gst+1 = . . . =
g,g+1 g g−1
q1,gst+1, where qk,l = qk,k+1qk+1,k+2 · · · ql−1,l, then at most one of xt+1 =t and s t+1
= dt+1 can be true
ps
1 k,k+1 k 1
for each k ≤ g −
1.

Proof. Assume both equalities hold. Adding together equations (1e) for levels k + 1, . . . , g in period t + 1,
plugging in xt+1 = pst and st+1 = qg−1,gst+1 = . . . = q1,gst+1, we have
k,k+1 k g g−1 1

t+
1 1
g g
s ∑ q1,i ≥ ps1 ∑ q1,i.
t
(9)
i=k+ i=k
1

Note that

∑g q1,k
q1,i
i=k = +1
g g
∑ i=k+ q1,i ∑i=k+
q1,i
1
1 1
= +1
∑gi=k+1 qk ,i
1
≥ g−k
+1
∑ i=11qi+
max
∑g−k qi
i=1 max
= .
51
 g−k−1 i
qmax(1 + ∑ q )
i=1 max

Combined with (9), this results in

1 + ∑g−k qi p p
i=1 max pst st dt ≥ dt+1
st+1
1
≥ > ≥
g−k−1 i
qmax(1
1 +∑ q ) q1 max qmax
i=1 max

52
by Assumption 2.6. We have arrived at a contradiction.

Claim A.4. For t ≥ 1, let k be an arbitrary level, and g = min{i : = 0, k + 1 ≤ i ≤ n}. If = pst
xt+1 i,i+1 t+1
x
k,k+1 k
and
t+1 s
t+1
= qg−1,gs = . . . = ql+1,gs , 0 ≤ l ≤ k − 1, then
t+1 t+1
> 0.
x
g g−1 l+1 l,l+1

Proof. The result follows directly by Lemma 3.1 if xt+1 > 0. Consider when xt+1 = 0. l = k − 1 is trivial
l+1 l+1
since xt+1 = st+1 > 0. For l ≤ k − 2, we show that > pql+1,kst by induction.
xt+1
k−1,k k l,l+1 l+1

Base case: Here l = k − 2. st+1 = —t st+1 = — qk−1,kst+1 implies st+1 =

pst +t+1 ps +t+1
x x
pst
k−1 k− k−2,k−1 k k−1 k−2,k−1 k−1 k−1
+x t+
1 1
k−1 k−2,k−1
. qk−1,gst+1 = st+1 = pst + > pst ≥ t implies st+1 > pst . Therefore,
xt+1 pqk−1,gs
1+qk−1,k k−1 g g g−1,g g k−1 k−1 k−1
pst t+1
+xk−2,k−1 pq
, and thus xt+1 t k−1,ks > 0.
>
k−1
> pst
1+qk−1,k k−1 k−2,k−1 k−1

Induction: Assume that the claim holds for l, 1 ≤ l ≤ k − 2, i.e.

l,l+1
> pql+1,kstl+1 , then
xt+1

st+1 = pst + xt+1 — xt+1

l l l−1,l l,l+1
< pst + xt+1 pql+1,kst
—
l l−1,l l+1
< ps + x t t+1
− pql,ks . t

l l−1,l l

On the other hand, ql,gst+1 = st+1 > pst ≥ pql,gst implies st+1 > pst . Therefore, xt+1 > pql,kst > 0,
l g g l l l l−1,l l
i.e. the claim holds for l − 1 as well.

A.3 A perturbation procedure for n ≥ 3

(4a) is a special case of (4b) if we define st = dt, xt = pxt−1, q0,1 = 1; the only difference is that
xt have
0 0,1 0 0,1

no upper bound. Pick the earliest period where (4b) is violated, as before we rewrite it as period 1
and redefine sucessive periods as 2, 3, Let m be any violated level in period 1. Below is the key notation
we
will use:

53
• jt: a level that has ever seen staff
′
reduction and has full promotions in period t; mathematically this
means x t
= ps and ∆s < 0 for some t′ < t.
t−1 t

jt , jt +1 jt jt

• lt: a level with ratio slack and for which there is a jt where all levels in between have positive
promotions in period t, i.e. and xtk,k+1 > 0, ∀ k = lt, . . . , jt.
st lt > qlt ,lt lt
t
+1 +1s

54
 • l¯t : an lt where the ratio relationship would be violated if not perturbed in period t + 1, i.e.l¯st+1 +
t
p∆st < q ¯ (st+1
t+1
+ p∆st ) and l¯t ∈ Lt. An implicit constraint is = q ¯ st+1. +1
¯ s ¯
l¯t +1 lt ,lt l l¯t l¯t +1 lt ,lt +1 l¯t
+1 ¯t

• Jt, Lt, L¯ t : the set of all jt, the set of all lt, and the set of all l¯t , respectively. L¯ t ⊆ Lt.

• max jt: the largest element in Jt, i.e. max{ j : j ∈ Jt}, with other maxima and minima defined
analo- gously.

Suppose we have perturbed periods 1, . . . , t − 1 and the current perturbed solution satisfies constraints in
these periods. Clearly all jt should be perturbed to guarantee feasibility. By Claims A.3 and A.4 there
must exist an lt for each jt, and thus it is a candidate for the perturbation in period t to stop. The
perturbation in period t − 1 also causes infeasibility at l¯t−1 in period t, which constitutes an additional
source for further perturbation. Obviously we can set J1 = L1 = {m}, L¯ 0 = ∅. Procedure 1
illustrates how to identify Jt, Lt and L¯ t −1 when t ≥ 2.

Procedure 1 Jt, Lt, and L¯ t −1 when t ≥ 2

1: Jt = Lt = L¯ t −1 = ∅,t− = {lt−1 : lt−1,lt−1+1 < 0}
t−1
′
L 1 ∆x
2: for lt−1 ∈ Lt−1 do
3:
t
if lt−1+1 + t−1
lt−1+1
< qlt−1 ,lt−1 (s tlt−1 + p∆st−1 ) then
+1 lt−1
s4: L¯ t−1 = L¯ t−1 ∪ {lt−1}
p∆s
5: end if
6: end for S
S ′
t′≤t−1 Jt ′ ∪ Lt′ } + 1
′
7: for k = min{i : i ′ J
t ≤t−1 t ′ ∪ L t ′ } to max{i :
t i ∈∆st′ < 0 for some t′ < t do
∈ if xk,k+1
8: = pskt−1 and k
9: Jt = Jthen
t ∪ {k}
10: g = min{i : i,i+1 = 0, k + 1 ≤ i ≤ n}
t+1
x
11: for i ≤ g − 1
do
12: if st > qi,i+1st and xt > 0, ∀ k = i, . . . , g − 1 then
i+1 i k,k+1
13: Lt = Lt ∪
{i}
14: end if
15: end for
16: end if
17: end for

We are now ready to construct perturbing operations. Since Lemma 3.1 and Claim A.2 have identified

55
several non-optimal cases, we only consider solutions that satisfy the conditions therein.
Perturbation for period 1 is trivial: ∆r1 = −∆r1 = ∆z1 = −ε. For an arbitrary t ≥ 2 before the
m+1 m m,m+1

possible end period, ∆ztk,k+1 consists of two parts: a change due to full promotions at jt, and a change due

56
to tight ratio relationships at l¯t−1. The calculation of ∆z consequently depends on the locations of jt, lt and
l¯t−1. If there is an l¯t−1 ∈/ Lt, then the levels between any pair of jt and l¯t−1 must have been reached at some
point before t, thus by Claim A.2 we know xt
k,k+1 > 0, ∀ k = l¯t−1, . . . , lt, ∀ lt ∈ Lt. Otherwise L¯ t−1 = ∅.
Hence although there may be multiple jt, lt and l¯t−1, and a large number of possible locations, it suffices to
check the following four cases:

lt
l¯t−1 − 1 l¯t−1 − 1 jt

l¯t−1 jt jt lt
jt lt
jt l¯t−1
lt
jt
jt
l¯t− jt jt
l¯t−1 + 1 1
l¯t−1 + 1 lt
lt

In the above graph, promotions are positive at levels connected by the vertical lines, a gap between levels
indicates zero promotions, and the direction of the arrows is consistent with the perturbation flows. Claims
A.5 through A.8 validate the operations we use in each case.

Claim A.5. Assuming a perturbed solution is feasible for periods 1, . . . , t − 1, if there exists an lt ∈ Lt with
lt ≥ max{max jt, max l¯t−1 + 1} and xt > 0 for k with min{min jt, min l¯t−1 + 1} ≤ k ≤ lt, then a solution
k,k+1

also feasible for period t can be obtained by sequentially applying equations

∆zt
= min{∆zt + ∆rt−1 − ∆rt qk−1,k, ∆rt−1}, (10a)
k,k+1 k−1,k k k−1 k

∆rt = ∆zt + ∆rt−1 − ∆zt , (10b)

k k−1,k k k,k+1

to k = kb, . . . , ke, and finally letting

t
t
∆r ke+1 = + ∆rt−1 , (10c)
ke,ke+1 ke+1
∆z

where kb = min{min jt, min l¯t−1 + 1}, ke = min{lt : lt ≥ max{max jt, max l¯t−1 + 1}}.

57
Proof. We can choose the smallest such lt as a common level for the perturbation driven by all jt and

l¯t−1 to stop in period t. To fix potential infeasibility caused by full promotions or tight ratio
relationships, it

is reasonable to decrease the promotions at these levels and update employment accordingly. The level to

start such operations should of course be min{min jt, min l¯t−1 + 1}. The decreasing effect will
finally be conveyed to the lt we choose.
We consider level k to determine ∆zt . Since ∆rt−1, ∆rt , and (in particular, ∆zt = 0)
∆zt
k,k+1 k k− k−1,k kb−1,kb
1
are all known as we reach node (k,t), we can solve


 ∆zt + ∆rt−1 =
+ ∆rt
∆zt
 k−1,k k k,k+1 k

∆rt ≥ ∆rt qk−1, ,
k
k k−1

 t
∆ k,k+1 ≤ ∆rkt−1

z
and choose the largest possible ∆zt (so that |∆zt | is as small as possible), which yields (10a). The
k,k+1 k,k+1
three constraints above represent flow conservation, ratio relationships, and promotion bounds respectively.
∆rt
k can then be determined via flow conservation, i.e. (10b). (10c) is a result of stopping perturbation at
ke.

Claim A.6. Assuming a perturbed solution is feasible for periods 1, . . . , t − 1, if there exists an lt ∈ Lt with
lt ≤ min{min jt, min l¯t−1 − 1} and xt > 0 for k with lt ≤ k ≤ max{max jt, max l¯t−1 − 1}, then a solution
k,k+1

also feasible for period t can be obtained by sequentially applying equations

∆zk,k+1
t
= min{∆zt
k+1,k+2 + ∆rk+
t
/qk+1,k+2 − ∆rk+1
t−1
, ∆rt−1
k }, (11a)
2

∆rt = ∆zt + ∆rt−1 − ∆zt , (11b)

k+1 k,k+1 k+1,k+2
k+
1

to k = kb, . . . , ke, and finally letting

∆rt = ∆rt−1 − ∆zt , (11c)

58
ke ke ke,ke+1

59
where kb = max{max jt, max l¯t−1 − 1}, ke = max{lt : lt ≤ min{min jt, min l¯t−1 − 1}}.

Proof. We can choose the largest such lt as a common level for the perturbation driven by all jt and

l¯t−1 to stop in period t. Again we fix potential infeasibility by decreasing promotions. But unlike the
previous claim, we operate in a top-down fashion since the perturbation is expected to end at the lt ∈ Lt
we choose.
Wet consider level k + 1 to determine . Since ∆rt−1 , ∆rt−1, ∆rt , and (in particular,
∆z ∆zt
k,k+1 k+1 k k+ k+1,k+2
2
t
∆z kb+1,kb+2 = 0) are all known as we reach node (k + 1,t), we can solve


 ∆zt + ∆rt−1 = k+1,k+2 + ∆rk+

t
 k,k+1 t k+ 1
∆z 1
∆rt ,
/
 1 ≤ k+ qk+1,k+2
k+ t
 ∆r
2
∆
t t−1
k,k+1 ≤ ∆rk
z
and choose the largest possible ∆ztk,k+1 , which yields (11a). ∆r can then be determined via flow conservation,

i.e. (11b) and (11c).

Claim A.7. Assuming a perturbed solution is feasible for periods 1, . . . , t − 1, if there exists an lt ∈ Lt with
min{min jt, min l¯t−1 + 1} ≤ lt ≤ max{max jt, max l¯t−1 − 1} and xt > 0 for k with min{min jt, min l¯t−1 +
k,k+1

1} ≤ k ≤ max{max jt, max l¯t−1 − 1}, then a solution also feasible for period t can be obtained by
sequen- tially applying equations (10) to levels min{min jt, min l¯t−1 + 1} ≤ k ≤ lt − 1, equations
(11) to levels lt + 1 ≤ k ≤ max{max jt, max l¯t−1 − 1}, and finally letting

=
∆zlt t,lt +1
(12a)
min{∆r t−1
+ —t ql −1,l , +t /ql +1,l +2 − ∆rt−1 , ∆rt−1}
∆zt ∆r ∆zt ∆r
lt lt −1,lt lt −1 t lt +1,lt lt +2 t lt +1 lt
+2 t
t

∆rt = ∆zt + ∆rt−1 − ∆zt (12b)

lt lt −1,lt lt lt ,lt +1

60
 ∆rt
=t + ∆rt−1 − ∆zt . (12c)
∆z
lt +1 lt ,lt lt lt +1,lt +2
+1 +1

Proof. This is a hybrid of the previous two claims. By a similar analysis the perturbation should only

61
cover levels from min{min jt, min l¯t−1 + 1} to max{max jt, max l¯t−1 − 1}. We can treat any such
lt as a breakpoint above which Claim A.6 applies and below which Claim A.5 applies. It only remains to
perturb
level lt itself. To guarantee feasibility we can choose the minimum of the values provided by Claims A.5
and A.6 to determine
∆zt lt ,lt
, and and can then be calculated by flow conservation. This yields
+1 ∆rt ∆rt
lt lt
(12). +1

Claim A.8. Assuming a perturbed solution is feasible for periods 1, . . . , t − 1, if there are zero
promotions between successive levels lt, . . . , jt or jt, . . . , lt, then a solution also feasible for period t
can be obtained by applying (10) to each succession lt, . . . , jt or (11) to each succession jt, . . . , lt, as
long as the jt values partition Jt.

Proof. By Claim A.2 there is no l¯t−1 and so we only consider the impact of full promotions at jt. The
applicability of the claims follows immediately from the fact that each succession here is an instance
of Claim A.5 or A.6.

We need to ensure, though, that each jt is included in exactly one succession. Starting from max jt, if
there exists some lt with lt ≥ max jt and xtk,k+1 > 0 for k with max jt ≤ k ≤ lt, there may be a lower jt that
satisfies this condition as well; hence we can decrease promotions from the lowest such jt to any such lt,
and use the formulas from Claim A.5 to determine ∆zt k,k+1 and ∆rkt . Otherwise, by Claims A.3 and A.4 there
must exist some lt with lt < max jt and k,k+1
xt > 0 for k with lt ≤ k ≤ max jt, and so we can use the formulas
from Claim A.6 to determine
and until reaching the highest such lt. After either case is done,
∆zt k,k+1
∆rt
k+
we can move downwards to the next jt that 1has not been visited, and apply the same argument again. This
process goes on until reaching min jt.

We now state our perturbation procedure as Procedure 2. Note that if in some period we find xt k+1,k+2 <
¯
k+ for each perturbed level k (including when Jt = L t −1 = ∅), the procedure can end in this period,
pst−1
1
and from then on the perturbed solution will remain the same as the initial solution; otherwise, the
procedure will iterate forward infinitely but converge to a new feasible solution. Claim A.9 justifies the
lower cost of the final perturbed solution.

62
Claim A.9. Procedure 2 modifies the given solution by adding to it a series of negative cost cycles or
infinite paths in the time-space network.

63
Proof. Except for the possible end period, all the perturbations are initiated by decreasing the z values. Pick
any node (k,t) with negative ∆r flows in the perturbation network. Each time we conduct an operation as in
Claims A.5 to A.8, by flow conservation ∆rt is passed to either ∆rt−1 or , resulting in a left arc and a
∆zt
k k k−1,k

downward arc, respectively. As this propagates, two cases may occur.

Case 1: We reach some lt (t ≤ t) where the flow turns right, then follows a right-down-right pattern,
and finally turns upwards at the perturbation’s end period. This constitutes a counterclockwise
cycle as described in Claim A.1.

Case 2: The flow may continue shifting in a right-down-right pattern perpetually, which constitutes a
coun- terclockwise infinite path as described in Claim A.1.

Procedure 2 Perturbation when n ≥ 3 and pk = pk+1, ∀ k = 1, . . . , n − 1

1: ∆r1 = −∆r1 = ∆z1 = −ε, J1 = L1 = {m}, L¯ 0 = ∅, t = 2, ep = 0
m+1 m m,m+1
2: while ep = 0 do
3: if tk+1,k+2 < psk+
t−1
for every perturbed level k then
x4: ep = 1 1
5: else
6: Run Procedure 1 to identify Jt, Lt, and L¯ t −1
7: if Claim A.5 is applicable then
8: Perturb according to Claim A.5
9: else if Claim A.6 is applicable then
10: Perturb according to Claim A.6
11: else if Claim A.7 is applicable
then 12: Perturb according to
Claim A.7 13: else
14: Perturb according to Claim A.8
15: end if
16: Update ∆x, ∆s, x, s
17: t=t+1
18: end if
19: end while
20: for perturbed levels k in increasing order do
21: ∆zt = ∆rt−1 + ∆zt
k,k+1 k k−1,k
22: Update ∆x, ∆s, x, s
23: end for

64
A.4 A technical note on the perturbation amount ε

We have constructed perturbation operations that are feasible for small enough ε. To obtain a valid perturbed
solution, however, we need to guarantee that ε > 0. Because the perturbation may range over infinitely many
periods, it could be that the required ε eventually converges to zero. We next argue why this is not the
case. If the perturbation ends in some period, it essentially works in finite dimensions and thus ε > 0.
On the other hand, if the procedure iterates infinitely, ε depends on the x and s values. In particular, the values
of the training variables, promotion variables, and slack between consecutive levels matter since we are
decreasing them. To eliminate the possibility of ε converging to zero, it suffices to bound those values
from below

wherever they are perturbed.

When n = 2, the proof of (4b) reduces the promotion in period 1 and increases it in later periods.
The increments depend on the reduction in period 1 and so depend on how much we can decrease there,
which clearly is positive. In the proof of (4a), ε depends on st+1 − dt+1 if xt+1 = p1st , and p1st − xt+1 if
st+1 = dt+1.
1 1,2 1 1 1,2 1
dIn the former case, s −
t+1
p2st +p1st ≥
(p +
d
pmin
)d − ≥p d , which is bounded away
= −
1 t+1
d 2
t+ min t t+ min t
1
qmax
1 1
q1,2

from zero since dt ≥ d1 > 0. In the latter case, the perturbation ends in period t + 1 so the
perturbation is finite.
Now consider when n ≥ 3. Recall that in all the cases considered for perturbation, we perturb
levels between jt and lt, or l¯t−1 and lt, or both. For any level k with lt ≤ k ≤ jt, the proof of Claim A.4
actually
provides a lower bound independent of t, i.e. xt+1 pqk+1, jt k+
k,k+11
≥ pq1, jt d1. Furthermore, if we redefine
>
t+1
s t t′

g = min{i : xi,i+1 ≤ C, jt + 1 ≤ i ≤ n}, where C > 0 can be any constant less than mint′=1,...,t {psi } for each
i (such as C = pq1,nd1), the bound still holds, and hence the x variables at levels k with jt ≤ k ≤ lt are
also bounded below by a constant independent of t.

The same trick can be applied to any level k between lt and l¯t−1, i.e. the lower bounds on the x
variables in Claim A.2 can be strengthened from zero to the same constant C above. Finally, Claim A.3 is
still correct
max
t+1
for g,g+1 ≤ C as long as C < pqn−1d1. This enables us to use a g that still satisfies the properties in both
xclaims as the level starting from which a search of lt is conducted in Procedure 1.

65
 Since the proofs of the claims only utilize the linear relationships between s and x, we can obtain similar
bounds for the slack between consecutive s variables. All the lower bounds depend only on d1 and
fixed parameters like n, p and q. It follows that the ε in the infinite case is indeed positive.

66
B Proof of Theorem 3.3

We will construct a one-period lookahead policy based on Lemmas 3.1 and 3.2, and then demonstrate that
the resulting solution is unique.

B.1 A one-period lookahead policy

The notation we use is summarized below:

t
i,i+1
•B : upper promotion bound for level i in period t.

• I: a list of levels where promotion bounds would be violated if hiring were not allowed.

Claim B.1. Consider a subsystem consisting of levels from j to k + 1, 0 ≤ j ≤ k ≤ n − 1 in period t. Assume

st−1, i = j, . . . , k + 1 and st are known. If p j+1st−1 < q j, j+1st and pi+1st−1 ≤ qi,i+1 pist−1, i = j + 1, . . . , k,
i j j+1 j i+1 i

then the unique solution to the following equations provides a solution that satisfies promotion bounds at
levels in I:

p j+1st−1 + xt —t +t =t q j, j+1(st − ) (13a)

x x x
j+1 j, j+1 j+1, j+ j j, j+1
j+2 1

pi+1st−1 + xt —t +t =t qi,i+1(pist−1 + — xt + xt),

x x x
i+1 i,i+1 i+1,i+2 i+ i i−1,i i,i+1
1
(13b)
i
i = j + 1, . . . , k − 1

pk+1st−1 + xt +t =t qk,k+1(pkst−1 + — xt + xt ), (13c)

x x
k+1 k,k+1 k+ k k−1,k k,k+1 k
1

where xt = 0 if i ∈/ I and xt = if i ∈ I.
Bt
i i,i+1 i,i+1

Proof. Clearly any solution to the above linear system satisfies promotion bounds at any i ∈ I. Nonnegativity
of the x variables is guaranteed by the deficiency of staff at level j and the tight ratio relationships
(with respect to retention from period t − 1) at levels j, . . . , k. For a specific solution x, a policy can
be obtained

67
by letting


st − xt , i= j
 i i,i+1

st i = pist−1 + xti−1,i — xi,i+1

t
+ xit, i = j + 1, . . . , k . (14)
i


i + x i−1,i+ x i,
pist−1 t t
i=k+1

Let the coefficient matrix be such that column l (l = 1, . . . , k − j +1) records the coefficients ofj+l−1,
xt j+l
if j + l − 1 ∈/ I
otherwise. The elements are
and xt j+l

−q j+l, j+l+1, m=l+1


aml 1 if j + l − 1 ∈ I and 1 + q j+l−1, j+l otherwise,m = l
= 0 if j + l − 1 ∈ I and − 1 otherwise, m=l − .
 1

0, otherwise

Define Dm (m = 1, . . . , k − j + 1) as the determinant of the submatrix composed of the first m

rows and the first m columns. For m ≥ 3, we have the recursion


Dm−1, j+ m− 1∈ I
D m=  .
(1 + qm+ j−1,m+ j )Dm−1 − qm+ j−1,m+ j Dm−2, otherwise

By induction we know Dk− j+1 ≥ Dk− j ≥ . . . ≥ D1 > 0, and hence the solution is unique.

We now construct a feasible one-period lookahead policy by solving subproblems composed of levels
1, . . . , k + 1 sequentially until k = n − 1. During each loop, we first check if the resulting solution is
feasible without promotion at level k, if yes then we are done. Otherwise, we try to get a solution which
uses only promotions, i.e. solve (13) with I = ∅. If this happens to be feasible, then we update the st
and xt values and exit; otherwise we calculate a feasible solution by allowing hiring, i.e. solve (13) with I
/= ∅. For each k, we keep iterating these steps for subproblems composed of levels j, . . . , k + 1 so that
we can stop at the highest j and the lower levels are not affected. When determining promotion and
hiring, we force the ratio constraints to be tight so that we use the smallest possible xt. In other words,

68
we promote and hire only if

69
necessary. A formal statement is described in Procedure 3.

Procedure 3 A one-period lookahead policy

1:t st = dt if i = 0 and pist−1 if i ≥ 1, =t 0 if t = 1 and ∞ if t ≥ 2, = pist−1, i = 1, . . . , n − 1
B B
i i 0,1 i,i+1 i
2: k=0
3: while k ≤ n − 1 do
4: j=k
5:t while j ≥ 0 and < q j, j+1st do
s
6: I= j+1 j

∅
7: Bt =
t
j, —
t
j, j+1
j+1
j, B x
j+1
8: Solve (13)
t
9: I = {i : i,i+1 > Bi,i+1 , j ≤ i ≤ k}
xt
10: if I /= ∅ then
11: while I is changed do
12: Solve (13)
13: Update I
14: end while
15: end if
16: Update s with (14)
17: j=j−1
18: end while
19: k=k+1
20: end while

We end this section with two comments. First, the procedure is applicable to both n = 2 and n ≥ 3.
Second, once i enters I at some iteration, it will be there forever: The first time i enters I, xt i,i+1 must be
decreased (from infeasibility) to full promotion and so sit must be increased in the next iteration (which
is true since the only possibility for st not to be increased is then to decrease xt or xt , but this would
i i i−1,i

induce infeasibility between sit and lower levels). To further satisfy the ratio relationships at levels i to k,
t
sl+1 (i ≤ l ≤ k) cannot be decreased either, which in turn forces the promotions at these levels to be full if
they were. This implies the procedure terminates.

B.2 Optimality of the one-period lookahead policy

Claim B.2. Recursively applying Procedure 3 yields the unique solution that satisfies Lemmas 3.1 and 3.2.

Proof. Clearly the solution satisfies the lemmas. Suppose there are multiple feasible solutions for which the

70

k,k+1 k
lemmas hold. We compare an arbitrary one of them, say (u, y), with (s, x) obtained from Procedure 3.

By
Lemma 3.1 = pkut−1. Start from the earliest period, say t, where there is a difference
> 0 only if
yt k+
1 yt

71
 ′ ′
between xt and yt. Pick the lowest different level, say i. We have st = ut , ∀ k ≤ i − 1, and st = ut , ∀ t′ ≤ t − 1.
k k
We first note that yt < xt or yt < xt cannot be true; otherwise we should be able to obtain a smaller
i,i+1 i,i+1 i i

x as Procedure 3 finishes since a feasible solution must satisfy the ratio and bound constraints at every level.
t

We next show that yt > xt or yt >

xt cannot be true, either. Since (u, y) is feasible, by (1e)
i,i+1 i,i+1 i i
ytt > xt or yt > xt implies ut > for some k ≥ i, so there is over promotion/hiring and the lemmas
s
i,i+1 i,i+1 i i k k
must be violated somewhere in (u, y).

C Extended model used in the computational examples

We modified (1) for our computational examples as follows. Let:

• T ≥ 1: Length of planning horizon.

• L ≥ 1: Length of the training program.

• a: Age of the student or worker, al ≤ a ≤ au.

• p0,i ∈ (0, 1): Per-period rate of continuing education for students in school year i = 1, . . . , L
− 1, or per-period rate of graduating and going to the workforce for students in school year i =
L.

• pk,a ∈ (0, 1): Per-period retention rate of workers of age a = al, . . . , au that stay in the system at level
k = 0, . . . , n from one period to the next.

• mk,a: The age distribution of students (k = 0) or workers (k = 1, . . . , n) of age a = al, . . . , au.

t
• s0,i,a : Students of age a = al, . . . , au in school year i = 1, . . . , L at end of period t = 1, . . . , T .

t
• sk,a : Workers of age a = al, . . . , au in level k = 1, . . . , n at end of period t = 1, . . . , T .

Our modified problem has the following formulation.

T n
min C s x γt−1 c xt n−1 n
h
au
st
ct
x
(, )= ∑ k k+ ∑ k,k+1 k,k+1
+ ∑ ∑
k k,a
∑ k= k=1 k=1 a=a
t=1 0 l
au
t
s.t. ∑ 1,a ≥ dt , ∀ t = 1, . . . , T
a=al

72
 au au
t
∑ k+1,a — qk,k+1 ∑ kst ≥ 0, ∀ k = 1, . . . , n − 1, ∀ t = 1, . . . , T
a=al
,a
a=al
t
s0,1,a — m0,ax = 0, ∀ a = al, . . . , au, ∀ t = 1, . . . , T
t

0
t
s0,i,a — p0,i−1s t−1
0,i−1,a−1 = 0, ∀ i = 2, . . . , L ∀ a = al + 1, . . . , au − 1, ∀ t = 1, . . . , T
t
s — p0,i−1(st−1
0,i,au t−1
+ 0,i−1,au−1 ) = 0, ∀ i = 2, . . . , L, ∀ t = 1, . . . , T
0,i−1,au s
st − p1,a−1st−1
— m1,axt − p0,Lst−1 + t = 0,
m2,ax
1,a 1,a−1 1 0,L,a−1 1,2

∀ a = al + 1, . . . , au − 1, ∀ t = 1, . . . , T
t
s − p1,a st−1 − p1,a −1st−1
— m1,a xt − p0,L(st−1 + st−1 )
1,au u u 1,au−1 u 0,L,au 0,L,au−1
1,au 1

+ m2,axt
1,2
= 0, ∀ t = 1, . . . , T

pk,a−1st−1
— t
k,a + — k,k+1
+ mk,axt = 0, ∀ k = 2 . . . , n − 1,
k,a−1
s mk,axt mk+1,ax t
k−1,k k
∀ a = al + 1, . . . , au − 1, ∀ t = 1, . . . , T
pk,a −1st−1
+ pk,a st−1 − st +t mk,a —t mk+1,a + mk,a xt = 0,
x x
u k,au−1 u k,au k,au u k−1,k u k,k+1 u k

∀ k = 2 . . . , n − 1, ∀ t = 1, . . . , T
pn,a−1st−1
+ t +t xt ) − = 0, ∀ a = al + 1, . . . , au − 1,
mn,a(x s
n,a−1 n−1,n n n,a

∀ t = 1, . . . , T
pn,a −1st−1
+t pn,a st−1 + mn,a +t xt ) − = 0, ∀ t = 1, . . . , T
(x s
u n,au−1 u n,au u n−1,n n n,au

pk,a−1st−1 — k,k+1 ≥ 0, ∀ k = 1, . . . , n − 1,
k,a−1
t
mk+1,ax

∀ a = al + 1, . . . , au, ∀ t = 1, . . . , T

s0,i,a = 0, ∀ i = 2, . . . , L ∀ a = al, . . . , al + i − 2, ∀ t = 1, . . . , T
t

stk,a = 0, ∀ k = 2, . . . , n − 1, ∀ a = al, . . . , al + k + L − 2, ∀ t = 1, . . . , T

xt, st ≥ 0, ∀ t = 1, . . . , T
 50

View publication stats