Professional Documents
Culture Documents
By
JIANG ZHANG
Department of Operations
August 2004
CASE WESTERN RESERVE UNIVERSITY
JIANG ZHANG
______________________________________________________
Matthew J. Sobel
(signed)_______________________________________________
(chair of the committee)
Lisa M. Maillart
________________________________________________
Yunzeng Wang
________________________________________________
Peter H. Ritchken
________________________________________________
________________________________________________
________________________________________________
06/16/2004
(date) _______________________
*We also certify that written approval has been obtained for any
proprietary material contained therein.
I grant to Case Western Reserve University the right to
use this work, irrespective of any copyright, for the
University’s own purposes without cost to the University
or to its students, agents and employees. I further agree
that the University may reproduce and provide single
copies of the work, in any format other than in or from
microforms, to the public for the cost of reproduction.
JIANG ZHANG
To my mother: Huishu Jiang
List of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ix
Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . x
Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xii
1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
v
2 Fill Rate of General Review Supply Systems . . . . . . . . . . . . . . . . . . . 25
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
in Inventory Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
3.5 Interchangeability . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
vi
3.6 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
3.8 Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
vii
List of Tables
3.6 S-optimal Base-Stock Levels and Fill Rates at which they are F-optimal 88
3.7 F-optimal Base-Stock Levels and Unit Stockout Costs at which they are
S-optimal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
viii
List of Figures
2.2 The Fill Rate Integral for a system with Normal Demand . . . . 38
negative Density . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
3.4 Locus of {(b, f)} with the Same Optimal Base-Stock Level: Non-
negative Density . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
ix
Acknowledgements
studies with patience. I would also like to thank him for his incredible effort and
sors Lisa Maillart, Peter Ritchken, and Yunzeng Wang for their generous insight,
comments, and support on this work. In addition, my thanks are also owed to Pro-
Western Reserve University, for their generous financial support. Special thanks
to department’s staff, Elaine Iannicelli, Sue Rischar, and Emily Anderson for their
I had a fabulous time at Case which would not have been possible without the
company of friends like Junze Lin, Zhiqiang Sun, Yuanjie He, Huichen Chiang,
x
Wei Wei, Xiang Fang, Qiaohai Hu, Will Millhiser, Ant Printezis, Halim Hans, and
Kang-hua Li who have always helped and cheered me up in every possible way.
Finally, I would like to thank my family for their unconditional love, support
and encouragement. My special thanks are for my wife Yan who is always there
xi
Three Essays on Inventory Management
Abstract
by
JIANG ZHANG
This dissertation consists of three essays that are related to inventory man-
agement.
The first essay models a single-product equity-owned firm which orders prod-
ucts from an outside supplier, borrows short-term capital for solvency, and issues
dividends to its shareholders while facing financial risks and demand uncertainty.
The firm maximizes the expected present value of the time stream of dividends. If
there is a setup cost in this model, we show that an (s, S) replenishment policy is
optimal by jointly optimizing the firm’s operational and financial decisions. The
this essay studies the same model with a smoothing cost (instead of a setup cost)
and shows that the optimal policy has the same form as the traditional smoothing
xii
cost model. Although operational decisions and financial decisions interact with
each other in these models, the optimal inventory policies have standard forms.
The second essay obtains fill rate formulas for general review inventory models
exact fill rate formulas for single-stage model with a general demand distribution.
A simple fill rate expression is derived for the model with normally distributed
each stage and then provide exact fill rate formulas for two-stage and three-stage
models.
There are parallel streams of literature which analyze identical models except
that one stream has backorder costs and the other has fill rate constraints. The
third essay clarifies redundancy in the two streams of dynamic inventory models
with linear purchase costs. We show that optimal policies for either kind of
model can be inferred from the other. That is, inventory fill rate constraints
xiii
Chapter 1
Financial Criterion
1.1 Introduction
Nearly all the literature on optimal inventory management uses criteria of cost
demands. If inventory decisions do not affect the revenue stream, these two crite-
ria result in the same optimal replenishment policy. Most of this literature treats
perhaps due to the perception that inventory managers in a large firm cannot in-
fluence the firm’s financial policy and that financial officers are usually detached
1
2
from the inventory decisions. This separate consideration of financial and opera-
tional decisions simplifies management and has its foundation in corporate finance.
The pathbreaking papers, Modigliani and Miller (1958) and Modigliani and Miller
(1963) (hereafter referred to as M-M), show that the firm’s capital structure and
its financial decisions should be independent of the firm’s investment and opera-
However, when market imperfections such as taxes and transaction fees are in-
troduced, the results characterized from these separate treatments may no longer
hold. “Treating real and financial decisions of the firm as independent may not
resources by different firms. For example, many small and medium-sized firms
are cash constrained, and their operational decisions are heavily dependent on
appealing and insightful results, there remains the question of whether joint opti-
mization of both the operational and financial decisions of a firm will generate new
insights regarding firm behavior and perhaps overturn or modify existing results.
The M-M theorem not only allows separation of operations and finance, but
also establishes that the firm’s optimal decisions are the same no matter if it
3
optimizes the value of the firm, dividends, or retained earnings. However, when
valid. The literature on agency theory finds that corporate managers, the agents
reduce the resources under managers’ control, thereby reducing managers’ control,
and making it more likely they will incur the monitoring of the capital markets
which occurs when firm must obtain new capital”(Jensen 1986, p. 1). So managers
have to seek external funds to finance their projects. Since the external funds are
usually unavailable for certain firms or available only at high prices, it somewhat
reduces the profitability (or increases the operating costs) of the projects and may
affect the performance evaluation of the managers. This conflict may discourage
managers from disgorging the cash to shareholders and cause organizational inef-
ficiencies.
among a firm’s decisions. Most of this literature focuses on the effects of mar-
ket imperfections on financial structures and decisions. Miller and Rock (1985)
outside investors and they show that there exists an equilibrium investment pol-
icy which leads to lower levels of investment than the optimum achievable under
4
full information. Myers (1974) shows how investment decisions, i.e., acceptance
or rejection of projects, affect the optimal financial structure of a firm and why
shows that the cost of capital of a competitive firm facing stochastic demand is
affected by the level of production. Hite (1977) examines the impact of leverage
on the optimal stock of capital by the firm and its capital-to-labor ratio. Datan
and Ravid (1958) analyze the interaction between the optimal level investment
and debt financing. In their model, a firm faces an uncertain price and has to
decide on its optimal level investment and debt simultaneously. They show that
a negative relationship exists between investment and debt. Dammon and Senbet
(1988) extend capital structure model in DeAngelo and Masulis (1980) and study
the effect of corporate and personal taxes on the firm’s optimal investment and
financing decisions under uncertainty. However, this literature does not address
and uses a dividend criterion. The primary purpose of this paper is to study how
gle product equity-owned retail firm which periodically reviews its inventory and
retained earnings. Every period the firm faces random demand and replenishes
5
ods, the firm issues a dividend to its shareholders. The firm seeks to maximize the
expected present value of the time stream of dividends (also called shareholders’
wealth).
the literature (cf. Hojgaard and Taksar 2000, Milne and Robertson 1996, Moyer,
McGuigan, and Kretlow 1990, Sethi 1996, and Taksar and Zhou 1998). This goal
states that management should seek to maximize the present value of the expected
future returns to the shareholders of the firm. These returns can take the form of
If the objective of the firm is to minimize the expected present value of or-
dering, holding, and shortage costs and the capital market is perfect, the optimal
replenishment policies have been characterized for a broad range of conditions. See
Porteus (1990), Graves et al. (1993), and Zipkin (2000) for details and references
to the literature.
decisions. Li, Shubik, and Sobel (2003) examine the relationship between de-
inventory models. They show that there are myopic optimal base-stock policies
associated with production decisions and dividend decisions. The present paper
proceeds directly from Li, Shubik, and Sobel (2003) and augments their model
6
with a setup cost and smoothing costs. Buzacott and Zhang (1998) look at the
interface of finance and production for small firms with limited borrowing. They
to optimize inventory and borrowing decisions, and they assume that the demand
for the product is known. Archibald, Thomas, Betts, and Johnston (2002) as-
sert that start-up firms are more concerned with the probability of survival than
with profitability. They present a sequential decision model of a firm which faces
Buzacott and Zhang (2003) incorporate financial capacity into production de-
period newsvendor model. They model the available cash as a function of assets
and liabilities that will be updated according to the dynamics of the production
activities. They analyze a leader-follower game between the bank and the retailer,
and illustrate the importance of jointly considering production and financial deci-
sions. Babich and Sobel (2002) consider capacity expansion and financial decisions
to maximize the expected present value of a firm’s IPO. They treat the IPO event
The rest of the paper is organized as follows. Section 1.2 formulates the finan-
cial inventory model and §1.3 analyzes the corresponding dynamic program. The
and extended to the infinite horizon optimal policy is discussed in §1.5. Section
1.6 studies the financial inventory model with smoothing costs and characterizes
We consider an equity-owned retail firm that sells a single product to meet uncer-
tain periodic demand and orders the product from a supplier with an ample supply.
The firm can make short-term loans, if necessary, to obtain working capital. As
discussed in the introduction, every R periods, dividends are issued to the share-
holders and the objective of the firm is to maximize the expected present value
occurs in each period. The firm observes the level of retained earnings, wn , and
the current physical inventory level, xn . A default penalty (or bankruptcy) p(wn )
assume that p(·) is convex nonincreasing on . Then the firm chooses the level
of its short-term loan, bn , and the order quantity, zn . The restriction R = 1 sim-
generality.
At the beginning of each period, the firm also decides on the amount of divi-
vn < 0 then it is a capital subscription. Also at the beginning of the period, the
paid, and the ordering decision is implemented at a cost of Kδ(zn ) + czn , where K
is an ordering setup cost, δ(zn) = 1 if zn > 0, and δ(zn ) = 0 otherwise. Then de-
mand Dn in period n is realized, and sales revenue net of inventory cost, denoted
where yn , r, h and π denote available goods in period n after delivery, unit sale
price, holding cost and shortage penalty cost, respectively. However, we only use
convexity of g(·, d) for each d ≥ 0. Finally, the loan principal bn is repaid. We as-
For convenience but without loss of generality, we assume that the order lead-
period n is
yn = xn + z n (1.1)
That is, In is working capital after the dividend is issued, loan interest and replen-
ishment costs are paid, and before the loan is made and revenue and inventory
costs are realized. The total working capital available at the beginning of period n
is bn + wn and the residual cash left in the firm before sales is bn + In . We assume
that there is an interest rate γ associated with In , that is, if working capital is
positive, the firm will gain γIn if In ≥ 0 or pay a penalty −γIn if In < 0.
xn+1 = xn + zn − Dn
The first equation balances the flow of physical goods and the second equation
balances the cash flow. Using (1.1) and (1.2), the balance equations become
xn+1 = yn − Dn (1.3)
bn ≥ 0 and zn ≥ 0 (1.5)
Given xn and wn , from (1.1) and (1.2) the decision variables in period n can be
Let B denote the present value of the time stream of dividends and let β
∞
B= β n−1vn (1.7)
n=1
Remark Discount factor β can be regarded as the risk neutral discount rate for
the shareholders and need not to be a constant every period. Letting β(xn , wn )
namely,
Hn = (x1 , w1, b1 , I1, y1, D1 , · · · , xn−1 , wn−1 , bn−1 , In−1 , yn−1 , Dn−1 , xn , wn )
Let Ωn be the set of all possible Hn sequences. A policy is a nonanticipative rule for
choosing y1, I1 , b1, y2, I2, b2 , . . . . That is, a policy is a rule that, for each n, chooses
for each η ∈ Ωn , for all n = 1, 2, . . .. Since a policy specifies the three decisions
each period (the amount of supply level, yn , the amount of internally generated
working capital, In , and the short-term loan, bn ), the firm can easily determine
the order size and the amount of dividends to issue to shareholders by using (1.1)
11
i.e., one that maximizes the expected value of (1.7) subject to (1.3)-(1.6).
This section gives the dynamic programming equations which correspond to the
the decision space from three to two. From (1.2) and (1.3), vn = wn − p(wn ) −
In − cyn + cxn − ρ(bn ) − Kδ(yn − xn ) and xn = yn−1 − Dn−1 (n > 1), and using
∞
B = cx1 − β ncDn
n=1
∞
+ β n−1 wn − p(wn ) − In − (1 − β)cyn − ρ(bn ) − Kδ(yn − xn )
n=1
∞
∞
= cx1 + w1 − p(w1 ) − β n cDn − β n−1Kδ(yn − xn )
n=1 n=1
∞
n−1
+ β −(1 − β)cyn − In + β (1 + γ)In + g(yn , Dn )
n=1
−βp (1 + γ)In + g(yn , Dn ) − ρ(bn )
L(b, I, y) = −(1 − β)(I + cy) + βγI + βE g(y, D) − p (1 + γ)I + g(y, D) − ρ(b)
(1.8)
12
∞
∞
E(B) = cx1 +w1−p(w1 )−E β n cDn +E β n−1 L(bn , In , yn )−Kδ(yn −xn )
n=1 n=1
(1.9)
As in Veinott and Wagner (1965) and in many other references since then, we
and penalties). Because the first four terms in (1.9) do not depend on the decision
variables, a policy maximizes (1.9) if and only if it maximizes the last term of (1.9).
∞
sup EH1 β n−1 [L(bn , In, yn ) − Kδ(yn − xn )]
n=1
where the supremum is over the set of all policies, H1 = (x1, w1 ) is the initial
state, and the constraints in (1.10) follow from (1.1), (1.2), (1.5), and (1.6).
N
sup EH1 β n−1 [L(bn , In , yn ) − Kδ(yn − xn )]
n=1
and n = 1, 2, . . .,
This dynamic program has one state variable, x, and three decision variables (b,
I, y), but the original problem has two state variables, x and w. This reduction
occurs because equation (1.4) allows embedding state variable w into the decision
Let bn (x), In (x) and yn (x) be optimal values of b, I and y in (1.12). The
following proposition states that borrowing should not exceed the amount needed
to cover current expenses. That is, bn (x) = [−In(x)]+ . The proof is similar to
ϕn (x) = sup −(1 − β)(I + cy) + βγI + βE g(y, D) − p[(1 + γ)I + g(y, D)]
b,I,y
−ρ(b) + βE[ϕn−1 (y − D)] − Kδ(y − x) : y ≥ x, I + b ≥ 0, b ≥ 0
ϕn (x) = sup − c(1 − β)y + βE g(y, D) + βE ϕn−1 (y − D) − Kδ(y − x)
y
+ sup −(1 − β)I + βγI − βE p((1 + γ)I + g(y, D))
I
+
−ρ[(−I) ] : I ∈ : y ≥ x .
The next section uses Proposition 1.3.1, (1.12), and (1.13) to analyze the
dynamic problem and establish conditions that guarantee the optimality of (s, S)
policies.
cies
This section establishes conditions under which the optimal ordering policy turns
out to be an (s, S)-type policy for the dividend criterion inventory model. An (s,
S) policy brings the level of inventory after ordering up to S if the initial inventory
level x is below s (where s ≤ S), and orders nothing otherwise. For a finite horizon
dynamic inventory problem in which the ordering cost is linear plus a fixed setup
cost and the other one-period costs are convex, Scarf (1959) and Zabel (1962)
show that the optimal ordering policy is (sn , Sn ). Iglehart (1963) shows that
the limiting (s, S) policy characterizes the optimal policy for the infinite horizon
problem. Scarf’s proof uses the important concept of K-convexity. In this paper,
we use K-concavity.
15
x ∈ , λ ≥ 0, and θ > 0,
λ
f(x + λ) − f(x) ≤ K + [f(x) − f(x − θ)] (1.14)
θ
tions.
0, α2 > 0);
Proof f(·) is K-concave if and only if −f(·) is K − convex. So (a) through (d)
For dynamic program (1.12) and (1.13), we use the following results to show
Let µ and Q(·) be the mean and distribution function of D. Let G(y) =
E[g(y, D)].
Lemma 1.4.2 (a) p[(1 + γ)I + g(y, d)] is convex with respect to (I, y) ∈ 2 (for
each d ≥ 0);
Proof (a) Since g(·, d) is a convex on for each d ≥ 0, (1 + γ)I + g(y, d) is jointly
convex in (I, d). The conclusion in (a) follows from monotonicity and convexity
of p(·). p[(1 + γ)I + g(y, d)] is convex on I ∈ , so is E{p[(1 + γ)I + g(y, D)]}.
Note that the hypothesis uses concavity of L(·, ·, ·) and ensures the existence of
maxima of L(b, I, ·). The proof of this theorem is not a paraphrasing of Scarf
(1959). Indeed, it exploits Proposition 1.3.1 to reduce the decision space from 3
each n.
Tn (y) = sup{Jn (b, I, y) : b ≥ 0, b+I ≥ 0}; Proposition 1.3.1 asserts that b = (−I)+
is optimal. So
(Proposition B-4) shows that ζ(·) is concave on by proving that supI∈ {L(0, I, y)−
ρ[(−I)+]} is concave in y.
The proof that an (s, S) policy is optimal for period n follows the next lemma
Lemma 1.4.4 Suppose T (·) is K-concave, attains its global maximum at S, and
the global maximum of Tn (·) is attained, say at Sn . Also, let sn be the smallest
optimal.
The analysis in §1.3 replaces the infinite planning horizon in (1.10) with a finite
planning horizon (n < ∞) in (1.11). This section shows that the earlier conclu-
sions regarding the qualitative properties of an optimal policy remain valid for
(1.10). We draw on Iglehart (1963) and Heyman and Sobel (1984, sections 8-5,
8-6), and proves that (a) there are upper and lower bounds for the sequences
{sn } and {Sn } of the finite horizon optimal policy, (b) the value function of the
finite horizon dynamic program converges as n → ∞, (c) the limit value function
nite horizon optimal policy converges to a policy that is optimal in the functional
equation, and (e) the limit policy inherits the qualitative properties of the finite
optimal policies.
Assuming that sup(I,b)∈2 {L(b, I, y)} ≥ 0 leads to ζ(·) > 0 on . Then the
Then −Tn(·) and −ϕ(·) are K-convex, and −ζ(·) is convex. The dynamic program
in (1.19) and (1.18) preserves all the properties (convex one-step cost function and
(Iglehart 1963) hold in our infinite horizon model. Therefore, an (s, S) policy is
optimal in (1.10).
In this section, we replace the setup cost in previous sections with smoothing
costs and provide another test of the following conjecture. For a broad class of
cost structures in inventory models, replacing traditional criteria with the dividend
criterion does not change the form of the optimal replenishment policy. Smoothing
Let e+ be the unit cost of increasing replenishment and e− be the unit cost of
∞
∞
B = cx1 − β n cDn + β n−1[wn − p(wn ) − In − (1 − β)cyn
n=1 n=1
e+ − e−
e+ · (zn − zn−1 )+ + e− · (zn−1 − zn )+ = e|zn − zn−1 | + (zn − zn−1 )
2
∞
B = β n−1 −(1 − β)(In + cyn ) + βγIn + βg(yn, Dn )
n=1
e+ − e−
− βp (1 + γ)In + g(yn , Dn ) − ρ(bn ) − e|zn − zn−1 | − (1 − β)2yn
2
∞
+ cx1 + w1 − p(w1 ) − β ncDn
n=1
e+ − e− ∞
+ (1 − β)x1 − β n (1 − β)Dn − z0
2 n=1
Let
M(b, I, y) = −(1 − β)(I + cy) + βγI + βE g(y, D) − p[(1 + γ)I + g(y, D)]
e+ − e−
−ρ(b) − (1 − β)2y
2
Therefore,
∞
E(B) = β n−1 E M(bn , In , yn ) − e|zn − zn−1 |
n=1
∞
+E cx1 + w1 − p(w1 ) − β n cDn (1.21)
n=1
e+ − e− N
+ (1 − β)x1 − β n (1 − β)Dn − z0
2 n=1
21
Since the last two rows of (1.21) depend only on the distribution of demand and
lem and the former converges to the latter. Henceforth, we optimize the following
objective:
N
sup EH1 { β n−1 [M(bn , In , yn ) − e|zn − zn−1 |]}
n=1
φn (x, z) = sup −e|y − x − z| + sup{Jn (b, I, y) : I + b ≥ 0, b ≥ 0}
y≥x
22
= sup −e|y − x − z| + hn (y) (1.25)
y≥x
The assumption that M(·, ·, ·) is a concave function on its domain 3 leads to the
concavity of hn (·) on y.
and b + I ≥ 0, then for each n and x, there are numbers un (x) and Un (x) with
and hn (·) are concave functions on their respective domains because M(·, ·, ·) is
their boundaries), let hn (·) denote the left-hand derivative of hn (·). Concavity
and a modification of Sobel (1969) leads to the structural result in Theorem 1.6.2.
Moreover,
His method is based on the convexity of the value functions, and can be directly
applied in our model. So the infinite horizon counterparts of (1.22), (1.23), (1.24),
and (1.26) are valid and the following stationary policy is optimal:
⎧
⎪
⎪ u(x) if x + z < u(x)
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪ if u(x) ≤ x + z < U(x)
⎨x + z
y= (1.27)
⎪
⎪
⎪
⎪ U(x) if x ≤ U(x) ≤ x + z
⎪
⎪
⎪
⎪
⎪
⎪
⎩
x if U(x) < x
where u(x) = sup{y : h (y) ≤ e}, U(x) = sup{y : h (y) ≤ −e}, and h(y) is the
limit of hn (y) as n → ∞.
In this paper, we consider periodic review inventory systems where the objective
consider joint financial and replenishment decisions and examine models with a
setup cost and with smoothing costs by embedding the state variable for working
24
analysis and allows the use of existing stochastic optimization techniques to obtain
qualitative results. In the setup cost case, the proof is not straightforward. In
both cases, the same form of replenishment policy is optimal as when the criterion
is cost minimization.
R is a positive integer, would lead to similar results. The only change would be to
to the model, Proposition 1.3.1 and Corollary 1.6.1 remain valid. So Theorems
1.4.3 and 1.6.2 remain essentially unchanged with only minor changes in their
proofs.
Chapter 2
Supply Systems
2.1 Introduction
All inventory systems face a difficult tradeoff between inventory costs and cus-
tomer service. The fill rate, the long-run average fraction of demand which is sat-
isfied immediately from on-hand inventory, is perhaps the most important measure
There is a literature on formulas for the fill rate under different inventory re-
plenishment policies. Most of it concerns the fill rate in a single-stage system with
25
26
for the item fill rate in an uncapacitated single-stage system with normally dis-
tributed demand, develop a new approximation, and evaluate the accuracy via
Sobel (2004), Glasserman and Tayur (1994), and Glasserman and Liu (1997)
consider the fill rate of capacitated periodic review multi-stage supply systems
in which each stage reviews its inventory periodically, and there is a constant
correction terms, for fill rate and optimal base-stock levels of multi-stage systems,
whereas the fill rate formulas in Sobel (2004) are exact and the bounds are valid
without asymptotics.
Research on fill rate of periodic review inventory systems usually assumes that
the system reviews its inventory every period. In practice, although customer
demand may arise every period, a firm may not review its inventory and make
who is supplied by a wholesaler who ships products to the retailer by truck once
a week. Although the retailer might prefer to replenish her inventory daily, she
should reorder goods only shortly before the truck leaves. One may argue that in
this situation, if we define the unit period as one week, the system becomes the
usual periodic review system. Indeed, this is a widely held perception. However,
our results show that the fill rate computed via a rescaled periodic review system
27
inventory policy subject to a lower bound on the fill rate induced by the policy.
Many authors consider optimization problems with service level constraints and
Tijms and Groenevelt (1984) consider both periodic review and continuous re-
view (s, S) inventory systems and present a practical approximation for the reorder
point s subject to a fill rate constraint and find that the normal approximation
gives good results for required service levels when the coefficient of variation of
the demand during lead time and review periods does not exceed 0.5.
Silver (1970), Yano (1985), and Platt, Robinson, and Freund (1997) propose
ordering. The system faces a discrete compound Poisson demand process in which
optimal policy consists of a mixed multistage echelon stock (R, nQ) policy with
models with setup costs, focus on (s, S) policies where the order quantities are
s such that the required service level is achieved. Schneider and Ringuest consider
28
Boyaci and Gallego (2001) and Shang and Song (2003) study a periodic re-
view service-constrained serial inventory system where the leadtime demand for
the end product is Poisson distributed. Their service measure, the limiting prob-
ability of having positive on-hand inventory at the last stage, differs from the fill
rate. Boyaci and Gallego focus on base-stock policies, develop heuristic solutions,
and discuss the relationship between stockout cost and service-constrained models.
Shang and Song study the same model, and develop closed-form heuristics to ap-
paper concerns exactly optimal policies for periodic review single-stage models
formance to various parameters. So fill rate equations are used to analyze the
sensitivity of inventory levels to alternative fill rate goals. In this sense, the role
of fill rate targets is similar to that of stockout costs, but practitioners seem to
prefer fill rate targets. Van Houtum and Zijm (2000) discuss the possible relations
between backorder cost and several types of service contraints. In particular, they
establish the one to one correspondence between backorder cost and modified fill
rate (one minus the ratio of the average backlog at the end of a period and the
mean demand per period) constraint. Chapter 3 of this dissertation show that fill
models and establish monotone mappings between the set of optimal polices with
backorder costs and the set of optimal policies with fill rate constraints.
The rest of this paper is organized as follows. Section 2.2 introduces the general
and §2.4. Section 2.3 provides fill rate formulas for general demand distribution.
Specific fill rate formulas are developed for Gamma and Normal distributions
system and has the fill rate formulas for general review two-stage systems. Section
2.6 has a formula for the fill rate of a three-stage system. We conclude the paper
in §2.7.
The following model describes a periodic review N-stage serial system which is
displayed in Figure 2.1. Materials, parts or products can be ordered from any
stage and are then shipped to the next downstream stage. The inventory level
for additional items, if any. An order for material placed at the beginning of a
review period t with destination stage n arrives at that stage at the beginning of
period t+Ln (Ln and Rn are positive integers.), if sufficient materials are available
at stage n + 1. The outside supplier preceding stage N has ample supplies and
can deliver any order that is placed by stage N. Customer demand for the end
30
At the beginning of period t, let xnt denote the number of items that are in
at stage 1 minus the number of items backlogged, if any, at the beginning of period
t. That is, x1t is the on-hand physical inventory if x1t ≥ 0, and −x1t is the amount
of backordered demand if x1t < 0. Let zN t be the number of items purchased from
an outside supplier in period t, and for n < N, let znt be the number of items
and finite mean µ. To avoid trivialities, it is assumed that G(0) < 1. Let G(k) (·)
k
denote the k-fold convolution of G(·), i.e., the distribution function of j=1 Dj ,
stages;
Let ynt be the inventory at stage n in period t after deliveries of previously ordered
Notice that znt = 0 if period t is not a review period for stage n. For expository
convenience, let period t with t|R = 0 (where a|b denotes a modulo b when a and
stage 1 with order size zt units which will be delivered at period t + L1. If t|R
= 0,
(y1t)+ . Because excess demand (if any) is backlogged, the inventory dynamic are
as follows:
The fill rate, β, is the long run average fraction of demand that can be satisfied
where (u)+ denote max{u, 0}. The expectation and limit exist in (2.31) for the
which products are ordered from an outside supplier every fixed R periods and
are available to satisfy demand in period t + L (where both R and L are positive
policies have been proved to be optimal for periodic review single-stage system
under general conditions and are very easy to implement in practice (cf. Zipkin
L+[(t−L)|R]
y1t = τ − Dt−k , for (t − L)|R = 0, 1, · · · , R − 1 (2.32)
k=1
In Sobel (2004), there is an inconsistency between the fill rate definition and
the proof of Theorem 1 . He uses y1t in the fill rate definition [(3) on page 43], but
uses x1t to derive the fill rate formula in the proof of Theorem 1 (line 9, page 44).
However, because of the chronology differences between his model and the present
33
model, substituting x1t in the proof by y1t of our model results in the same fill
rate formulas. Formulas (4) and (5) in Sobel (2004) remain valid here if R = 1.
The following theorem gives an exact fill rate formula for a general review
Theorem 2.3.1
τ
1
β= [G(L)(b) − G(L+R) (b)]db (2.33)
Rµ 0
Proof From (2.31), the fill rate is the long run average fraction of demand that
T T
+
β = lim E min{(y1t) , Dt }/ Dt
T →∞
t=1 t=1
T T
= lim E [ min{(y1t)+ , Dt }/T ]/[ Dt /T ]
T →∞
t=1 t=1
1 T
= lim E min{(y1t)+ , Dt } /T
µ T →∞ t=1
1 T
= lim E min{(y1t )+ , Dt } /(T /R)
Rµ T →∞ t=1
R−1
1
= lim E min{(y1t)+ , Dt } /(T /R)
Rµ T →∞ i=0 t∈{t:(t−L)|R=i}
1
R−1
= lim E min{(y1t)+ , Dt }/(T /R)
Rµ T →∞ i=0 t∈{t:(t−L)|R=i}
+
Let Hj = t∈{t:(t−L)|R=j} min{(y1t ) , Dt }, j = 0, 1, · · · , R − 1. Using (2.32), for all
L+j
lim E[Hj /(T /R)] = E min (τ − Dk )+ , DL+j+1
T →∞
k=1
L+j
= E DL+j+1 − [DL+j+1 − (τ − Dk )+ ]+
k=1
34
L+j
= µ − E [DL+j+1 − (τ − Dk )+ ]+
k=1
L+j
1 R−1
1 R−1
β(τ ) = Hj = µ − E [DL+j+1 − (τ − Dk )+ ]+
Rµ j=0 Rµ j=0 k=1
L+j
1 R−1
= 1− E [DL+j+1 − (τ − Dk )+ ]+
Rµ j=0 k=1
L+j
1 R−1
+
K(τ ) = E [DL+j+1 − (τ − Dk )+ ]
R j=0 k=1
∞ τ
1 R−1
= (a + b − τ )dG(L+j) (b)dG(a)
R j=0 0 τ −a
∞ ∞
+ a dG(L+j) (b)dG(a)
0 τ
∞ τ
1 R−1
= µ 1 − G(L+j) (τ ) + (a + b − τ )dG(L+j) (b)dG(a)
R j=0 0 τ −a
∞ τ
1 R−1
K (τ ) = − dG(L+j) (b)dG(a)
R j=0 0 τ −a
R−1 τ ∞
1
= − dG(a)dG(L+j) (b)
R j=0 0 τ −b
R−1
τ
1 (L+j)
= G(τ − b) − 1 dG (b)
R j=0 0
1
R−1
= G(L+j+1) (τ ) − G(L+j) (τ )
R j=0
1 (L+R)
= G (τ ) − G(L) (τ )
R
τ
= 1 − [K(0) + K (a)da]/µ
0
τ
1
= [G(L) (b) − G(L+R) (b)]db.
Rµ 0
Theorem 2.3.1 characterizes the dependence of the system fill rate on the base-
stock level (τ ), demand distribution (G), review period (R), and leadtime (L).
Systems
This subsection specializes (2.33) for the gamma distribution and the normal
distribution.
Let Γ(j, λ) denote the sum of j independent, identically distributed random vari-
ables, each one exponential with parameter λ, i.e., Γ(j, λ) is a gamma random
λe−λa (λa)γ−1
Γ(γ)
γ−1
G(a) = P (D ≤ a) = 1 − e−λa (λa)j /j!
j=0
36
Consequently, G(L) and G(L+R) are Γ(Lγ, λ) and Γ[(L + R)γ, λ], respectively. So,
Lγ−1
G(L) (a) = 1 − e−λa (λa)j /j!
j=0
(L+R)γ−1
G(L+R) (a) = 1 − e−λa (λa)j /j!
j=0
τ (L+R)γ−1
Lγ−1
1
β = e−λa(λa)j /j! − e−λa(λa)j /j! da
Rµ 0 j=0 j=0
τ (L+R)γ−1
λ
= e−λa (λa)j /j! da
Rγ 0 j=Lγ
(L+R)γ−1 τ
1
= λe−λa (λa)j /j! da
Rγ j=Lγ 0
Therefore,
(L+R)γ
1
β= P {Γ(j, λ) ≤ τ } (2.34)
Rγ j=Lγ+1
Many researchers investigate the fill rate of an inventory system with normally
distributed demands. In order to compare our results with others, we analyze the
general review inventory systems when demand (D) is normally distributed with
Let Φ(·) and φ(·) denote the distribution and density function, respectively,
of a standard normal random variable (with mean 0 and variance 1), and let
√
b(a, j) = (a − jµ)/(σ j). The normality and independence of demand imply
37
that sums of demands are normally distributed; that is, G(S) (a) = Φ[b(a, S)] for
1 τ
β = Φ[b(a, L)] − Φ[b(a, L + R)] da
Rµ 0
√ b(τ,L) √ b(τ,L+R)
1
= σ L Φ(x)dx − σ L + R Φ(x)dx (2.35)
Rµ b(0,L) b(0,L+R)
The evaluation of the integral in (2.35) exploits the following equation (Hadley
∞
[1 − Φ(x)]dx = φ(t) + tΦ(t) − t
t
s
Φ(x)dx = φ(s) − φ(t) + sΦ(s) − tΦ(t). (2.36)
t
Using (2.36) in (2.35) yields the following equation for the fill rate that uses only
the standard normal density and tabulated standard normal distribution function:
1 √
β = σ L φ[b(τ, L)] − φ[b(0, L)] + b(τ, L)Φ[b(τ, L)] − b(0, L)Φ[b(0, L)]
Rµ
√
− σ L + R φ[b(τ, L + R)] − φ[b(0, L + R)]
+ b(τ, L + R)Φ[b(τ, L + R)] − b(0, L + R)Φ[b(0, L + R)]
√
1 √
= σ L φ[b(τ, L)] − Φ[b(0, L)] − σ L + R φ[b(τ, L + R)] − φ[b(0, L + R)]
Rµ
+ (τ − Lµ) Φ[b(τ, L)] − Φ[b(τ, L + R)]
+ R Φ[b(τ, L + R)] + L Φ[b(0, L)] − (L + R)Φ[b(0, L + R)] . (2.37)
b
xφ(x)dx = φ(a) − φ(b). (2.38)
a
38
Figure 2.2: The Fill Rate Integral for a system with Normal Demand
τ b(a,L) b(a,L+R)
1
β= φ(x)dx − φ(x)dx da
Rµ 0 −∞ −∞
whose integrand in the a − x plane covers the area in the northeast and southeast
b(a, L), and x = b(a, L + R). As shown in Figure 2.2 , this area is the union of
formula (2.38), and integrating A1, A2, and A3 individually leads to the same fill
tribution
Although (2.37) can be calculated easily, the following simple approximation yields
valuable insights:
√
1
βA = − σ L + Rφ[b(τ, L + R)]
Rµ
+ (τ − Lµ) 1 − Φ[b(τ, L + R)] + RΦ[b(τ, L + R)] (2.39)
Approximation (2.39) is derived from (2.37) by observing from Figure 2.2 that
φ[b(0, L)] 0, φ[b(τ, L)] 0, φ[b(0, L+R)] 0, Φ[b(τ, L)] 1, Φ[b(0, L+R)] 0,
and Φ[b(0, L)] 0. Our numerical results show that the fill rate approximation
(2.39) is very accurate when τ ≥ (L + R)µ. The numerical comparisons are shown
in Table 2.1.
Table 2.1 use the same normal demand distribution as Table 1 in Sobel (2004),
compute the exact value and approximation of the fill rate in a general review
single-stage system with µ = 2000 and L + R = 5. The last columns of the tables
report the absolute error (100%) of fill rate and its approximation.
• The fill rate increases as the variance (σ 2) goes down or base-stock level (τ )
goes up.
(σ/µ) decreases. There are three cases (with high variance, long leadtime) in
our example that cause large errors, but the errors diminish when base-stock
level increases.
with linear inventory holding costs at all stages and linear backorder costs at stage
one (Clark and Scarf 1960, Federgruen and Zipkin 1984). It is quite natural to use
• echelon inventory level of a stage is the inventory on hand at this stage plus
sbnt = the (beginning) echelon inventory level at stage n before any order is
received
41
abnt = the (beginning) echelon inventory position at stage n before any order
is placed
The evolution of the system and dynamics of these echelon variables can be
specified as:
Ln
abnt = sbnt + zn,t−k ant = abnt + znt snt = sbn,t + zn,t−Ln (2.40)
k=1
and
L
n −1
formulation in echelon variables because xnt = ynt − znt, ynt = snt − an−1,t (let
Ln −1
s0t = a0t = 0), and znt = ant − k=1 zn,t−k − snt . Because s1t = y1t, (2.31) can
be written
T
t=1 min{(s1t)+ , Dt }
β = lim E T (2.42)
T →∞
t=1 Dt
for n < N, ⎧
b
⎪
⎨ ant if τn ≤ abnt
ant = (2.44)
⎪
⎩
min{τn , sn+1,t } otherwise,
and ant = abnt if period t is not a review period for all n ≤ N. The order quantity
The subsequent sections discuss the fill rate of general review two-stage and
This subsection considers a two-stage general review system in which both stages
have the same review intervals (R1 = R2 = R) and order leadtimes are L1 = L
review period for stage two. Both stages have the same length of review intervals.
system; in essence, the only difference is that the order quantity in stage one at
characterize the relationship between the echelon inventory level at stage one and
base-stock levels τ1 and τ2 for stage one and 2, respectively. It is without loss
of generality to assume that the initial echelon inventory positions are no higher
43
than the echelon base-stock levels (cf. Sobel 2004, page 47):
At the beginning of each period t, abnt , sbnt for n = 1, 2, z2,t−1, and z1,t−k for
k = 1, · · · , L are known. In fact, there are at most Ln /R ( a/b is the minimum
integer that is greater than or equal to a/b ) outstanding orders that in transit to
Suppose t is a review period at stage two and R > 1. Since R > 1, stage
one does not place an order in period t. Then a2t = τ2 , s2t = sb2t, a1t = ab1t, and
s1t = sb1t + zt−L. When demand Dt occurs at stage one, period t + 1 is not a review
period for stage two, so ab2,t+1 = sb2,t+1 = τ2 −Dt = s2,t+1 (order has been received).
Period t + 1 is a review period for stage one, ab1,t+1 = ab1,t − Dt , sb1,t+1 = s1t − Dt ,
and because of ordering, a1,t+1 = min{τ1 , s2,t+1}. Order quantity will be delivered
L period later, until then the changes of echelon inventory level (s1t) depend only
The following result is consistent with comments in several papers [e.g. Cheng
Lemma 2.5.1
L
s1t = a1,t−L − Dt−k
k=1
Proof
L−1
s1t = a1t − zt−k
k=0
L−1
= a1t − (a1,t−k − ab1,t−k−1 )
k=0
44
L−1
= a1t − a1,t−k − (a1,t−k−1 − Dt−k−1 )
k=0
L
= a1,t−L − Dt−k
k=1
Lemma 2.5.1 states that the echelon inventory level at stage one in period t is
periods’ demands.
As a consequence [similar to the derivation of (2.32)], s1t for any period t can
L+(t−L)|R
s1t = τ1 + min{0, ∆ − Dt−[L+(t−L)|R]−1 } − Dt−k (2.45)
k=1
its inventory one period ahead of stage one. Lemma 2.5.1 assures
L
s1L = a1,L − Dt−k
k=1
L
= min{τ1 , τ2 − DL−1 } − Dt−k
k=1
L+0
= τ1 + min{0, ∆ − DL−0−1 } − Dt−k
k=1
L
s1,t+1 = a1,t+1−L − Dt+1−k
k=1
L
= min{τ1 , τ2 − Dt+1−L−1 } − Dt+1−L−k
k=1
L+(t+1−R)|R
= τ1 + min{0, ∆ − Dt+1−[L+(t+1−L)|R]−1 } − Dt+1−k
k=1
If (t + 1 − L)|R
= 0, then zt+1−L = 0, (t + 1 − L)|R = (t − L)|R + 1, so
s1,t+1 = s1,t − Dt
L+(t−L)|R
= τ1 + min{0, ∆ − Dt−[L+(t−L)|R]−1 } − Dt−k − Dt
k=1
L+(t+1−R)|R−1
= τ1 + min{0, ∆ − Dt−[L+(t+1−L)|R−1]−1 } − Dt−k − Dt
k=1
L+(t+1−R)|R
= τ1 + min{0, ∆ − Dt+1−[L+(t+1−L)|R]−1 } − Dt+1−k
k=1
Lemma 2.5.2 leads to a relatively simple fill rate formula. Notice that expres-
Theorem 2.5.3
1 τ1
β = G(∆) G(L)(a) − G(L+R) (a)
Rµ 0
∞
+ G(L) (a + ∆ − c) − G(L+R) (a + ∆ − c) dG(c) da (2.46)
∆
Proof From (2.31), the fill rate is the long run average fraction of demand that
T T
+
β = lim E min{(s1t) , Dt }/ Dt
T →∞
t=1 t=1
46
T T
= lim E [ min{(s1t )+ , Dt }/T ]/[ Dt /T ]
T →∞
t=1 t=1
1 T
= lim E min{(s1t )+ , Dt } /T
µ T →∞ t=1
1 T
= lim E min{(s1t)+ , Dt } /(T /R)
Rµ T →∞ t=1
R−1
1
= lim E min{(s1t)+ , Dt } /(T /R)
Rµ T →∞ i=0 t∈{t:(t−L)|R=i}
1
R−1
= lim E min{(s1t)+ , Dt }/(T /R)
Rµ T →∞ i=0 t∈{t:(t−L)|R=i}
+
Let Hj = t∈{t:(t−L)|R=j} min{(s1t ) , Dt }, j = 0, 1, · · · , R− 1. Using (2.45) and
1 R−1
β(τ1, ∆) = β(Hj )
Rµ j=0
L+j+1 +
1 R−1
= 1 − E DL+j+2 − [τ1 − (D1 − ∆)+ − Dk ]+
Rµ j=0 k=2
L+j+1 +
1 R−1
K(τ1 , ∆) = E DL+j+2 − [τ1 − (D1 − ∆)+ − Dk ]+
R j=0 k=2
47
τ1 +∆−c ∞
1 R−1
∞
=
R j=0 ∆ 0 τ1 +∆−b−c
Notice that the sum of the second and third lines in (2.47) is zero and the sum of
1 ∞ τ1 +∆−c
R−1
= G(L+j+2) (τ1 + ∆ − b − c) − 1 dG(L+j) (b)dG(c)
R j=0 ∆ 0
∆ τ1
+ G(L+j+2) (τ1 − b) − 1 dG(L+j) (b)dG(c)
0
0
1 R−1
∞
= G(L+j+1) (τ1 + ∆ − c) − G(L+j) (τ1 + ∆ − c) dG(c)
R j=0 ∆
48
∆
(L+j+1) (L+j)
+ G (τ1) − G (τ1 ) dG(c)
0
1 ∞
R−1
= G(L+j+1) (τ1 + ∆ − c) − G(L+j) (τ1 + ∆ − c) dG(c)
R j=0 ∆
(L+j+1) (L+j)
+G (τ1 ) − G (τ1 ) G(∆)
1 ∞
= G(L+R) (τ1 + ∆ − c) − G(L) (τ1 + ∆ − c) dG(c)
R ∆
(L+R) (L)
+G (τ1) − G (τ1 ) G(∆) (2.48)
Therefore,
time
The formula for a two-stage general review system with a single period leadtime
in stage two can be easily extended to a two-stage general review system in which
the review intervals are R1 = R2 = R and order leadtimes are L1 and L2 . We use
stage two. Both stages have the same length of review intervals.
Lemma 2.5.1 still holds for this general two-stage system. We give the following
L2 L1 +(t−L1 )|R
s1t = τ1 + min 0, ∆ − Dt−[L ]−k − Dt−k (2.49)
1 +(t−L1 )|R
k=1 k=1
The proof of Lemma 2.5.4 is similar to that of Lemma 2.5.2 except that the order
placed by stage two has L2 delay before it can be used to satisfy stage one’s order.
L2
Replacing D1 and G(c) by k=1 Dk and G(L2 ) (c), respectively, in the proof of
Theorem 2.5.5
τ1
1
β = G(L2 )(∆) G(L1 ) (a) − G(L1 +R) (a)
Rµ 0
∞
+ G(L1 )(a + ∆ − c) − G(L1 +R) (a + ∆ − c) dG(L2 ) (c) da (2.50)
∆
In what follows, we present an alterative formula for (2.50) using the single
stage fill rate formulas. Let β1(τ, L) be the fill rate of a single stage system
incomplete convolutions (cf. van Houtum et al. 1996). Let G(k) be the distribution
k
functions of j=1 . Let δ > 0 and define
⎧
(k)
⎪
⎨ G (x + δ) if x ≥ 0
(k)
Gδ (x) =
⎪
⎩
0 if x < 0,
50
[m,n]
The incomplete convolution, denoted Gδ , is
δ
[m,n]
Gδ (x) = G(m) (δ + x − u)dG(n) (u).
0
[m,n]
note that G0 (x) = G(m+n) (x)
Corollary 2.5.6
1
β = G(L2 ) (∆)β1(τ1 , L1 ) + β1(τ1 + ∆, L1 + L2 ) − β1 (∆, L1 + L2 )
Rµ
τ1
[L1 ,L2 ] [L +R,L2 ]
− G∆ (a) − G∆ 1 (a) da (2.51)
0
τ1 ∞
1
β = G(L1 ) (a + ∆ − c) − G(L1 +R) (a + ∆ − c) dG(L2 ) (c)da
Rµ 0 ∆
τ1
(L2 ) (L1 ) (L1 +R)
+G (∆) G (a) − G (a) da
0
τ1 a+∆
1
= G(L1 )(a + ∆ − c) − G(L1 +R) (a + ∆ − c) dG(L2 )(c)da
Rµ 0 ∆
+G(L2 ) (∆)β1(τ1 , L1 )
1 τ1 a+∆ (L1 )
= G (a + ∆ − c) − G(L1 +R) (a + ∆ − c) dG(L2 )(c)da
Rµ 0 0
τ1 ∆
− G(L1 ) (a + ∆ − c) − G(L1 +R) (a + ∆ − c) dG(L2 ) (c)da
0 0
(L2 )
+G (∆)β1(τ1 , L1 )
τ1
1
= G(L1 +L2 ) (a + ∆) − G(L1 +L2 +R) (a + ∆) da
Rµ 0
τ1
[L ,L ] [L +R,L2 ] (L2 )
− [G∆ 1 2 (a) − G∆ 1 (a) da + G (∆)β1(τ1 , L1 )
0
τ +∆
1 1
= G(L1 +L2 ) (u) − G(L1 +L2 +R) (u) du
Rµ ∆
τ1
[L ,L ] [L +R,L2 ] (L2 )
− [G∆ 1 2 (a) − G∆ 1 (a) da + G (∆)β1(τ1 , L1 )
0
τ1 +∆
1
= G(L1 +L2 ) (u) − G(L1 +L2 +R) (u) du
Rµ 0
51
∆
− G(L1 +L2 ) (u) − G(L1 +L2 +R) (u) du
0
τ1
[L ,L2 ] [L +R,L2 ]
− [G∆ 1 (a) − G∆ 1 (a) da + G(L2 )(∆)β1(τ1 , L1 )
0
1
= G(L2 ) (∆)β1(τ1 , L1 ) + β1(τ1 + ∆, L1 + L2 ) − β1(∆, L1 + L2 )
Rµ
τ1
[L1 ,L2 ] [L +R,L2 ]
− G∆ (a) − G∆ 1 (a) da (2.52)
0
It is clear from (2.50) that the fill rate of a two-stage general review system depends
• The fill rate increases as variance goes down, and as echelon base-stock level
• The fill rate decreases as leadtime at stage one and stage two increase.
• The fill rate increases as echelon base-stock level at stage two increases.
This section considers a three-stage general review system in which the review
interval is the same at each stage (R1 = R2 = R3 = R) and order leadtimes are
52
period at stage three. We assume that an echelon base-stock policy with echelon
L3
s1t = τ1 + min 0, ∆1 + min 0, ∆2 − Dt−[L ]−L2 −k
1 +(t−L1 )|R
k=1
L2
L1 +(t−L1 )|R
− Dt−[L ]−k − Dt−k (2.53)
1 +(t−L1 )|R
k=1 k=1
Proof sketch: Suppose t is a review period at stage one. Then the echelon
position is
L2
a1t = min{τ1 , s2t } = min{τ1 , a2,t−L2 − Dt−k }
k=1
L2
= min τ1 , min{τ2 , s3,t−L2 } − Dt−k
k=1
L3 L2
= min τ1 , min{τ2 , a3,t−L2 −L3 − Dt−L2 −k } − Dt−k
k=1 k=1
L3
L2
= min τ1 , min{τ2 , τ3 − Dt−L2 −k } − Dt−k
k=1 k=1
L3
L2
= min τ1 , τ2 + min{0, ∆2 − Dt−L2 −k } − Dt−k
k=1 k=1
L3
L2
= τ1 + min 0, ∆1 + min{0, ∆2 − Dt−L2 −k } − Dt−k .
k=1 k=1
Theorem 2.6.2
τ1
∞ ∞
1
β = G(L1 )(a + ∆1 + ∆2 − c − e)
Rµ 0 ∆2 ∆1 +∆2 −e
−G(L1 +R) (a + ∆1 + ∆2 − c − e) dG(L2 ) (c)dG(L3 )(e)
∞
(L1 ) (L1 +R)
+G (a) − G (a) G(L2 ) (∆1 + ∆2 − e)dG(L3 ) (e)
∆2
∞
(L3 )
+G (∆2) G(L1 )(a + ∆1 − c) − G(L1 +R) (a + ∆1 − c) dG(L2 ) (c)
∆1
(L1 ) (L1 +R) (L2 ) (L3 )
+G (a) − G (a) G (∆1)G (∆2 ) da. (2.54)
T T
β = lim E min{(s1t)+ , Dt }/ Dt
T →∞
t=1
t=1
1
R−1
= lim E min{(s1t)+ , Dt }/(T /R)
Rµ T →∞ i=0 t∈{t:(t−L)|R=i}
+
Let Hj = t∈{t:(t−L)|R=j} min{(s1t ) , Dt }, j = 0, 1, · · · , R − 1, use (2.53), and
notice that the demand random variables are independent. Let D(j) have the
j
distribution of k=1 Dk . Then limT →∞ E[Hj /(T /R)] is equal to
(L3 ) (L2 ) (L1 +j) +
E min τ1 + min 0, ∆1 + min {0, ∆2 − D }−D −D ,D
Let β(τ1 , ∆1, ∆2) make explicit the dependence of β on τ1 , ∆1, and ∆2. Then
1 R−1
β(τ1, ∆1, ∆2) = β(Hj )
Rµ j=0
54
K(τ1 , ∆1 , ∆2)
1 R−1
+ + +
(L3 ) + (L2 ) (L1 +j)
= E D − τ1 − [D − ∆2 ] + D − ∆1 −D
R j=0
1 R−1
∞ ∞ τ1 +∆1 +∆2 −e−c ∞
=
R j=0 ∆2 ∆1 +∆2 −e 0 τ1 +∆1 +∆2 −e−c−b
(a + b + c + e − τ1 − ∆1 − ∆2 )dG(a)dG(L+j)(b)dG(L2) (c)dG(L3)(e)
∞ ∞ ∞ ∞
= adG(a)dG(L1+j) (b)dG(L2) (c)dG(L3)(e)
∆∞2 ∆∆11+∆2 −e τ1 +∆1 +∆2 −e−c 0
+∆2 −e τ1 ∞
+ (a + b − τ1 )dG(a)dG(L1+j) (b)dG(L2) (c)dG(L3) (e)
∆∞
2
0 ∆1 +∆2 −e 0 τ1 −b
∞ ∞
+ adG(a)dG(L1+j) (b)dG(L2)(c)dG(L3) (e)
∆∆2 2 0 ∞ τ1 +∆1τ−c
1
0∞
+
0 ∆1 0 τ1 +∆1 −c−b
−G(L1 ) (τ1 + ∆1 + ∆2 − c − e) dG(L2 ) (c)dG(L3 ) (e)
∞
+[G(L1 +R) (τ1 ) − G(L1 ) (τ1 )] G(L2 ) (∆1 + ∆2 − e)dG(L3 ) (e)
∆2
∞
+G(L3 ) (∆2) G(L1 +R) (τ1 + ∆1 − c) − G(L1 ) (τ1 + ∆1 − c) dG(L2 ) (c)
∆1
G(L1 +R) (τ1 ) − G(L1 ) (τ1 ) G(L2 ) (∆1)G(L3 ) (∆2 ) (2.55)
[m,n,q] [m,n,q]
An alternate to (2.54) uses (2.33) and (2.50). Let Gδ1 ,δ1+δ2 and Gδ1 ,δ2 be the
Corollary 2.6.3
1
β = G(L3 )(∆2 )β2(τ1 , ∆1, L1 , L2 )
Rµ
[L ,L3
+β1(τ1 , L1 ) G(L2 +L3 ) (∆1 + ∆2) − G∆22 (∆1)
The proof of Corollary 2.6.3 is similar to that of Corollary 2.5.6 and is omitted.
56
2.7 Conclusion
This paper develops formulas for the fill rates of single-stage and multi-stage
supply systems that use base-stock-level policies and have general review intervals.
We provide fill rate formulas for a single-stage general review system and general
rate expression uses only the standard normal distribution function and density
function. For the general review multi-stage systems, we first discuss how each
stage reviews its inventory and provide a general approach to compute the system
fill rate.
57
σ σ/µ τ R L β βA %Error
200 0.1 11000 1 4 0.999014 0.999014 0
200 0.1 11000 2 3 0.999507 0.999507 0
200 0.1 11000 3 2 0.999671 0.999671 0
200 0.1 11000 4 1 0.999754 0.999754 0
1000 0.5 11000 1 4 0.789395 0.760095 2.93
1000 0.5 11000 2 3 0.880264 0.880047 0.0217
1000 0.5 11000 3 2 0.919869 0.920032 0.0163
1000 0.5 11000 4 1 0.938963 0.940024 0.1061
2000 1 11000 1 4 0.590938 0.335729 25.5209
2000 1 11000 2 3 0.686992 0.667864 1.9128
2000 1 11000 3 2 0.766131 0.778576 1.2445
2000 1 11000 4 1 0.815568 0.833932 1.8364
200 0.1 12000 1 4 1 1 0
200 0.1 12000 2 3 1 1 0
200 0.1 12000 3 2 1 1 0
200 0.1 12000 4 1 1 1 0
1000 0.5 12000 1 4 0.895048 0.886563 0.8485
1000 0.5 12000 2 3 0.943282 0.943282 0
1000 0.5 12000 3 2 0.962025 0.962188 0.0163
1000 0.5 12000 4 1 0.97058 0.971641 0.1061
2000 1 12000 1 4 0.679695 0.520189 15.9506
2000 1 12000 2 3 0.765023 0.760095 0.4928
2000 1 12000 3 2 0.826923 0.840063 1.314
2000 1 12000 4 1 0.861683 0.880047 1.8364
200 0.1 13000 1 4 1 1 0
200 0.1 13000 2 3 1 1 0
200 0.1 13000 3 2 1 1 0
200 0.1 13000 4 1 1 1 0
1000 0.5 13000 1 4 0.95544 0.953442 0.1998
1000 0.5 13000 2 3 0.976695 0.976721 0.0026
1000 0.5 13000 3 2 0.984318 0.984481 0.0163
1000 0.5 13000 4 1 0.987299 0.98836 0.1061
2000 1 13000 1 4 0.758474 0.664425 9.4049
2000 1 13000 2 3 0.829454 0.832213 0.2759
2000 1 13000 3 2 0.874769 0.888142 1.3373
2000 1 13000 4 1 0.897742 0.916106 1.8364
Table 2.1: Fill Rate and its Approximation for Normal Demand
58
σ σ/µ τ1 ∆1 R L1 L2 β σ σ/µ τ1 ∆1 R L1 L2 β
2 0.2 50 0 1 4 1 0.1747 2 0.2 80 0 1 4 1 1
2 0.2 50 0 1 4 2 0.0037 2 0.2 80 0 1 4 2 0.9940
2 0.2 50 0 2 3 1 0.4985 2 0.2 80 0 2 3 1 1
2 0.2 50 0 2 3 2 0.0892 2 0.2 80 0 2 3 2 0.9970
2 0.2 50 0 3 2 1 0.6654 2 0.2 80 0 3 2 1 1
2 0.2 50 0 3 2 2 0.3336 2 0.2 80 0 3 2 2 0.9980
2 0.2 50 0 4 1 1 0.7491 2 0.2 80 0 4 1 1 1
2 0.2 50 0 4 1 2 0.5 2 0.2 80 0 4 1 2 0.9985
5 0.5 50 0 1 4 1 0.3030 5 0.5 80 0 1 4 1 0.9749
5 0.5 50 0 1 4 2 0.1052 5 0.5 80 0 1 4 2 0.8546
5 0.5 50 0 2 3 1 0.4702 5 0.5 80 0 2 3 1 0.9868
5 0.5 50 0 2 3 2 0.2041 5 0.5 80 0 2 3 2 0.9147
5 0.5 50 0 3 2 1 0.6200 5 0.5 80 0 3 2 1 0.9912
5 0.5 50 0 3 2 2 0.3485 5 0.5 80 0 3 2 2 0.9427
5 0.5 50 0 4 1 1 0.7141 5 0.5 80 0 4 1 1 0.9933
5 0.5 50 0 4 1 2 0.4932 5 0.5 80 0 4 1 2 0.9571
5 0.5 50 20 1 4 1 0.6357 5 0.5 80 20 1 4 1 0.9987
5 0.5 50 20 1 4 2 0.5395 5 0.5 80 20 1 4 2 0.9944
5 0.5 50 20 2 3 1 0.7773 5 0.5 80 20 2 3 1 0.9993
5 0.5 50 20 2 3 2 0.6948 5 0.5 80 20 2 3 2 0.9971
5 0.5 50 20 3 2 1 0.8503 5 0.5 80 20 3 2 1 0.9994
5 0.5 50 20 3 2 2 0.7906 5 0.5 80 20 3 2 2 0.9979
5 0.5 50 20 4 1 1 0.8868 5 0.5 80 20 4 1 1 0.9986
5 0.5 50 20 4 1 2 0.8419 5 0.5 80 20 4 1 2 0.9975
Table 2.2: Fill Rate of Two-stage Systems for Normal Demand (a)
59
σ σ/µ τ1 ∆1 R L1 L2 β σ σ/µ τ1 ∆1 R L1 L2 β
5 0.5 50 30 1 4 1 0.6373 5 0.5 50 50 1 4 1 0.6373
5 0.5 50 30 1 4 2 0.6284 5 0.5 50 50 1 4 2 0.6373
5 0.5 50 30 2 3 1 0.7785 5 0.5 50 50 2 3 1 0.7785
5 0.5 50 30 2 3 2 0.7714 5 0.5 50 50 2 3 2 0.7785
5 0.5 50 30 3 2 1 0.8512 5 0.5 50 50 3 2 1 0.8512
5 0.5 50 30 3 2 2 0.8462 5 0.5 50 50 3 2 2 0.8512
5 0.5 50 30 4 1 1 0.8874 5 0.5 50 50 4 1 1 0.8874
5 0.5 50 30 4 1 2 0.8837 5 0.5 50 50 4 1 2 0.8874
5 0.5 80 30 1 4 1 0.9987 5 0.5 80 50 1 4 1 0.9987
5 0.5 80 30 1 4 2 0.9985 5 0.5 80 50 1 4 2 0.9987
5 0.5 80 30 2 3 1 0.9993 5 0.5 80 50 2 3 1 0.9993
5 0.5 80 30 2 3 2 0.9991 5 0.5 80 50 2 3 2 0.9993
5 0.5 80 30 3 2 1 0.9994 5 0.5 80 50 3 2 1 0.9994
5 0.5 80 30 3 2 2 0.9993 5 0.5 80 50 3 2 2 0.9994
5 0.5 80 30 4 1 1 0.9986 5 0.5 80 50 4 1 1 0.9986
5 0.5 80 30 4 1 2 0.9985 5 0.5 80 50 4 1 2 0.9986
Table 2.3: Fill Rate of Two-stage Systems for Normal Demand (b)
Chapter 3
in Inventory Models
3.1 Introduction
Inventories which encounter uncertain demand lead to risks of both excess sup-
ply and unsatisfied demand. The associated research on inventory models with
stochastic demand studies how best to balance these risks. Initially, the rela-
tive importance of the two risks was often parameterized with holding costs and
stockout costs. For the past twenty years, research has paid much attention to
the service role of inventories and focused on a service system’s fill rate, namely
60
61
the fraction of demand which is immediately met from on-hand inventory. During
this latter period, the relative importance of the two risks has often been param-
eterized with holdings costs and a lower bound on the fill rate. As a result, there
are parallel streams of literature which analyze identical models except that one
stream has stockout costs and the other has fill rate constraints.
the two approaches do not always yield the same results. This paper investigates
whether there is redundancy in the two streams of dynamic inventory models with
linear purchase costs, namely dynamic newsvendor models. We show the extent
to which optimal policies for either kind of model can be inferred from the other.
optimal for short, if it minimizes the long-run average sum of holding and stockout
minimizes the long-run average holding cost per unit time subject to a fill-rate
has a density function, then S-optimality and F-optimality are shown to be equiv-
(a) Corresponding to any unit stockout cost b, there is a base-stock level y and a
62
fill-rate f such that a base-stock policy with parameter y is both S-optimal with
level y and a stockout cost b such that a base-stock policy with parameter y is
both S-optimal with stockout cost b and F-optimal with constraint parameter f.
stockout cost b, a randomized base-stock level policy is F-optimal for most con-
mality can be accomplished via the other kind in the following sense:
(c) Corresponding to any unit stockout cost b, there is a base-stock level y and a
fill-rate f such that a base-stock policy with parameter y is both S-optimal with
(d) There are sequences of fill-rate constraint parameters 0 < f1 < f2 < · · · < 1,
base-stock level parameters y1 < y2 < · · ·, and stockout cost parameters b1 <
b2 < · · ·, with the following property. For any fill-rate constraint parameter f, say
fk ≤ f < fk+1 . Then the base-stock policy with parameter yk is both S-optimal
If demand is continuous, the consequence of (a) and (b) is that the struc-
ture and parametric analysis of F-optimal inventory policies are implicit in the
63
fore, it is redundant to analyze one if the other has been solved. If demand is
discrete, the consequence of (c) and (d) is that there are discrete sets of fill-rate
are interchangeable.
other than fill rate and stockout cost. For example, in §3.7 we sketch results for
balancing (i) inventory turnover ratios versus fill rates, and (ii) inventory costs
We provide a few portals to the large literature that is relevant to these issues.
Porteus (2002) and Zipkin (2000) are treatises on stochastic inventory models.
Also, see Porteus (1990) for a review of dynamic newsvendor models. Most of the
For example, Silver (1970), Yano (1985), and Platt, Robinson, and Freund (1997)
with batch ordering. The system faces a discrete compound Poisson demand
shows that an optimal policy consists of a mixed multistage echelon stock (R, nQ)
models with setup costs, focus on (s, S) policies where the order quantities are
s such that the required service level is achieved. Schneider and Ringuest consider
Boyaci and Gallego (2001) and Shang and Song (2003) study a periodic re-
view service-constrained serial inventory system where the leadtime demand for
the end product is Poisson distributed. Their service measure, the limiting prob-
ability of having positive on-hand inventory at the last stage, differs from the fill
rate. Boyaci and Gallego focus on base-stock policies, develop heuristic solutions,
and discuss the relationship between stockout cost and service-constrained models.
Shang and Song study the same model, and develop closed-form heuristics to ap-
paper concerns exactly optimal policies for periodic review single-stage models
Van Houtum and Zijm (2000) discuss the extent to which a cost-optimal policy
They consider several types of service contraints. When the demand distribution
optimal for an appropriate fill rate lower bound. We confirm their assertion under
more general conditions and examine when an F-optimal policy is also S-optimal.
We also show that the fill rate converges to their modified fill rate, characterize
65
the F-optimal policies with continuous and discrete demand distributions, and
establish the correspondences between backorder costs and fill rate contraints.
Chen and Zheng (1993) present nonlinear backorder costs in the setup cost in-
ventory model and shows that certain demand distributions guarantee the quasi-
convexity of cost rate function. Therefore, results derived from quasi-convex cost
functions remain valid. Rosling (2002a, 2002b) extends this inventory model
(Chen and Zheng 1993). Rosling (2002a) gives more general conditions than
that of Chen and Zheng (1993) on the demand process for which the cost func-
algorithm to find the optimal inventory policy. It also investigates the inventory
shortage costs and service constraints. Our paper focus on the dynamic newsven-
dor model and assumes one linear shortage cost which is the case in nearly all of
single-stage models with base-stock level policies. Glasserman and Tayur (1994),
Glasserman and Liu (1997), Glasserman (1997), Sobel (2004), and Chapter 2 of
this dissertation have other approximations and exact expressions for fill rates of
The paper has the following organization. §3.2 introduces the model and nota-
tion and §3.3 characterizes F-optimal policies when demand has a density function.
algorithms. §3.6 contains numerical examples and §3.7 generalizes and summa-
has a single product that faces independent and identically distributed nonneg-
0 < µ < ∞. Ordering decisions are made at the beginning of each period, and
each order takes L periods to deliver, i.e., items ordered in period n are deliv-
order-up-to level in period n, then the expected excess inventory and excess back-
L
ordered demand in period n + L are E[(a − j=0 Dn+j )+ ] (define u+ = max{u, 0})
L L
and E[( j=0 Dn+j − a)+ ], respectively. Let Bn+L (a) = ( j=0 Dn+j − a)+ and
L
Hn+L (a) = (a− j=0 Dn+j )+ , and let B(a) = E[Bn+L (a)] and H(a) = E[Hn+L (a)].
shipments due in periods 1, ..., L, let F N , KIN , and KSN denote, respectively, the
67
fill rate, the average end-of-period inventory level, and the average end-of-period
Let yn denote the inventory position after placing an order in period n, namely
the physical inventory (net of backlog) at the beginning of the period plus the sum
of shipments which will arrive at the beginning of periods n, n+1, · · · , n+L. Since
L−1
yn − j=0 Dn+j is the amount of goods that will be available to satisfy Dn+L ,
N L−1
N
FN = E min {Dn+L , yn − Dn+j } Dn+L
n=1 j=0 n=1
N
KIN = E Hn+L (yn ) /N
n=1
N
KSN = E Bn+L (yn ) /N
n=1
We assume that excess demand is backordered and order quantities are non-
negative. Let xn denote the inventory position net of backlog at the start of period
for all n. We assume that there are no planned backorders, so, yn ≥ max{xn , 0}.
inf π lim inf Eπ|x1 (hKIN ) : F ≥ f (3.56)
N →∞
inf π lim inf Eπ|x1 (hKIN + bKSN ) (3.57)
N →∞
68
inf π lim inf Eπ|x1 (hKIN − λF N ) (3.58)
N →∞
Problem (3.56) minimizes the average inventory cost subject to a fill rate con-
straint. Problem (3.57) minimizes the average sum of inventory and stockout
costs. Problem (3.58) is a Lagrangean with terms for the average inventory cost
and the fill rate. In (3.56) - (3.58) there is no loss of generality in the assumption,
henceforth made, that h = 1. In each problem, the primary issues are the form
that (3.58) is a special case of (3.57). The identity z = (z)+ − (−z)+ implies
L
Bn (a) = Hn (a) − a + Dn+j and B(a) = H(a) − a + (L + 1)µ
j=0
Rewriting F N as
N
N
FN = E [Dn+L − Bn+L (yn )] Dn+L
n=1 n=1
KIN − λF N
N L
=E [Bn+L (yn ) + yn − Dn+j ]/N
n=1 j=0
N
N
− λE [Dn+L − Bn+L (yn )] Dn+L
n=1 n=1
N L
=E [Bn+L (yn ) + yn − Dn+j ]/N
n=1 j=0
N N
N
− λE Dn+L − Bn+L (yn ) Dn+L
n=1 n=1 n=1
N L
=E [Bn+L (yn ) + yn − Dn+j ]/N
n=1 j=0
N
N
− λE 1 − Bn+L (yn ) Dn+L
n=1 n=1
N
The law of large numbers implies that n=1 Dn+L /N → µ with probability one
N N
n=1 Bn+L (yn )/N D limN →∞ inf n=1 Bn+L (yn )/N
lim inf N −→
N →∞ µ
n=1 Dn+L /N
D
where −→ denotes convergence in distribution.
Therefore,
N B N
Bn+L (yn )
n=1 n+L (yn ) 1 n=1
lim inf Eπ|x1 N = lim inf Eπ|x1 (3.59)
N →∞ Dn+L µ N →∞ N
n=1
1963, p. 182) implies that the right-side expectation in (3.59) exists if demand
70
observation
N
N n+L
0≤ Bn+L (yn )]/N ≤ Dj /N
n=1 n=1 j=n
ity (by the strong law of large numbers). Since 0 < µ < ∞, it follows that
N
E{ n=1 Bn+L (yn )]/N} exists and is finite for every N and has a convergent sub-
Employing (3.59),
Proposition 3.2.1
1 N
F = 1− lim sup Eπ|x1 B(yn)/N (3.60)
µ N →∞ n=1
71
Proof
N N
N n=1 [Dn+L − Bn+L (yn )] n=1 Bn+L (yn )/N
F =E N = 1−E N
n=1 Dn+L n=1 Dn+L /N
We note that Proposition 3.2.1 shows that the modified fill rate of Van Houtum
Let G(·) be the distribution function of DL+1 and let m = sup{a : G(a) = 0}
and Zipkin 2000) that yn = max{xn , y∗} for all n is an S-optimal policy with
b
y ∗ = G−1 ( ) (3.61)
b+1
real numbers.
3.2.1 permits replacement of the fill rate constraint with a constraint on the aver-
age backorder level. Therefore, an F-optimal policy solves the following problem:
N
inf lim inf Eπ|x1 H(yn )/N :
π N →∞
n=1
N
lim sup Eπ|x1 B(yn )/N ≤ µ(1 − f) (3.62)
N →∞
n=1
The following properties of H(·) and B(·) are useful in section §3.3. The
Lemma 3.2.2 H(·) is convex on + and is strictly increasing on (m, ∞), and
B(a) = 0 for a ≥ M.
η}, 0 ≤ η ≤ M − ν. Since H(·) and B(·) are strictly monotone and continuous
(due to convexity) on (m, ∞) and (0, M), respectively, they are injections; i.e.,
First we analyze the following N-period version of (3.62) and then let N → ∞:
N
N
inf
π
Eπ|x1 H(yn ) : Eπ|x1 B(yn ) ≤ Nµ(1 − f) (3.63)
n=1 n=1
The single-period newsvendor model with a fill rate constraint is the special
case of N = 1 in (3.63):
inf
π
Eπ|x1 [H(y1 )] : Eπ|x1 [B(y1)] ≤ µ(1 − f) (3.64)
The following result states that a base-stock policy is F-optimal for (3.64).
73
Proof Since H(·) is convex and increasing and B(·) is covex and decreasing,
the smallest y that satisfies the constraint achieves the constrained minimum of
H(·).
of the value of N.
Proof The claim is valid for N = 1 (Proposition 3.3.1). For any N > 1, rewrite
(3.63) as
N
N
inf Eπ H(yn ) : Eπ B(yn ) ≤ Nµ(1 − f) (3.66)
π
n=1 n=1
Since H (·) > 0 and B (·) < 0 on (m, M), if γ = 0, then (3.67) implies that
n = 1, · · · , N. But then,
N
E[B(yn )] = Nν = N(L + 1)µ > Nµ(1 − f) for all f > 0
n=1
So γ
= 0, and (3.68) yields
N
E[B(yn )] = Nµ(1 − f). (3.70)
n=1
So, y1 = y2 = · · · = yN because H(·) and B(·) are convex and monotone. There-
y1 = y2 = · · · = yN = y ∗ = B −1[µ(1 − f)]
all n. Therefore, adding the constraints yn ≥ xn for all n reduces the feasibility set
of (3.66), but does not affect the optimality of yn = y ∗ for all n (given x1 ≤ y ∗).
75
The transition from (3.63) to (3.62) is consistent with the large literature which
connects finite horizon and infinite horizon inventory models. We exploit the fact
that eventually the inventory level is at least as low as the back-stock level y ∗
(regardless of the initial inventory level). The proof is brief, straightforward, and
omitted.
Lemma 3.3.3 With probability one there is a period n∗ < ∞, such that xn ≤ y ∗
for all n ≥ n∗ .
Proposition 3.3.4 If demand has a density, then the base stock policy yn =
max{y ∗, xn } for all n, with y ∗ specified in (3.65), is F-optimal for (3.62), the
N
Proof For all N and for all π such that Eπ|x1 [ n=1 B(yn )] ≤ Nµ(1 − f), if
x1 ≤ y ∗ ,
N
NH(y ∗ ) ≤ Eπ|x1 H(yn ) (Lemma 3.3.2)
n=1
1 N
∗
H(y ) ≤ Eπ|x1 H(yn )
N n=1
∗ 1N
H(y ) ≤ lim inf Eπ|x1 H(yn )
N →∞ N n=1
∗
N 1 n
−1 N
lim inf Eπ∗ |x1 H(yn ) = lim inf Eπ∗ |x1 H(xn ) + H(y ∗ )
N →∞ N →∞ N n=1
n=1 n=n∗
76
1
N
= lim inf Eπ∗ |x1 H(y ∗)
N →∞ N n=n∗
= H(y ∗)
base-stock level B −1 [µ(1 − f)], is F-optimal when demand has a density. The
Demand and order quantities are integer-valued in this section. Let ξj = P {DL+1 =
j}, j = 0, 1, ..., M. For a target fill rate f < 1, identify k < M such that
B(k + 1) < µ(1 − f) ≤ B(k). If B(k) = µ(1 − f), it is clear from §3 that
the typical situation in which B(k + 1) < µ(1 − f) < B(k). Then yn = k for all
n is infeasible because the resulting fill rate is lower than f. If yn = k + 1 for all
n, then the fill rate is strictly higher than f and there is an opportunity to lower
the long-run average holding cost per period and still meet the fill rate constraint.
That is, a randomized policy dominates the best unrandomized policy. We note
prove that the following policy, labeled π ∗, is F-optimal among randomized poli-
Proof Without loss of generality let x1 ≤ k. Then {yn } is a Markov chain with
(1 − β)(1 − ξ0 ) β
λ∗k = λ∗k+1 =
1 − (1 − β)ξ0 1 − (1 − β)ξ0
Using (3.71),
1 ∗
F = 1− [λ B(k) + λ∗k+1 B(k + 1)]
µ k
1 [µ(1 − f) − B(k + 1)]B(k) + [B(k) − µ(1 − f)]B(k + 1)
= 1− =f
µ B(k) − B(k + 1)
Let Π(f) be the set of all randomized stationary policies which result in F ≥ f,
for π ∈ Π(f) let {λπj } be the stationary distribution of {yn } from initial state 0,
and write Hλπ for j H(j)λπj .
∗
Proof Since π ∗ ∈ Π(f) from Lemma 3.4.1, we show that Hλπ = min{Hλπ : π ∈
Π(f)}.
any stationary policy π, the inventory position {yn } is a Markov chain with states
The following linear program minimizes the average holding cost subject to a
λa ≥ 0, a = 0, · · · , M
The proof is completed with the following lemma which is proved in the Appendix.
Lemma 3.4.3 The optimal solution of (3.73) is λk = λ∗k , λk+1 = λ∗k+1 , and
λa = 0 for a
= k and a
= k + 1.
Let wja be the probability of ordering a − j units when the inventory level
λ∗k+1 ξk+1−j ), wj,k+1 = β(λ∗k ξk−j + λ∗k+1 ξk+1−j ) if j ≤ k wk+1,k+1 = λ∗k+1 ξ0 , and
wja = 0, otherwise.
79
We note that π ∗ is not uniquely optimal because other policies can induce the
3.5 Interchangeability
For b ≥ 0 and f ∈ [0, 1], let y S (b) = G−1 [b/(b + 1)] and y F (f) = B −1 [µ(1 − f)].
If b is the unit stockout cost, then y S (b) = G−1 [b/(b + 1)] is an S-optimal base-
stock level. If f constrains the fill rate and demand is continuous, then y F (f) =
focuses on two questions. First, for a unit stockout cost b, is there an f such
that y F (f) is an S-optimal base stock level? Second, if f constrains the fill rate,
y S (b) and y S (b) + 1 if demand is discrete)? Both questions have positive answers
Although the relation between b and f is more complicated with discrete demand,
if B[y F (f)] = µ(1 − f), then an F-optimal policy employs base stock level y F (f)
(with probability one). We say that demand has a strictly positive density g(·) if
Let F (y) be the fill rate induced by the base-stock policy with base-stock level
y F F [y S (b)] = y S (b) = G−1 [b/(b + 1)] (3.76)
y S G[y F (f)]/{1 − G[y F (f)]} ≥ y F (f) (3.77)
[0, 1] because both B(·) and B −1 (·) are continuous and strictly monotone. There-
fore, equating y S (b) = G−1 [b/(b + 1)] with y F (f) = B −1 [µ(1 − f)] and applying
B(·) to both sides yields (3.74) and (3.75). Similarly, Proposition 3.2.1 implies
F [y S (b)] = 1 − B{y S (b)}/µ. With the definitions of y F (·) and y S (·), this implies
(3.76). For (3.77), let a = y F (f) in y S {G(a)/[1 − G(a)]} = G−1 [G(a)] ≥ a with
equality if G has a strictly positive density. Finally, if demand has a strictly posi-
tive density, then y S (·) is an injection because both G(·) and G−1 (·) are continuous
If demand has a strictly positive density function, Proposition 3.5.1 states that
(a) Corresponding to any unit stockout cost b, a base-stock policy with base-
a density that is not strictly positive. In both instances, G(·) is not strictly
monotone which causes discontinuities in G−1 (·). If demand is discrete, recall the
1. If G[y F (f) − 1]/{1 − G[y F (f) − 1]} < b ≤ G[y F (f)]/{1 − G[y F (f)]}, then
B[y S (b)]/µ ≤ f < 1 − B[y S (b) + 1]/µ, then a randomized policy that employs
Proof For part 1, let b1 = G[y F (f)−1]/{1−G[y F (f)−1]} and b2 = G[y F (f)]/{1−
82
b1
G−1 ( ) = G−1 {G[y F (f) − 1]} = y F (f) − 1
b1 + 1
b2
G−1 ( ) = G−1 {G[y F (f)]} = y F (f)
b2 + 1
Since b1 < b ≤ b2, G−1 [b/(b + 1)] = y F (f). Part 2 follows from Proposition
3.4.2.
are random only in the former case. Nevertheless, a parametric analysis of either
kind of optimality can be accomplished via the other kind in the following sense:
optimal with constraint parameter f = 1 − B[y S (b)]/µ. Also, if 1 − B[y S (b)]/µ <
f < 1 − B[y S (b) + 1]/µ, then a policy which randomizes between base-stock levels
y F (f) is S-optimal for any stockout cost b in the interval G[y F (f) − 1]/{1 −
3.6 Examples
This section illustrates the results with numerical examples having continuous
positive density, and discrete demand. We shall clarify that qualitative results
In this subsection, the density function is g(a) = 1/10 and G(a) = a/10, a ∈
[0, 10]; so µ = 5. Therefore, G−1 (θ) = 10θ, B(y) = E[(D − y)+ ] = (10 − y)2/20,
√
and B −1(θ) = 10 − 20θ. Using (3.61), an S-optimal policy employs base-stock
level
b 10b
y S (b) = G−1 ( )=
b+1 b+1
In this example, y S (·) is concave and increasing. Using (3.65), an F-optimal policy
y F (f) = B −1[µ(1 − f)] = 10 − 20µ(1 − f) = 10 − 10 1 − f
In this example, y F (·) is convex and increasing. Simplifying y S (b) = y F (f) yields
Each equation in (3.78) uniquely defines the one-to-one mapping between b and
In this subsection, g(a) = 1/10 for a ∈ [0, 5] and a ∈ [6, 11], and otherwise
g(a) = 0. Therefore, µ = 5.5 and G(a) = a/10 if a ∈ [0, 5], G(a) = 1/2 if
a ∈ (5, 6], and G(a) = (a − 1)/10 if a ∈ (6, 11]. So G−1 (θ) = 10θ if 0 ≤ θ ≤ 1/2
and G−1 (θ) = 10θ + 1 if 1/2 < θ ≤ 1. That is, G−1 (·) is strictly increasing but
level ⎧
⎪
⎨ 10b/(b + 1), if 0 ≤ b ≤ 1
S −1b
y (b) = G ( )=
b+1 ⎪
⎩
(11b + 1)/(b + 1), if b > 1.
Although y S (·) is strictly increasing as in Section 3.6.1, it is neither concave nor
continuous due to the discontinuity of G−1 (·). As Figure 3.3 shows, no stockout
cost b induces y S (b) ∈ (5, 6]. However, y S (b) = G−1 [b/(b + 1)] is not necessarily
Continuing,
⎧ 2
⎪
⎪ (y − 20y + 110)/20 if 0 ≤ y ≤ 5
⎪
⎪
∞ ⎪
⎨
B(y) = (a − y)g(a)da = (17 − 2y)/4 if 5 < y ≤ 6
y ⎪
⎪
⎪
⎪
⎪
⎩
(11 − y)2/20 if 6 < y ≤ 11
Using (3.65), an F-optimal policy has base-stock level
⎧ √
⎪ 10 − 100 − 110f if 0 ≤ f ≤ 15/22
⎪
⎪
⎪
⎪
⎨
y F (f) = B −1 [µ(1 − f)] = ⎪ 11f − 5/2 if 15/22 < f ≤ 17/22
⎪
⎪
⎪
⎪
⎩ √
11 − 110 − 110f if 17/22 < f ≤ 1
As in Section 3.6.1, y F (·) is convex and increasing on [0, 1].
10b
= 10 − 100 − 110f for y ∈ [0, 5]
b+1
11b + 1
= 11 − 110 − 110f for y ∈ (6, 11]
b+1
b = [10 − 100 − 110f ]/ 100 − 11f ∈ [0, 1].
There are similar relationships for b ∈ (0, ∞] and f ∈ (17/22, 1], where
f = [11(b2 + 2b) + 1]/[11(b + 1)2 and b = [10 − 110 − 110f ]/ 110 − 110f .
86
(15/22, 17/22], y F (f) = 11f − 5/2 is an F-optimal base-stock level but there is no
Figure 3.4: Locus of {(b, f)} with the Same Optimal Base-Stock Level: Nonneg-
ative Density
(0 ≤ a ≤ 9), and G−1 (θ) = inf{a : G(a) ≥ θ}(0 ≤ θ ≤ 1). Define G−1 (0) = 0
S-optimal base stock levels. That is, y S (·) is piece-wise constant. Here, using
(3.65),
9
y F (f) = B −1[µ(1 − f)] = sup{y : (a − y) ≥ 45(1 − f)}
a=
y
That is, y F (·) is piece-wise constant (and discontinuous) although B(·) is con-
tinuous. Table 3.4 specifies G(·) and B(·) and Table 3.5 tabulates G−1 (·) and
B −1 (·). Tables 3.6 and 3.7 illustrate Proposition 3.5.2. Table 3.6 gives y S (b) as
b varies, together with the values of f for which y S (b) = y F (f). Table 3.7 gives
y F (f) as f varies, together with the values of b which cause y S (b) = y F (f). It is
apparent from Tables 3.6 and 3.7 that the parametric analysis of optimal policies
For example, when 7/3 < b ≤ 4, Table 3.6 shows that the S-optimal policy
has base-stock level 7 and induces a fill rate of 0.933. Table 3.7 confirms that
the F-optimal base-stock level is indeed 7 when f = 0.933. However, if 0.933 <
S B[yS (b)] S
b y S (b) 1 − B[yµ (b)] [1 − µ
− B[y µ(b)+1] )
,1
[0, 1/9] 0 0 [0, 0.2)
(1/9, 1/4] 1 0.2 [0.2, 0.377)
(1/4, 3/7] 2 0.377 [0.377, 0.533)
(3/7, 2/3] 3 0.533 [0.533, 0.667)
(2/3, 1] 4 0.667 [0.667, 0.778)
(1, 3/2] 5 0.778 [0.778, 0.867)
(3/2, 7/3] 6 0.867 [0.867, 0.933)
(7/3, 4] 7 0.933 [0.933, 0.978)
(4, 9] 8 0.978 [0.978, 1)
(9, ∞) 9 1 1
Table 3.6: S-optimal Base-Stock Levels and Fill Rates at which they are F-optimal
if f = 0.95, then (cf. Proposition 3.5.2), the following randomized policy strictly
dominates all deterministic policies. Order nothing when the inventory level is at
β. Using (3.71) and Table 3.4, γ = [B(7) − 4.5(1 − 0.95)]/[B(7) − B(8)] = 0.375
F
G[y (f )−1] F
G[y (f )]
f y F (f) ( 1−G[y F (f )−1] , 1−G[yF (f )] ]
Table 3.7: F-optimal Base-Stock Levels and Unit Stockout Costs at which they are
S-optimal
Numerous pairs of constraints and costs - not only fill rate constraints and stockout
costs - are interchangeable in the sense of §3.5. The key prerequisite is convexity
inventory turnover ratio versus fill rate. We model the turnover ratio as the ratio
of the demand to the average of the inventory levels at the beginning and ending
ratio, then the argument that leads to Proposition 1 shows that the constraint
corresponds to
N
lim sup E{ [yn + H(yn )]/N}/µ ≤ 2µ/τ
N →∞
n=1
We consider (i) maximization of the fill rate subject to this constraint, and (ii)
without loss of generality, and let Q(y) = H(y) + y. Obvious analogues of the
results in Sections 3 through 5 are valid for these problems because Q(·) is convex
and strictly increasing on (m, M). For example, if demand has a density, then the
following base-stock policy is optimal for problem (i): yn = max{y H , xn } for all n,
with y H = Q−1 (2µ/τ ). The following base-stock policy is optimal for problem (ii):
follows from the fact that y H and y T span the same set as τ and b span the
nonnegative numbers.
sections is the task of balancing inventory costs and stockout frequency. The task
can be formalized in several ways including (iii) minimizing the long-run average
function and m is the upper bound on stockout probability, then the following
base-stock policy is optimal for problem (iii): yn = max{y u, xn } for all n, with
to unity and c is the weight on stockout probability, then the following base-stock
policy is optimal for problem (iv): yn = max{y c, xn } for all n, where y c minimizes
H(y) + c[1 − G(y)]. Interchangeability follows from the fact that y u and y c span
the same set as m spans [0, 1] and c spans the nonnegative numbers.
91
stock levels. Also, there is a weakly monotone mapping from the set of S-optimal
policies to the set of F-optimal policies. For practical purposes, the parametric
3.8 Appendix
Since det(B) = B(k) − B(k + 1) > 0, for each λa the reduced cost is
Let ∆a be the numerator of (3.79); since B(k) − B(k + 1) > 0, δa and ∆a have
If a < k,
a = k − n − 1 and n = 0, · · · , k − 2
Since H(·) and B(·) are convex and monotone, ∆k−n−1 ≥ 0 because
If a > k + 1,
94
95