Académique Documents
Professionnel Documents
Culture Documents
i
probability of being in state i
F(i) cumulative distribution function for the failure states
S
D
optimal base-stock level for each warehouse in the decentralized system
S
C
optimal base-stock level for the warehouse in the centralized system
E[C] expected cost for an individual facility
E[C
D
] expected cost for the decentralized system
E[C
C
] expected cost for the centralized system
V [C] variance of the cost for an individual facility
V [C
D
] variance of the cost for the decentralized system
V [C
C
] variance of the cost for the centralized system
that it has been disrupted for i periods or fewer. Disruptions may be governed by a Markov
chain or a more general process. They follow the same process at every location and they are
independent over time and, in the decentralized system, across locations (each location can be
disrupted independently). In both systems, disruptions pause the ow of supply to the disrupted
stage, but any inventory at that stage may still be used during the disruption; therefore, disruptions
aect a stages supply-receiving function but not its demand-receiving function.
A holding cost of h per unit per period is incurred at the warehouse in the centralized system
or at each of the warehouses in the decentralized system. During non-disrupted period, this cost
is simply h times the base-stock level, less one periods worth of demand, d, observed that period.
However, during disruptions, the inventory continues to deplete by d each period until inventory is
exhausted and demand cannot be met. Unmet demands are backordered, and a stockout penalty
of p is incurred in both systems.
The parameters, decision variables (base-stock levels at each location), and performance mea-
sures for the system are summarized in Table 1. We use an additional subscript s to denote the
disrupted-supply model discussed in this section. Note that since S
D
is dened as the optimal
base-stock level for an individual warehouse in the decentralized system, the total inventory for
that system is nS
D
.
6
3.2 Mean and Variance of Optimal Cost
For a base-stock inventory policy at a single warehouse, the costs in a given period depend on the
state of the system, dened as the number of consecutive periods for which it has been disrupted.
Tomlin [2006] derives the expected cost for such a system, given in (1), and shows that the optimal
base-stock level is as given in (2):
E
s
[C] =
i=0
i
[h(S
s
(i + 1)d)
+
+p((i + 1)d S
s
)
+
] (1)
S
s
= jd, where j is the smallest integer such that F(j 1)
p
p +h
(2)
Schmitt et al. [2010] discuss these results, demonstrating that the optimal base-stock level is an
increasing step function of the newsboy fractile and the disruption probability. As the cost of
disruptions increases, the optimal solution increases by discrete jumps to the next whole periods
worth of demand.
We make use of these results in our analysis of the system. The relationships between the
inventory levels, expected costs, and cost variances for the centralized and decentralized systems
are given in the following theorem.
Theorem 1 In the multi-location system subject to supply disruptions, the decentralized, central-
ized, and single-facility optimal base-stock levels and performance measures are related as follows:
1. S
Cs
= nS
Ds
= nS
s
2. E
s
[C
C
] = E
s
[C
D
] = nE
s
[C
]
3. V
s
[C
C
] = nV
s
[C
D
] = n
2
V
s
[C
]
Proof : See Appendix, Section A.2.
Thus the total system inventory and expected costs are equal in the centralized and decentralized
systems, but the variance is n times greater for the centralized system, suggesting that, at least for
a risk-averse decision maker, the decentralized system is the better design for this system.
The intuition behind Theorem 1 is as follows. Because the total inventory is the same in both
systems, the total inventory devoted to each of the n sets of customers is the same in both systems.
Therefore each customer set feels the eects of disruptions for the same percentage of periods, on
average, in both systems. That is, a given disruption in either system causes a customer set to
experience the same number of stock-out periods. We state this formally in the following Corollary.
7
Corollary 2 In the multi-location system subject to supply disruptions, if base-stock levels are set
optimally, the distribution of consecutive periods for which a given warehouse is stocked out is
identical for the centralized and decentralized systems.
This causes the expected holding and stock-out costs to be the same. However, disruptions are
less frequent but more severe in the centralized system, and therefore that system has a greater
cost variance.
The relationships presented in Theorem 1 demonstrate the risk-diversication eect in the
multi-location system. If demand is deterministic, there is no risk-pooling benet from centraliza-
tion (since there is no risk from uncertain demand), thus expected costs for the centralized and
decentralized systems are equal. However, the risk from uncertain supply is better mitigated in the
decentralized system, as the lower variance for that system demonstrates. Snyder and Shen [2006]
demonstrate a similar result using simulation, though their model diers slightly from ours in that
they assume that inventory at a given stage may not be used when that stage is disrupted.
4 Multi-location System with Stochastic Demand
We now consider the classical model discussed by Eppen [1979]: the multi-location system with
deterministic supply and stochastic demand. Following Eppen, we assume that demand is normally
distributed with the same mean and variance
2
at each of the decentralized warehouses. There-
fore, the demand at the central warehouse in the centralized system has mean n and variance
2
C
= n
2
. We employ a subscript d to denote this model.
In Section 4.1, we review the results concerning the optimal base-stock levels and expected cost
presented by Eppen [1979]. We then evaluate the cost variance at optimality in Section 4.2. To
the best of our knowledge, ours is the rst study to consider the cost variance at optimality in this
system.
4.1 The Risk-Pooling Eect
The expected cost and optimal base-stock level for a single warehouse subject to stochastic demand
are well known:
E
d
[C] = h(S
d
) +(p +h)
1
_
S
d
_
(3)
S
d
= +
1
_
p
p +h
_
(4)
8
where
1
(x) =
_
x
(vx)(v)dv is the standard normal loss function. Therefore, for the centralized
system,
E
d
[C
C
] = h(S
Cd
n) +
C
(p +h)
1
_
S
Cd
n
C
_
(5)
S
Cd
= n +
C
1
_
p
p +h
_
(6)
Since the decentralized system functions as n single-wasehouse systems, E[C
Dd
] = nE
d
[C] and
S
Dd
= S
d
. We have the following Theorem:
Theorem 3 (Eppen, 1979) The optimal expected costs and inventory levels for the centralized
and decentralized multi-location systems subject to stochastic demand are as follows:
E
d
[C
D
] =
nE
d
[C
C
] (7)
nS
Dd
S
Cd
= (n
n)
1
_
p
p +h
_
(8)
Proof: Given by Eppen [1979].
This is the classical risk-pooling eect: by serving the warehouses demand from a centralized
inventory site, we pool the risk from demand uncertainty. As a result, the optimal expected cost
is
n times smaller for the decentralized system, and less inventory is required.
4.2 Cost Variance
The following theorem presents the relationship of cost variance for this system.
Theorem 4 At optimality, the cost variance for the centralized and decentralized multi-location
systems subject to stochastic demand are equal:
V
d
[C
C
] = V
d
[C
D
] (9)
Proof: See Appendix, Section A.3.
The theorem suggests that, under stochastic demand and deterministic supply, centralization
is optimal due to the risk-pooling eect, which aects the expected cost only and has no impact on
cost variance. In contrast, under supply disruptions and deterministic supply, decentralization is
optimal due to the risk-diversication eect, which aects the cost variance only and has no impact
on the expected cost. A natural question is, when both demand uncertainty and disruptions are
present, which system is optimal, i.e., which prevails: risk pooling or risk diversication? We
address this question in the next section.
9
5 Multi-location System with Supply Disruptions and Stochastic
Demand
In this section we compare the cost means and variances for centralized and decentralized multi-
location systems in which each warehouse is subject to random demand and the whole system is
subject to disrupted supply. We assume that when supply is not disrupted, the yield is deterministic.
We use a subscript b to denote this model since both supply and demand are stochastic. We
formulate the expected costs and cost variances in Sections 5.1 and 5.2, respectively. We perform
numerical analysis to compare the performance of the two systems in Section 5.3. Finally, in Section
5.4, we introduce a meanvariance objective that allows us to choose optimally between the two
systems when risk aversion is accounted for. We summarize our ndings in Section 5.5.
5.1 Expected Costs
At a single warehouse, a single periods demand is denoted D and is distributed as N(,
2
), while
the total demand in i periods is distributed as N(i, i
2
). The cost for a single warehouse subject
to normally distributed demand and disrupted supply is given by Schmitt et al. [2010] as:
E
b
[C] =
i=1
i1
_
h
_
S
b
(S
b
id)f
i
(id)dd +p
_
S
b
(id S
b
)f
i
(id)dd
_
=
i=1
i1
_
h(S
b
i) +
i(p +h)
1
_
S
b
i
i
__
(10)
Schmitt et al. [2010] prove that E
b
[C] is convex and argue that the optimal base-stock level cannot
be found in closed form because
_
Si
i
_
appears in the rst-order condition for i = 0, 1, . . . , .
Using (10), we derive the expected costs for the two systems in the following Proposition.
Proposition 5 When demand is normally distributed at the warehouses and supply is subject to
disruptions, the expected costs for the centralized and decentralized multi-location systems are:
E
b
[C
D
] = n
i=1
i1
_
h(S
Db
i) +
i(p +h)
1
_
S
Db
i
i
__
(11)
E
b
[C
C
] =
i=1
i1
_
h(S
Cb
in) +
in(p +h)
1
_
S
Cb
in
in
__
(12)
Proof: See Appendix, Section A.4.
Clearly, for the decentralized system, S
Db
= S
b
. For the centralized system, we have conrmed
numerically that S
Cb
= nS
b
in general, as in the deterministic-demand model in Section 3. However,
10
unlike that model, S
Cb
is neither consistently greater than nor consistently less than nS
b
when
demand is stochastic.
5.2 Cost Variance
We present the cost variance for arbitrary (not necessarily optimal) base-stock levels in the following
Proposition.
Proposition 6 When demand is normally distributed at the warehouses and supply is subject to
disruptions, the cost variances for the decentralized and centralized multi-location systems are:
V
b
[C
D
] = n
_
i=1
i1
_
h
2
(S
2
Db
2S
Db
i +i
2
+i
2
2
) + 2i
2
(p
2
h
2
)
2
_
S
Db
i
i
_
_
i=1
i1
_
h(S
Db
i) +
i(p +h)
1
_
S
Db
i
i
__
_
2
_
_
(13)
V
b
[C
C
] =
i=1
i1
_
h
2
(S
2
Cb
2S
Cb
ni +ni
2
+n
2
i
2
2
) + 2ni
2
(p
2
h
2
)
2
_
S
Cb
ni
ni
_
_
i=1
i1
_
h(S
Cb
ni) +
ni(p +h)
1
_
S
Cb
ni
ni
__
_
2
, (14)
where
1
() is the standard normal loss function and
2
() is the standard normal second-order loss
function.
Proof: See Appendix, Section A.5.
We have been unable to prove analytically that V
b
[C
C
] V
b
[C
D
]. However, we demonstrate
numerically in Section 5.3 that this inequality holds for every instance we tested. Thus, the risk-
diversication eect appears to prevail under supply disruptions, even when demand is stochastic.
5.3 Numerical Study
In this section, we perform a numerical study to compare the performance of the two systems
when supply disruptions and stochastic demand are present simultaneously. We consider n = 4
decentralized warehouses. We model disruptions as a two-state, discrete-time Markov process with
disruption probability and independent recovery probability . (That is, the probability of being
disrupted [not disrupted] next period given that the system is not disrupted [disrupted] in the
current period is given by [].)
11
Table 2: Parameter Levels
Variable Values
(failure probability) 0.05, 0.1, 0.2, 0.3, 0.4
(recovery probability) 0.25, 0.5, 0.75
(demand st. dev.) 5, 15, 25
p (penalty cost) 2, 5, 10, 15, 20
The steady-state probabilities for the disruption process are given by [Schmitt et al., 2010]:
0
=
+
;
i
=
+
(1 )
i1
, i 1 (15)
In Section 5.3.1, we vary the input parameters one by one to determine the impact of each
on the performance of the two systems. In Section 5.3.2 we examine the impact of the disruption
characteristics on the systems.
5.3.1 Impact of Parameters
We conducted a full-factorial experiment to determine the eect of the input parameters on the
optimal base-stock levels and the cost mean and variance at optimality in the two multi-location
systems. We varied , keeping xed to 100, and varied p, keeping h xed to 1. (This can be
done with nearly no loss of generality since the model is most sensitive to changes in the ratios
/ and p/(p + h) rather than to changes in the individual parameters.) In addition, we varied
the disruption parameters and . Table 2 lists the values tested for each parameter. The total
number of parameter combinations is 5 3 3 5 = 225.
For each combination of parameters, we optimized the expected cost functions (11) and (12)
numerically to solve for the optimal base-stock levels in both systems. Table 3 lists the mean,
median, minimum, and maximum, over the 225 instances, of several performance measures: dif-
ference in total inventory (column 2), dierence in expected cost (column 3), percent dierence in
expected cost (column 4), coecient of variation (CV) of the cost for the centralized and decen-
tralized systems (columns 5 and 6, resp.), and dierence in CV (column 7). The CV is given by
(C
) =
V [C
]
E[C
]
.
From Table 3 it is evident that, as suggested in Section 5.1, the decentralized system does
not always stock more than the centralized one, despite the risk-pooling eect; in 80 out of 225
cases (36%), the centralized system has a higher total base-stock level. This occurs because of
the discrete nature of disruptions. If demand is deterministic, then the optimal base-stock level is
always an integer multiple of the demand (see Section 3.2), and demand stochasticity perturbs these
12
Table 3: Comparison of Performance Measures
Inv. Di. E[Cost] Di. % E[Cost] Di. CV
C
CV
D
CV Di.
nS
Db
S
Cb
E
b
[C
D
] E
b
[C
C
]
E
b
[C
D
]E
b
[C
C
]
E
b
[C
C
]
(C
Cb
) (C
Db
) (C
Cb
) (C
Db
)
Mean 12 13.27 2.5% 2.12 1.03 1.09
Med. 14 9.10 0.9% 1.95 0.97 1.01
Min. -78 0.00 0.0% 0.89 0.44 0.38
Max. 89 53.64 41.9% 4.14 1.97 2.17
$-
$2,000
$4,000
$6,000
$8,000
$10,000
0.980 0.985 0.990 0.995 1.000
p/(p+h)
Eb[Cc]
SDb[Cc]
Eb[Cd]
SDb[Cd]
Figure 1: Centralized and Decentralized Expected Cost and Cost Standard Deviation vs.
p
p+h
discrete base-stock levels only slightly. Since the total standard deviation of demand is smaller in
the centralized system, the jumps are less smooth (closer to the deterministic-demand solution) and
may jump slightly above the optimal decentralized system levels. However, the optimal expected
cost of the centralized system is always less than or equal to that of the decentralized system.
Therefore, the expected-cost benet of the risk-pooling eect still exists when supply disruptions
are present, even though the ordering of the magnitude of base-stock levels does not.
We found that the dierence in expected cost is most sensitive to changes in and : it increases
as increases (disruptions become shorter) or as increases (demand is more uncertain). In these
cases, the impact of disruptions is smaller compared to that of the demand uncertainty, and so the
benets from risk pooling become more pronounced.
The cost variance is always greater for the centralized system, conrming the risk-diversication
eect. We found that the dierence in cost variance between the two systems generally increases as
p, , and the percent down-time (1
0
) increase, or as decreases. As the impact of disruptions
increases (either because they last longer increases or decreasesor because stockouts cost
more), the variance increases more for the centralized system than it does for the decentralized sys-
tem, because the risk-diversication eect mitigates the impact of disruptions in the decentralized
system.
13
Figure 1 plots the mean and standard deviation (SD) of cost for both systems as a function
of the newsboy fractile, p/(p + h). In this plot, we xed = 0.05, = 0.8, and = 15. The
mean and SD for both systems generally increase with the newsboy fractile. However, there is a
sharp decrease in the SD for the centralized system at p/(p +h) = 0.99. At this value, the optimal
base-stock level increases sharply from S
Cb
= 891 to S
Cb
= 1154. This large increase in inventory
results in a signicantly greater protection against disruptions and, as a result, in a reduction in
cost SD. The decentralized system has a similar, but less pronounced, decrease at the same point.
Note also from Figure 1 that the expected costs for the two systems are extremely close, while the
SD is substantially larger for the centralized system. We revisit this issue in Section 5.4.
5.3.2 Impact of Disruption Prole
Two pairs of (, ) values may result in the same fraction of periods disrupted, dened as
1
0
= /( +), (16)
but still have very dierent characteristics. For example, if and are both close to 0, then
disruptions are rare but long, while if they are both close to 1, then disruptions are frequent but
short. Snyder and Tomlin [2008] refer to the frequency/duration characteristics of a disruption as
the disruption prole. In this section, we study the eect of the disruption prole on the optimal
base-stock levels and performance measures of the two systems.
We generated ve groups of three (, ) pairs each. Within each group, the (, ) pairs have
the same percentage down-time (16) but a dierent disruption prole. The 15 instances are listed
in Table 4. The percent down-time increases with the group number (I, II, ...), while the disruption
length decreases with the letter within a group (a, b, c). In all instances, we set = 15 and p = 10
(the middle value from Table 2) and, as before, set h = 1 and = 100. Table 5 presents the optimal
base-stock levels and the mean and CV of cost for both systems.
The disruption prole has a signicant impact on the optimal base-stock levels and performance
measures. As the percent down-time increases, expected costs and optimal base-stock levels gen-
erally increase for both systems. Moreover, within a group (xed percent down-time), the optimal
base-stock levels and cost mean and variance are signicantly greater for rare/long disruptions than
for frequent/short ones. For example, instance I(a), which has a recovery probability () of less than
0.1, has particularly large CVs: 5.1 for the decentralized system and 11.4 for the centralized system.
Disruptions are rare in instance I(a), so the expected cost is small, but when disruptions occur, the
cost spikes, increasing the variance signicantly. Note also that in groups I and II, disruptions are
14
Table 4: Disruption Proles
Group % Down-Time
a 0.001 0.099
I b 1% 0.005 0.495
c 0.01 0.99
a 0.01 0.19
II b 5% 0.025 0.475
c 0.05 0.95
a 0.01 0.09
III b 10% 0.05 0.45
c 0.1 0.9
a 0.01 0.057
IV b 15% 0.05 0.283
c 0.1 0.567
a 0.01 0.04
V b 20% 0.05 0.2
c 0.1 0.4
Table 5: Base-Stock Levels and Performance Measures vs. Disruption Prole
Group S
Cb
E
b
[C
C
] (C
C
) nS
Db
E
b
[C
D
] (C
D
)
a 441.8 402.57 11.4 483.6 452.05 5.1
I b 441.8 130.17 7.3 483.6 179.53 2.6
c 441.8 89.76 4.0 483.6 139.12 1.2
a 451.5 1,075.26 5.7 502.8 1,103.96 2.7
II b 451.5 449.73 4.9 502.8 478.41 2.2
c 451.5 239.21 3.5 502.8 267.89 1.4
a 810.9 3,691.54 3.8 821.7 3,693.17 1.9
III b 764.6 854.80 3.0 729.2 860.72 1.5
c 745.9 411.87 1.2 691.7 419.30 0.8
a 2774.6 5,782.11 2.5 2751.4 5,784.26 1.3
IV b 1205.8 1,791.29 2.7 1211.5 1,798.20 1.4
c 821.7 812.71 2.4 843.3 826.61 1.2
a 4374.9 6,844.85 1.9 4357.5 6,847.41 1.0
V b 2009.5 2,994.01 2.4 2017.9 3,000.02 1.2
c 1214.0 1,402.56 2.2 1227.8 1,413.12 1.1
15
Table 6: Comparison of Performance Measures vs. Disruption Prole
Inv. Di. E[Cost] Di. % E[Cost] Di. CV Di.
Group nS
Db
S
Cb
E
b
[C
D
] E
b
[C
C
]
E
b
[C
D
]E
b
[C
C
]
E
b
[C
C
]
(C
C
) (C
D
)
a 41.8 49.48 12.29 % 6.4
I b 41.8 49.36 37.92 % 4.7
c 41.8 49.36 54.99 % 2.8
a 51.5 28.70 2.67 % 2.9
II b 51.5 28.68 6.38 % 2.7
c 51.5 28.68 11.99 % 2.1
a 10.8 1.62 0.04 % 1.9
III b -35.4 5.92 0.69 % 1.5
c -54.1 7.43 1.80 % 0.5
a -23.3 2.15 0.04 % 1.3
IV b 5.7 6.91 0.39 % 1.4
c 21.6 13.89 1.71 % 1.3
a -17.4 2.56 0.04 % 1.0
V b 8.4 6.01 0.20 % 1.2
c 13.8 10.57 0.75 % 1.1
rare enough that it is optimal to barely protect against them at allthe optimal base-stock levels
are almost the same as in the no-disruption case (in which S
C
= 440.1 and S
D
= 480.1). Therefore,
disruptions are quite costly when they occur. These results conrm results by Snyder and Tomlin
[2008], Tomlin [2006], and others that suggest that rare/long disruptions are more dicult to plan
for than frequent/short ones.
Table 6 compares the performance of the centralized and decentralized systems for these in-
stances. Either system may have a greater total base-stock level (as in Table 3), and there is no
clear relationship between the dierence in base-stock levels and the disruption prole. However, as
the mean disruption duration increases (within a group), the dierence in expected cost decreases
and the dierence in cost variance increases. In other words, when disruptions are longer, the
risk-pooling eect is less pronounced and the risk-diversication eect is more pronounced.
5.4 Optimal Risk-Averse Design
We have shown that for the OWMR system with both supply disruptions and stochastic demand,
inventory decentralization reduces the cost variance via the risk-diversication eect while central-
ization reduces the expected cost via the risk-pooling eect. It is natural to ask which eect is
more pronounced, and therefore which system a risk-averse decision maker should prefer.
16
5.4.1 MeanVariance Objective
We employ the classical meanvariance approach, minimizing the following objective function:
(1 )E[C] +V [C] (17)
where [0, 1]. The larger is, the more risk-averse the decision maker is. Many other objectives
for risk-averse decision making have been proposed in the nance and operations literature. We
chose to focus on the meanvariance objective primarily for analytical tractability. Numerical
studies using the Conditional Value-at-Risk (CVaR) objective produced similar insights, but a more
formal theoretical analysis was precluded by the increased complexity of the objective function.
We optimized (17) numerically using a line-search procedure. This function is neither convex
nor concave for all values of . There are values of and S
b
such that (17) is convex and values for
which it is not convex. By testing the function numerically, we found that for 0.05, the range
we used in our tests below, (17) is convex at nearly every value of S
b
, with the exception of S
b
0.
We also found that the function is convex for all S , in the instances we tested, and this range
typically includes the optimal base-stock level. Moreover, the second derivative of (17) is positive
at the solution found by our optimization procedure for every instance we tested. Therefore, we
are condent that our results represent the globally optimal solutions for each instance.
5.4.2 Numerical Study
Our numerical study shows that even for very small values of , the decentralized system is almost
always optimal under the meanvariance objective. We set = 100, = 15, = 0.025, and
= 0.5, resulting in a percent down-time of 4.76%. Figure 2 plots the regions in which each
system is optimal for [0, 0.05] and p/(p + h) [0.5, 1) (keeping h xed to 1, i.e., p [1, )).
The region in which the centralized system is optimal is barely visible. Figure 3 zooms in on the
range [0, 0.001] so that this region is more visible. Clearly, the centralized system is optimal
only for very risk-neutral decision makers operating under low service levels: for 0.0008 and
p/(p +h) 0.5, the decentralized system is always optimal.
The centralized system can also be optimal for systems with very low disruption risk. For
example, Figure 4 plots the optimal systems for 0.12% down-time ( = 0.001 and = 0.8).
However, even in this case, the decentralized system is optimal for all 0.02 and p/(p+h) 0.75.
We next examined the impact of and on the optimal design. In Figure 5, we set p/(p+h) =
0.9 and = 0.001. The graph shows that the decentralized system is always optimal when the
17
0.5
0.6
0.7
0.8
0.9
1
0.00 0.01 0.02 0.03 0.04 0.05
kappa
n
e
w
s
b
o
y
f
r
a
c
t
i
l
e
Decentralized
Optimal
Centralized
Optimal
Figure 2: Optimal Inventory System for 4.76% Down-Time
0.5
0.6
0.7
0.8
0.9
1
0.0000 0.0002 0.0004 0.0006 0.0008 0.0010
kappa
n
e
w
s
b
o
y
f
r
a
c
t
i
l
e
Centralized
Optimal
Decentralized
Optimal
Figure 3: Optimal Inventory System for 4.76% Down-Time: Narrower Range
0.5
0.6
0.7
0.8
0.9
1
0.00 0.01 0.02 0.03 0.04 0.05
kappa
n
e
w
s
b
o
y
f
r
a
c
t
i
l
e
Centralized
Optimal
Decentralized
Optimal
Figure 4: Optimal Inventory System for 0.12% Down-Time
18
0
0.2
0.4
0.6
0.8
1
0.000 0.002 0.004 0.006 0.008 0.010
alpha
b
e
t
a
Centralized
Optimal
Decentralized
Optimal
Figure 5: Optimal Inventory System for Varying and Values
failure probability exceeds 0.7%, regardless of the recovery probability.
These gures all employ small values of and the newsboy fractile and reasonably small dis-
ruption probabilities. They demonstrate that the decentralized system is far more likely to be the
optimal inventory design for a risk-averse rm.
Since the variance is typically several orders of magnitude greater than the expected cost for
our model, one could argue that the mean-variance objective is naturally biased toward the lower-
risk strategy, i.e., the decentralized system. Therefore, we also performed similar tests using the
standard deviation of the cost in place of the variance, and with values roughly equal to the
square root of the original values. The results were very similar. For the same parameter values
discussed above, we found that when the disruption probability is 4.76%, the decentralized system
is optimal for all 0.0075 and p/(p + h) 0.5, and when the disruption probability is 0.12%,
the decentralized system is optimal for all 0.1 and p/(p + h) 0.75. We also varied and
as above and found that when = 0.001 and p/(p + h) = 0.9, the decentralized system is always
optimal when the failure probability exceeds 1.8%, regardless of the recovery probability. A more
detailed description of these results is omitted due to space considerations.
5.5 Summary of Numerical Comparisons
For the multi-location system subject to both demand uncertainty and supply disruptions:
The risk-diversication eect exists, resulting in lower cost variance in the decentralized
system
The risk-diversication eect is more pronounced when disruptions are longer and more
costly
The risk-pooling eect exists, resulting in lower expected cost in the centralized system
The risk-pooling eect does not always result in lower total base-stock levels
19
The risk-pooling eect is more pronounced when disruptions are less frequent and
shorter, or when demand is more variable
Disruption proles aect the optimal solutions
For the same percent down-time, longer disruptions result in higher optimal base-stock
levels, higher expected costs, and higher cost variances
The dierence in cost variances between the centralized and decentralized systems is
greater for longer disruptions, but the dierence in expected cost is greater for shorter
disruptions.
For a risk-averse rm, the decentralized system is typically the optimal inventory design
This holds unless service level (newsboy fractile), value, or failure probabilities are
very low
6 Conclusions
In this paper, we consider a multi-location system with both supply and demand uncertainty. We
investigate the risk-diversication eect, in which the expected cost is the same in the centralized
and decentralized systems but the cost variance is smaller in the decentralized system. The intuition
behind this eect is that by distributing inventory at multiple sites, the impact of any one disruption
is smaller, even though each site is still aected by the same number of disruptions. The rm
benets from not putting all its eggs in one basket; although the same number of eggs may be
destroyed, they are not all destroyed at once. We proved that the risk-diversication eect occurs
in multi-location systems with supply disruptions.
In contrast, the classical risk-pooling eect prevails under demand uncertainty; under this eect,
the expected cost is smaller in the centralized system, and we prove that the variance is equal in
the two systems. We showed numerically that, when disruptions and demand uncertainty are
both present, both eects occur to a certain extent, but that in most cases the risk-diversication
eect strongly dominates the risk-pooling eect. Therefore, a decentralized system is optimal when
disruptions and demand uncertainty are both present, except when the service level (newsboy
fractile) is very low, the rm is very risk neutral, and/or the system is very reliable.
Natural extensions to this research can be made by relaxing the assumptions in our models. For
example, we assumed zero xed-ordering costs. An argument could be made that this ignores an
additional cost benet from centralization; synchronizing orders could reduce the total xed costs.
However, as we have shown, risk diversication would suggest that a risk-averse rm should not
synchronize orders if individual shipments may be disrupted at the supplier.
20
Another extension could relax our assumption of independently distributed disruptions. Eppen
[1979] shows that demand correlation aects the magnitude of the risk pooling eect; positively
correlated demand decreases the risk pooling benets of centralization and negatively correlated
demand increases them. A similar study could be conducted for supply uncertainty. We would
expect that positively correlated supply disruptions would decrease the risk diversication benets
of decentralization, whereas negatively correlated disruptions would increase those benets.
Another obvious extension to our work could address the assumption that inventory can only
be held at one echelon of the supply chain, which we made to simplify our analysis and clarify our
conclusions. Of course, centralized and decentralized systems are not the only choices available to a
rm operating in a multi-location environment; the rm may also choose a hybrid system in which
inventory is held at both echelons. Such a system may well provide a desirable balance between
the risk-diversication and risk-pooling eects. These systems are signicantly harder to analyze
(as the literature on multi-location system attests), but an investigation of these two competing
eects in this context is an important avenue for future research.
7 Acknowledgments
This research was supported in part by National Science Foundation grants DGE-9972780, DMI-
0522725, and DMI-0621433. This support is gratefully acknowledged.
References
S. Axsater. Inventory Control. Kluwer Academic Publishers, Boston, MA, rst edition, 2000.
V. Babich, A.N. Burnetas, and P. H. Ritchken. Competition and diversication eects in supply
chains with supplier default risk. Manufacturing & Service Operations Management, 9(2):123
146, 2007.
E. Berk and A. Arreola-Risa. Note on Future supply uncertainty in EOQmodels. Naval Research
Logistics, 41:129132, 1994.
X. Chen, M. Sim, D. Simchi-Levi, and P. Sun. Risk aversion in inventory management. Operations
Research, 55(5):828842, 2007.
S. Chopra, G. Reinhardt, and U. Mohan. The importance of decoupling recurrent and disruption
risks in a supply chain. Naval Research Logistics, 54(5):544555, 2007.
M. Dada, N. Petruzzi, and L. Schwarz. A newsvendors procurement problem when suppliers are
unreliable. Manufacturing & Service Operations Management, 9(1):932, 2007.
L. Eeckhoudt, C. Gollier, and H. Schlesinger. The risk-averse (and prudent) newsboy. Management
Science, 41(5):786794, 1995.
21
G. D. Eppen. Eects of centralization on expected costs in a multi-location newsboy problem.
Management Science, 25(5):498501, 1979.
A. Federgruen and N. Yang. Selecting a portfolio of suppliers under demand and supply risks.
Operations Research, 56(4):916936, 2008.
W. J. Hopp and Z. Liu. Protecting supply chain networks against catastrophic failures. Work-
ing paper, Dept. of Industrial Engineering and Management Science, Northwestern University,
Evanston, IL, 2006.
M. Parlar and D. Berkin. Future supply uncertainty in EOQ models. Naval Research Logistics,
38:107121, 1991.
A. J. Schmitt and L. V. Snyder. Innite-horizon models for inventory control under yield uncertainty
and disruptions. Computers and Operations Research, 39:850862, 2012.
A. J. Schmitt, L. V. Snyder, and Z. J. M. Shen. Inventory systems with stochastic demand and
supply: Properties and approximations. European Journal of Operational Research, 206(2):313
328, 2010.
L. V. Snyder. A tight approximation for a continuous-review inventory model with supplier dis-
ruptions. Working paper, P.C. Rossin College of Engineering and Applied Sciences, Lehigh
University, Bethlehem, PA, 2008.
L. V. Snyder and Z. J. M. Shen. Supply and demand uncertainty in multi-echelon supply chains.
Working paper, P.C. Rossin College of Engineering and Applied Sciences, Lehigh University,
Bethlehem, PA, 2006.
L. V. Snyder and B. Tomlin. Inventory management with advanced warning of disruptions. Working
paper, P.C. Rossin College of Engineering and Applied Sciences, Lehigh University, Bethlehem,
PA, 2008.
L. V. Snyder, Z. Atan, P. Peng, Y. Rong, A. J. Schmitt, and B. Sinsoysal. Supply chain disruptions:
A review. Working paper, P.C. Rossin College of Engineering and Applied Sciences, Lehigh
University, Bethlehem, PA, 2010.
B. Tomlin. On the value of mitigation and contingency strategies for managing supply chain
disruption risks. Management Science, 52(5):639657, 2006.
B. Tomlin. Disruption-management strategies for short life-cycle products. Naval Research Logis-
tics, 56:318347, 2009.
B. Tomlin and Y. Wang. On the value of mix exibility and dual sourcing in unreliable newsvendor
networks. Manufacturing & Service Operations Management, 7(1):3757, 2005.
J. A. Van Meighem. Risk mitigation in newsvendor networks: Resource diversication, exibility,
sharing and hedging. Management Science, 53(8):12691288, 2007.
C. A. Yano and H. L. Lee. Lot sizing with random yields: A review. Operations Research, 43(2):
311334, 1995.
P. H. Zipkin. Foundations of Inventory Management. McGraw-Hill Higher Education, Boston,
MA, rst edition, 2000.
22
A APPENDIX: PROOFS
A.1 Loss Functions
Let X be a random variable with pdf f(x) and cdf F(x). The loss function for X is given by:
G(x) =
_
x
(t x)f(t)dt =
_
x
(1 F(t))dt (18)
and the second-order loss function is:
H(x) =
1
2
_
x
(t x)
2
f(t)dt =
_
x
G(t)dt. (19)
Let
1
(z) be the standard normal loss function and
2
(z) as the standard normal second-order
loss function.
The following lemma presents properties we use in the proofs below. The proof of the lemma is
omitted; it follows from well known results concerning loss functions [e.g., Axsater, 2000, Zipkin,
2000].
Lemma 7 1. Let X have cdf F and loss function G. Then for any function (x),
d
dx
G((x)) =
2
_
S
_
. (21)
A.2 Proof of Theorem 1, Section 3.2
A.2.1 Optimal Base-Stock Levels
S
Ds
= S
s
since each warehouse acts as a single-facility system. Moreover, from (2) we know that
S
s
= jd, where j is the smallest integer such that F(j 1)
p
p+h
. Similarly, in the centralized
system, the optimal base-stock level is S
Cs
= jnd for the same j (since the denition of j does not
depend on the demand). Therefore, S
Cs
= nS
s
.
A.2.2 Expected Costs at Optimality
In the decentralized system, the total expected cost is simply the sum of the single-facility costs,
so E
s
[C
D
] = nE
s
[C].
23
In the centralized system, the optimal expected cost at the centralized warehouse, using nS
s
in
place of S
Cs
in (1), is given by:
E
s
[C
C
] =
i=1
i
[h(nS
s
(i + 1)nd)
+
+p((i + 1)nd nS
s
)
+
]
= n
i=1
i
[h(S
s
(i + 1)d)
+
+p((i + 1)d S
s
)
+
]
= nE
s
[C
] = E
s
[C
D
] (22)
A.2.3 Cost Variances at Optimality
For the decentralized system, since each warehouse operates independently, the variance of the cost
is equal to the sum of the single-warehouse variances, or V
s
[C
D
] = nV
s
[C].
Now, for a single warehouse,
E
s
[(C
)
2
] =
i=1
i
_
h
2
_
(S
s
(i + 1)d)
+
_
2
+p
2
_
((i + 1)d S
s
)
+
_
2
_
(23)
In the centralized system, using S
Cs
= nS
s
,
E
s
[(C
C
)
2
] =
i=1
i
_
h
2
_
(nS
s
(i + 1)nd)
+
_
2
+p
2
_
((i + 1)nd nS
s
)
+
_
2
_
= n
2
i=1
i
_
h
2
_
(S
s
(i + 1)d)
+
_
2
+p
2
_
((i + 1)d S
s
)
+
_
2
_
= n
2
E
s
[(C
)
2
] (24)
Thus, using (22) and (24),
V
s
[C
C
] = E
s
[(C
C
)
2
] E
s
[C
C
]
2
= n
2
E
s
[(C
)
2
] (nE
s
[C
])
2
= n
2
(E
s
[(C
)
2
] E
s
[C
]
2
)
= n
2
V
s
[C
] (25)
Thus V
s
[C
C
] = nV
s
[C
D
].
A.3 Proof of Theorem 4, Section 4.2
At optimality, the centralized and decentralized base-stock levels satisfy
S
Dd
=
_
p
p +h
_
=
S
Cd
n
C
. (26)
Let
X
_
p
p +h
_
(27)
24
Using (21) from Lemma 7,
E
d
[C
2
] = h
2
_
S
d
(S
d
d)
2
f(d)dd +p
2
_
S
d
(d S
d
)
2
f(d)dd (28)
= h
2
_
(S
2
d
2S
d
d +d
2
)f(d)dd + (p
2
h
2
)
_
S
d
(d S
d
)
2
f(d)dd (29)
= h
2
(S
2
d
2S
d
+
2
+
2
) + 2
2
(p
2
h
2
)
2
_
S
d
_
(30)
At optimality, we have
E
d
[(C
)
2
] = h
2
((S
d
)
2
2S
d
+
2
+
2
) + 2
2
(p
2
h
2
)
2
(X) (31)
In addition, from (3),
E
d
[C
] = h(S
d
) +(p +h)
1
(X) . (32)
Therefore,
V
d
[C
] = E
d
[(C
)
2
] E
d
[C
]
2
= h
2
((S
d
)
2
2S
d
+
2
+
2
) + 2
2
(p
2
h
2
)
2
(X)
_
h(S
d
) +(p +h)
1
(X)
2
= h
2
((S
d
)
2
2S
d
+
2
+
2
) + 2
2
(p
2
h
2
)
2
(X)
_
h
2
((S
d
)
2
2S
d
+
2
) + 2h(p +h)(S
d
)
1
(X) +
2
(p +h)
2
1
(X)
2
= h
2
2
+ 2
2
(p
2
h
2
)
2
(X) 2h(p +h)(S
d
)
1
(X)
2
(p +h)
2
1
(X)
2
=
2
_
h
2
+ 2(p
2
h
2
)
2
(X) 2h(p +h)X
1
(X) (p +h)
2
1
(X)
2
(33)
Let Y equal the sum of the terms inside the brackets. Note that Y is a function of X, not of or
alone; therefore, Y is the same for an individual warehouse or for the centralized warehouse in
the centralized system.
Since the decentralized system functions as n individual warehouses, we have
V
d
[C
D
] = nV
d
[C
]. (34)
Applying (33),
V
d
[C
C
] =
2
C
Y = n
2
Y = nV
d
[C
] = V
d
[C
D
]. (35)
as desired.
25
A.4 Proof of Proposition 5, Section 5.1
E
b
[C
D
] is simply equal to nE
b
[C]. Similarly, E
b
[C
C
] = nE
b
[C], but with the pooled demand
standard deviation equal to
i=1
i1
_
h
2
_
S
b
(S
b
id)
2
f
i
(id)dd +p
2
_
S
b
(id S
b
)
2
f
i
(id)dd
_
=
i=1
i1
_
h
2
_
(S
2
b
2S
b
id +i
2
d
2
)f
i
(id)dd + (p
2
h
2
)
_
S
b
(id S
b
)
2
f
i
(id)dd
_
=
i=1
i1
_
h
2
(S
2
b
2S
b
i +i
2
+i
2
2
) + 2i
2
(p
2
h
2
)
2
_
S
b
i
i
__
(36)
V
b
[C] = E
b
[C
2
] (E
b
[C])
2
=
i=1
i1
_
h
2
(S
2
b
2S
b
i +i
2
+i
2
2
) + 2i
2
(p
2
h
2
)
2
_
S
b
i
i
__
i=1
i1
_
h(S
b
i) +
i(p +h)
1
_
S
b
i
i
__
_
2
(37)
In the decentralized system, the total variance is just the sum of the individual-warehouse variances
since each retailer operates independently. In the centralized system, the total demand seen by the
centralized warehouse is distributed as N(ni, ni
2
). Therefore,
V
b
[C
C
] =
i=1
i1
_
h
2
(S
2
Cb
2S
Cb
ni +ni
2
+n
2
i
2
2
) +
2ni
2
(p
2
h
2
)
2
_
S
Cb
ni
ni
_
_
i=1
i1
_
h(S
Cb
ni) +
ni(p +h)
1
_
S
Cb
ni
ni
__
_
2
(38)
26