PROBABILISTIC ANALYSES AND PRACTICAL ALGORITHMS FOR

INVENTORY-ROUTING MODELS
LAP MUI ANN CHAN
Philips Laboratories, Briarcliff Manor, New York

AWI FEDERGRUEN
Columbia University, New York

DAVID SIMCHI-LEVI
Northwestern University, Evanston, Illinois
(Submitted February 1994; revisions received June 1995, November 1995; accepted January 1996)

We consider a distribution system consisting of a single warehouse and many geographically dispersed retailers. Each retailer faces
demands for a single item which arise at a deterministic, retailer specific rate. The retailers’ stock is replenished by a fleet of vehicles
of limited capacity, departing and returning to the warehouse and combining deliveries into efficient routes. The cost of any given
route consists of a fixed component and a component which is proportional with the total distance driven. Inventory costs are
proportional with the stock levels. The objective is to identify a combined inventory policy and a routing strategy minimizing
system-wide infinite horizon costs. We characterize the asymptotic effectiveness of the class of so-called Fixed Partition policies and
those employing Zero Inventory Ordering. We provide worst case as well as probabilistic bounds under a variety of probabilistic
assumptions. This insight is used to construct a very effective algorithm resulting in a Fixed Partition policy which is asymptotically
optimal within its class. Computational results show that the algorithm is very effective on a set of randomly generated problems.

I n many distribution systems, important cost savings can

be achieved by integrating inventory control and routing
decisions, i.e., by determining simultaneously the timing
incoming orders from outside vendors, but which do not
keep stock themselves.
The distribution planning problem associated with a
and sizes of the retailer deliveries as well as efficient vehi- cross-docking strategy can be modeled as follows: a single
cle schedules so as to minimize total transportation and warehouse serves many retailers which are geographically
inventory carrying costs. In this type of systems the “ware- dispersed in a given area. Stock for a single item is deliv-
house” and the “retailers” may represent (part of) consec- ered to the retailers by a fleet of vehicles of limited capac-
utive layers in the distribution network of a single ity. Each retailer faces a deterministic, retailer specific,
company; alternatively, customers may be external, as in demand rate. Inventory holding costs are accrued at a con-
the increasingly popular “vendor managed” or “direct re- stant rate, which is assumed to be identical for all retailers.
plenishment” arrangements in which vendors assume the No inventory is kept at the warehouse. Each time a vehicle
responsibility of maintaining their customers’ inventories is sent out to replenish inventory, it incurs a fixed cost
instead of responding to customer generated orders. (independent of the specific route driven) plus a cost pro-
The impact of integrated inventory and routing strate- portional to the total distance traveled by the vehicle. The
gies was recently emphasized by Stalk et al. (1992) who objective is to determine an inventory policy and a routing
review the evolution of the discount retailing industry. strategy such that each retailer can meet its demands and
They observe Wal-Mart developing into the largest and the long-run average transportation and inventory costs
highest profit retailer in the world. This success story was are minimized.
attributed by Stalk et al. to a relentless focus on satisfying In a distribution system of this type, one may have an
retailer needs by efficient logistical design and planning. additional constraint limiting the frequency with which
“The key to achieving these goals was to make the way the each retailer is visited. Such a constraint may, for example,
company replenished inventory the centerpiece of its compet- be due to limited material handling capacity and/or due to
itive strategy.” Stalk et al. identify a number of major com- the set-up time required for unloading deliveries at the
ponents in this strategic vision, most importantly, a retailers.
logistics technique referred to as “cross-docking.” This re- It is highly improbable that an optimal strategy will ever
fers to a distribution strategy in which the stores are sup- be identified for this model; such attempts have long been
plied by central warehouses which act as coordinators of abandoned even for far simpler models, e.g., the special
the supply process, and as transshipment points for case where the cost of dispatching a vehicle to a group of

Subject classifications: Transportation: vehicle routing. Inventory/production: multiechelon.

Area of review: OPTIMIZATION.

Operations Research 0030-364X/98/4601-0096 $05.00

Vol. 46, No. 1, January–February 1998 96 q 1998 INFORMS
 CHAN, FEDERGRUEN, AND SIMCHI-LEVI / 97
retailers consists only of the fixed component and is inde- packing, that a 5 1, in which case an FP (ZIO) strategy is
pendent of the distance traveled. Models, with joint re- asymptotically optimal while for other common distribu-
plenishment costs of this type, are often referred to as tions a is close to one, see Coffman and Lueker (1991) and
Joint Replenishment Problems, see Jackson et al. (1985) Rhee (1988) for a more detailed discussion and character-
and Federgruen and Zheng (1992). Most importantly, the ization of the packing constant.
structure of a (fully) optimal strategy is so complex that it To put these results in perspective we now review the
would fail to be implementable even if it could be deter- different existing approaches developed for the model con-
mined in a reasonable amount of time. As a consequence, sidered here. This model was first introduced by Anily and
various authors have restricted themselves up front to spe- Federgruen (1990). These authors, restricted their analysis
cific classes of strategies and developed methods to iden- to a class of replenishment strategies C with the following
tify optimal or asymptotically optimal rules within the properties: a replenishment strategy in C specifies a collec-
chosen class. tion of regions (subset of retailers); if a retailer belongs to
To date, nothing appears to be known on how much is several regions a specific fraction of its sales is assigned
lost by restricting oneself to any of these classes of strate- to each of these. Each time one of the outlets in a given
gies. It is noteworthy that all of the proposed classes of region gets a delivery, this delivery is made by a vehicle
policies are subsets of the class of Zero Inventory Order- which visits all other outlets in the region as well. Observe
ing (ZIO) policies, under which a retailer is replenished if that a large amount of flexibility is preserved by allowing
and only if its inventory is down to zero. In the absence of for overlapping regions of retailers but this may overesti-
constraints on the vehicle capacity or the frequency with mate inventory costs for split retailers. The generated par-
which retailers can be served, it is easily verified that a titions have an asymptotically insignificant number of split
ZIO policy is optimal. However, in the presence of these retailers; see Hall (1991) and Anily and Federgruen (1991)
constraints, ZIO policies may fail to be optimal, as we shall for details. With n retailers, Anily and Federgruen show
demonstrate shortly.
that regions can be formed by a simple regional partition-
Even the structure of an optimal ZIO policy is too com-
ing scheme and a combined inventory and routing strategy
plex to permit implementation or identification by a rea-
can thus be computed in O(n log n) time, which is asymp-
sonable algorithm; this is why all the literature on this
totically optimal within the class C.
model has restricted itself to specific subclasses of the ZIO
Subsequent work considers restrictions to other classes
policies. One attractive such class are the Fixed Partition
of strategies. Gallego and Simchi-Levi (1990) show that
Policies (FP) introduced by Bramel and Simchi-Levi
Direct Shipping policies, i.e., policies in which each vehicle
(1992). A FP strategy partitions the set of retailers into a
visits a single retailer, are within 6% of optimality under
number of regions such that each region is served sepa-
certain restricted parameter settings. Herer and Roundy
rately and independently from all other regions. Moreover,
(1997) and Viswanathan and Mathur (1997) show good
whenever a retailer in a set is visited by a vehicle, all other
empirical performance for the so-called power-of-two
retailers in the set are visited as well. FP policies are easy
strategies under which each retailer is replenished at con-
to implement: they allow for an easy integration of the
distribution, marketing, and customer service functions. stant intervals which are power-of-two multiples of a com-
The main objective of this paper is to characterize the mon base planning period. Power-of-two policies have
asymptotic effectiveness of the class of ZIO and the class been shown (See Federgruen et al. 1992) to be within 2%
of FP policies. Interestingly enough, the cost of solutions of optimality for general submodular joint replenishment
produced by an optimal ZIO policy is directly related to cost structures, but vehicle routing costs may fail to be
the optimal solution of an associated bin-packing problem submodular, as shown in Anily and Federgruen (1990).
in which the retailers need to be packed into unit size bins Finally, as mentioned above, Bramel and Simchi-Levi
and their “sizes” are proportional to their relative demand (1992) analyzed the class of Fixed Partition strategies.
rates. Indeed, we demonstrate that as the number of retail- They show good empirical performance for medium size
ers grows, the cost of an optimal FP policy as well as that problems in the absence of frequency constraints. For a
of an optimal ZIO policy exceeds a lower bound for the more detailed literature review on inventory-routing prob-
minimum cost value under any strategy by no more than a lems see Federgruen and Simchi-Levi (1992a and 1992b).
factor =a: here a denotes the so-called packing constant The remainder of this paper is organized as follows. In
in the associated bin packing problem. The packing con- Section 1 we specify the model assumptions, introduce the
stant a Ä 1 is defined as the asymptotic reciprocal of the notation and provide an example in which ZIO policies
average utilization of a bin in an optimal solution. Since may fail to be optimal (even in an asymptotic sense). In
for any sequence of the retailers demand rates, 1 ¶ a ¶ 2, Section 2 we develop a lower bound B* for the cost under
0.41 5 =2 2 1 represents a worst case bound for the any feasible policy. In Section 3 we construct a close-to-
(asymptotic) optimality gap for FP (ZIO) policies. In par- optimal Fixed Partition Policy and develop worst-case
ticular, when demand rates are generated independently bounds for the gap between its cost and the lower bound
from a common distribution, we have for many distribu- B* derived in Section 2. Section 4 is devoted to probabilis-
tions of the demand rates allowing for so-called “perfect” tic analyses of this optimality gap and of the optimal cost
98 / CHAN, FEDERGRUEN, AND SIMCHI-LEVI
value Z*. A by-product of our analysis is a practical algo- An Example
rithm for the combined inventory routing problem. For Consider an inventory routing problem in which there are
this purpose we describe, in Section 5, an alternative, 3n retailers, each one with demand rate 2, located at the
mathematical programming based heuristic for identifying same point, at a distance d 5 1 from the warehouse. Let
close-to-optimal FP policies which is of superior perfor- f 5 1 and Q 5 3. The fixed cost of sending out a vehicle,
mance. A numerical study reported in Section 6, demon- c, equals 1 and similarly the holding cost rate, h, is 1.
strates that this heuristic comes close to the lower bound
B* even for problems with a moderate number of retailers, Lemma 1. There exists a feasible policy with long-run aver-
n. Finally, in Section 7, we discuss generalizations and age cost Z 5 Z*zi 2 0.5n.
variations of our model.
Proof. Consider policies which satisfy the Zero-Inventory
property. Let w be the size of a single delivery to a retailer
1. NOTATION AND MODEL ASSUMPTIONS in a policy of this type. The frequency constraint implies
that w Ä 2/f 5 2, and hence each delivery to a retailer
Consider a distribution system with a set N 5 {1, 2, . . . ,
must be made by a separate vehicle. Since 2(2di 1 c) D/
n} of geographically dispersed retailers. A central ware-
h 5 12 . 9 5 Q2, the optimal ZIO policy delivers a full
house with an unlimited supply of a given product serves
truck load (three units) to each retailer every 1.5 units of
the retailers using vehicles of limited capacity, Q. Retailer
time. The long-run average transportation cost of this pol-
i, located at a distance di from the warehouse, faces a
icy is (3n)(2d 1 c)/1.5 5 6n while the long-run average
deterministic demand rate Di per unit of time and accrues
holding cost is 3n(1.5) 5 4.5n.
a linear holding cost at a constant rate, h, per unit of
product stored there per unit of time. Demand at each
Consider now a different policy which fails to satisfy the
retailer must be met over an infinite horizon without short-
Zero-Inventory Ordering property. Under this policy, each
ages or backlogging. The frequency with which a given
retailer receives a delivery every unit of time. The fre-
retailer can be visited is bounded from above by f, i.e., the
quency constraint is clearly satisfied. Without loss of gen-
time that elapses between two consecutive deliveries to a
erality, assume the system starts with zero-inventory at
retailer should be at least 1/f. This upper bound on the
each retailer. Partition the retailers into groups of three
delivery frequency to each retailer may be due to the set-up
retailers each. For each such group of three retailers, let
time required for unloading at the retailers or may be due
the delivery sizes be (2, 2, 3) at time 0, (2, 3, 1) at time
to other material handling constraints.
2t 2 1 and (2, 1, 3) at time 2t for each t 5 1, 2, 3, . . . .
Following Anily and Federgruen (1990), we assume that
Hence, for each t 5 1, 2, 3, . . . , only two fully loaded
the demand rates are rational, i.e., for all i 5 1, . . . , n, Di
vehicles are needed to visit each group of three retailers. It
is an integer multiple of some common quantity D, or
is easy to see that the long run average transportation cost
Di 5 kiD with ki a positive integer. We refer to the quan-
of this policy is (2d 1 c)2n 5 6n while the long run
tity ki as the multiplier of retailer i, i [ N. Assuming that
average holding cost is n[1 1 1.5 1 1.5] 5 4n. □
Q and f are rational as well we choose D sufficiently small
that the vehicle capacity Q is an integer multiple of q 5
D/f, the smallest possible delivery quantity for any retailer. 2. A LOWER BOUND FOR THE COST OF ANY
As a consequence, b# 5 Qf/D is integer. FEASIBLE POLICY
Each time a vehicle is sent out to replenish inventory to In this section we develop a lower bound for the minimum
a set of retailers S, it incurs a fixed cost c plus a cost long run average cost of every policy.
proportional to the total distance traveled by the vehicle, Lemma 2.
i.e., a cost proportional to L(S), the length of the optimal

O F k D~2dQ 1 c! 1 G
traveling salesman tour through the warehouse and the n
i i
hk i q
retailers in the set S. Without loss of generality, we set B* 5 ,
i51 2
the cost per mile equal to one. We seek a combined inven-
tory control and routing strategy that procures retailers in is a lower bound for the minimum long-run average cost
time to meet their demands and minimizes the long-run among all feasible policies.
average total inventory holding and transportation cost per Proof. Let Ii Ä 0 be the initial inventory level at retailer i
unit of time. As in traditional joint replenishment inven- for every i. Consider an arbitrary policy 3 over an infinite
tory models, it is not clear that an “optimal” policy always horizon. Let #(3, t) be the average cost per unit of time
exists. So, let Z* denote the infimum of the long-run average incurred by this policy over the interval [0, t). It suffices to
cost values over all feasible policies. Similarly, let Z*zi de- show that #(3, t) Ä (t/(t 1 1/f )) B* 2 c9/t for some
note the infimum of the long-run average cost over all Zero- constant c9 for all t . maxi Ii/kiD.
Inventory Ordering policies. The following example shows
that Z* may be strictly smaller than Z*zi even in an asymp- Assume the retailers are ordered such that d1 Ä d2
totic sense, i.e., in a sequence of problem instances in Ä . . . Ä dn. Let M be the number of vehicles sent out from
which n 3 `, we may have limn3` Z*/n , limn3` Z*zi/n. the warehouse during the interval [0, t); Sj be the set of
 CHAN, FEDERGRUEN, AND SIMCHI-LEVI / 99
retailers visited by vehicle j, j 5 1, 2, . . . , M; and w ji the 3. A CLOSE-TO-OPTIMAL FIXED PARTITION
number of units of product received by retailer i from POLICY
vehicle j during [0, t). Let Qj be the amount of product In this section we construct a FP policy which comes close
delivered by the jth vehicle during the interval [0, t), i.e., to being optimal. In particular, we show that the cost of
Qj 5 ¥i51
n
w ji. this FP policy asymptotically (as n 3 `) exceeds the lower
We first construct a lower bound for the total transpor- bound B* by no more than a factor =a, where 1 Ä a Ä 2
tation cost incurred by policy 3. Consider the jth vehicle denotes the so-called packing constant associated with
and a retailer i [ Sj. Clearly, L(Sj) 1 c Ä 2di 1 c and packing customers of “size” {ki: i 5 1, 2, . . . , n} into bins
hence of size b# 5 Q/q 5 Qf/D (see below as well as the introduc-
Q j @L~S j ! 1 c# 5 O w @L~S ! 1 c# > O
i[S j
j
i j
i [ Sj
w ij ~2d i 1 c!. tion for precise definitions.)
We construct the FP policy using the following two-step
Since Qj ¶ Q, procedure. In the first step, we partition the given area A
where the retailers are distributed into subregions. The

O wQ ~2d
j
i retailers in each such subregion are then partitioned into
L~S j ! 1 c > i 1 c!.
i[S j
sets of retailers by solving the bin-packing problem defined
by the multipliers of the retailers and bins of size b# . Each
Hence the total transportation cost is no smaller than such set is then served in an efficient way.

O @L~S ! 1 c# > O O wQ ~2d

M M j
i
1 c! The Region Partitioning Scheme
j i
j51 j51 i[S j
Let G(u) be an infinite grid of squares with edges parallel
5O O n

i51 j:i[S j
w ij
Q
~2d i 1 c!
to the coordinate axes and side length u/=2. Intersecting
each of these squares with A, let {A1, A2, . . . , Am} denote

>O
k Dt 2 I
n the resulting collection of nonempty intersections. Accord-
i i
~2d i 1 c!. ingly, each subregion Aj, j 5 1, 2, . . . , m, is either a square
Q
of side u/=2 or the intersection of such a square with A.
i51

Consider now the holding cost for each retailer i. Let ri Let Nj, be the set of retailers in subregion Aj with nj 5
be the number of deliveries received by retailer i over the uNju, j 5 1, 2, . . . , m. Given subregion Aj, let d j be the
interval [0, t). Due to the upper bound for the frequency distance from the warehouse to its closest point in Aj, j 5
with which each retailer receives deliveries, ri ¶ (t 1 1/f ) f. 1, 2, . . . , m.
Hence, the holding cost incurred by retailer i is no smaller To construct the fixed partition policy, we group all the
than when the total delivery quantity to retailer i in [0, t) retailers in subregion Aj, j 5 1, 2 . . . , m, into sets by
is the minimum required, i.e., kiDt 2 Ii, and the quantity is solving the bin-packing problem defined by the multipliers
delivered at ri equidistant epoches when inventories are (the numbers ki) of the retailers in Nj and bins of capacity
down to zero (see Carr and Howe 1962 for a rigorous b# . Each such set S of retailers is served together and is
proof). In this case the average inventory level equals visited using a reorder interval that depends on k(S) [
(kiDt 2 Ii)/ 2ri. The total holding costs incurred by retailer ¥i[S ki and the subregion where the retailers in S are
i in [0, t) are thus bounded from below by located. If S is in the subregion Aj for some j 5 1, 2, . . . ,
k i Dt 2 I i hk i t tD k i Dt 2 I i m, then the reorder interval is
th > 2
2r i 2 ~t 1 1/f ! f 2r i tS
2

5
hk i t q k i Dt 2 I i
if Î2k~S! D~2d j 1 c!/h < k~S!q
1
5 2 . ,
2~t 1 1/f ! 2r i f
Let 5 Î 2~2d j 1 c!
k~S! Dh
, if k~S!q , Î2k~S! D~2d j 1 c/h < Q,

O QI ~2d
M
i Q
c9 5 i 1 c!. , otherwise.
j51 k~S! D
Combining the lower bounds on the transportation and the That is, the reorder interval is chosen so that qS 5 k(S) DtS
holding costs, we have is the value of q achieving

O F k D~2dQ 1 c! 1 G H k~S! D~2dq J

t n
i i
hk i q j 1 c! hq
#~3, t! > min 1 . (1)
t 1 1/f 2 k~S!q¶q¶Q 2
i51

c9 h O I 1 i i
Consequently, these reorder intervals satisfy the capacity
2 2 as well as the frequency constraints.
t 2f ~t 1 1/f !

D B* 2 c9 2 h O I
For any set of retailers S, S # Nj, we use the following
5S
t i i 1 routing strategy. The vehicle travels from the warehouse to
. □
t 1 1/f t 2f ~t 1 1/f ! its closest point in Aj, visits the retailers in S in any order,
100 / CHAN, FEDERGRUEN, AND SIMCHI-LEVI
and then returns to the warehouse. It is clear that the total F(b, d) 5 x 1 ab, where x 5 D(2d 1 c)/q and a 5 hq/ 2.
distance traveled is no more than 2d j 1 (uSu 1 1)u. Note that

Analysis of the Upper Bound

For each subregion Aj, let b(Nj) be the optimal solution to
b# q .
b#
b
Î2bD~2dh 1 c! > Îbb# Î2bD~2dh 1 c!
the bin-packing problem defined by the multipliers of the
retailers in Nj, j 5 1, 2, . . . , m. Let Sj(l ), l 5 1, 2, . . . , 5 Î
2b# D~2d 1 c!
h
,
b(Nj) be the lth set of retailers assigned to one bin in this
optimal solution. and F(b# , d) 5 x 1 ab# . Thus
We first need the following technical lemma. F~b, d! x 1 ab
5 f~ x! 5 .
Lemma 3. (a) The function #
F~b , d! x 1 ab#

F~b, d! 5 min
bq¶q¶Q
F bD~2dq 1 c! 1 hq2 G , Observe that f( x) achieves its maximum subject to the
constraint x ¶ ab (implied by case 2) at x 5 ab. Thus,

is concave in b for all b [ [1, b# ]. F~b, d! 2ab 2 2 Îb/b# Îb/b# ,

< 5 5 <
(b) F(b, d) ¶ F(b# , d)=b /b# 5=b /b# [ f(2d 1 c) 1 hQ/ 2] F~b# , d! a~b 1 b# ! 1 1 b# /b Îb/b# 1 Îb# /b
for all b [ [1, b# ].
since ~1 2 =b /b# !2 Ä 0.
Proof. (a) We consider two cases:
Case 3: =2bD(2d 1 c)/h . Q.
Case (i): 2D(2d 1 c)/h Ä Qq. This implies that for any F(b, d) 5 x 1 ab, where x 5 hQ/ 2 and a 5 D(2d 1 c)/Q.
1 ¶ b ¶ b# , 2bD(2d 1 c)/h Ä (bq)2 and hence we have Note that

Î
5
bD~2d 1 c! hQ 2bD~2d 1 c!

F~b, d! 5
Q
1
2
, for
h
. Q 2,
Q 2h
Q< Î
b#
b
Q,
2b# D~2d 1 c!
h
,
i.e. b . b 0 5 ,
2D~2d 1 c! and F(b# , d) 5 x 1 ab# with x ¶ ab (implied by case 3).
Î2bD~2d 1 c!h, otherwise. Thus
Thus, F(b, d) is piecewise concave and since its left deriv-
ative in the breakpoint b 5 b0 equals its right derivative,
F~b, d!
F~b# , d!
5 f~ x! 5
x 1 ab
x 1 ab#
< Î b
b#
,
we conclude that F(b, d) is concave in b.
by the argument used in Case 2. □
Case (ii): 2D(2d 1 c)/h , Qq. Since 2bD(2d 1 c)/h ,
Q2 for any 1 ¶ b ¶ b# , we have
We are now able to derive an upper bound for the cost
F~b, d! of the above defined FP policy and hence for ZFPP, the

5
D~2d 1 c! hbq infimum of the cost values among all FP policies. This
2bD~2d 1 c!
1 , for bound depends on the number of routes b(Nj) into which
q 2 h
each of the customer sets {Nj; j 5 1, 2, . . . , m} in the
, ~bq! 2 ,
5 collection of subregions {Aj: j 5 1, 2, . . . , m} is parti-
2D~2d 1 c! tioned. For each subregion j 5 1, 2, . . . , m, we express
i.e. b . b 0 5 ,
hq 2 the number of routes generated in the subregion relative
Î2bD~2d 1 c!h, otherwise. to the minimum possible number of routes, i.e., the num-
ber of routes required if the demand multipliers {ki: i [
The concavity proof of F(b, d) is analogous to that in the
Nj} allow for perfect packing; in other words, we express
first case.
the number of routes employed by the FP policy in terms of
(b) We consider three cases depending on the quantity
=2bD(2d 1 c)/h, the unconstrained minimizer of the b~N j !
function bD(2d 1 c)/q 1 hq/ 2. bj 5
O b~N j !
l51 k~S j ~l !!/b#
> 1.

Case 1: bq ¶ =2bD(2d 1 c)/h ¶ Q.

Î
Theorem 1.
Î2bD~2d 1 c!h 5 b Î2b# D~2d 1 c!h
O Îb O F k D~2dQ 1 c! 1 G
F~b, d! 5
b# m
i i
hk i q
Z FPP < 1 2nuf.
Î
j
b 2
< F~b# , d!, j51 i[N j
b#
Proof. We bound ZFPP by the cost value of the above
since b ¶ b# and =2b# D~2d 1 c!h represents the uncon-
described FP policy. Under this policy, the reorder interval
strained minimum of the function b# D(2d 1 c)/q 1 hq/ 2.
for every subset of retailers Sj(l ), l 5 1, 2, . . . , b(Nj), is
Case 2: bq . =2bD(2d 1 c)/h. tSj(l ) Ä 1/f. Hence, ZFPP is bounded by
 CHAN, FEDERGRUEN, AND SIMCHI-LEVI / 101

OO
m b~N j ! By comparing the upper bound in Theorem 1 and the
Z FPP < $@2d j 1 c 1 u~uS j ~l !u 1 1!#/t S j ~l ! lower bound in Section 2, we immediately obtain the fol-
j51 l51
lowing asymptotic worst-case bound for the optimality gap
1 hk~S j ~l !! Dt S j ~l ! / 2} of ZFPP and hence for Z*zi.

< OO
m b~N j !

j51 l51
H 2d j 1 c
t S j ~l !
1
hk~S j ~l !! Dt S j ~l !
2
J 1 2nuf
Theorem 2. Consider an arbitrary sequence of retailer loca-
tions { x1, x2, . . .} and associated retailer multipliers {k1,

H J
k2, . . .}. Let Z*(n), Z*zi(n) and ZFPP(n) denote the infi-
k~S j ~l !! D~2d j 1 c!
OO
m b~N j !
hq mum of the costs incurred to serve the first n retailers
5 min 1
j51 l51 k~S j ~l !!q¶q¶Q q 2 among all possible strategies, all zero-inventory strategies
and all FP policies, respectively. Then
1 2nuf
Z *zi ~n! Z FPP~n!
OO
m b~N j !
5 F~k~S j ~l !!, d j ! 1 2nuf. lim < lim < Î2 5 1.41.
n3` Z*~n! n3` Z*~n!
j51 l51
Proof. Consider the FP policy obtained by the above two-
By Lemma 3 (a), F(b, dj) is a concave function of b for
step procedure with a given grid size u. Let J1 5 { j:
every j 5 1, 2, . . . , m, and therefore we have for every j,
limn3` uNju 5 `} and J0 5 {1, 2, . . . , m}\J1 5 { j: limn3`

O
b~N j !
F~k~S j ~l !!, d j ! < b~N j ! F SO b~N j ! k~S j ~l !!
, dj D uNju , `} so that

O Îb O F k D~2dQ 1 c! 1 G
l51 l51 b~N j ! hk i q
i i

S bb# , d D .
Z 0 ~n! ; j ,
5 b~N j ! F j[J 0 i[N j 2
j
j
is bounded in n. Observe from Lemma 2 that Z* ¶ n
Hence, we have by Lemma 3 (b),
[Dc/Q 1 hq/ 2]. Thus,

F S b#
, dj D Z FPP~n!
< lim
Z 0 ~n!
O b~N !~b# / b !
m bj lim
n3` Z*~n! n3` n@Dc/Q 1 hq/ 2#
Z FPP < 1 2nuf
O O
j j
j51 b# / b j Îb j @k i D~2d i 1 c!/Q 1 hk i q/ 2#

S D
j[J 1 i[N j
b#
F , dj
1 lim
n3` O O m
@k i D~2d i 1 c!/Q 1 hk i q/ 2#
OO
m b~N j ! bj j51 i[N j
5 k~S j ~l !! 1 2nuf 2uf
j51 l51 b# / b j 1
Dc/Q 1 hq/ 2
F~b# , d j !
OO
m
k i Îb j Îb j % 1 2uf
< 1 2nuf < lim $max (2)
Dc/Q 1 hq/ 2
b#
n3` 1
j51 i[N j j[J

F~b# , d i ! 2uf
O O k Îb < Î2 1
m ,
< i j 1 2nuf Dc/Q 1 hq/ 2
j51 i[N j b#
since bj ¶ 2 (the latter follows from the well-known fact
O OF G
m k i D~2d i 1 c! hk i q
5 Îb j 1 1 2nuf. □
that the optimal solution to the bin-packing problem is no
j51 i[N j Q 2 more than twice the sum of the fraction of the bin capacity
taken by each item). Since ZFPP denotes the infimum over
We now relate the upper bound for ZFPP to the lower all fixed partition policies, the theorem follows by consid-
bound for Z* obtained in Section 2. Note that, in case all ering a sequence of FP policies generated by the two-step
bj 5 1, j 5 1, 2, . . . , m, the two bounds coincide except procedure corresponding to a sequence of grid sizes {ul}
for the term 2nuf which can be made arbitrarily small by with liml3` ul 5 0. □
employing a small enough grid size u. The upper bound in
Theorem 2 also exhibits the fundamental tradeoff to be
considered in implementing the above regional partition- 4. PROBABILISTIC ANALYSIS OF OPTIMALITY
ing scheme: the second term in the upper bound (2nuf ) GAPS AND OPTIMAL COST VALUES
can be reduced to zero by adopting an increasingly small Significantly sharper bounds for the asymptotic optimality
grid size u in the first step of the procedure and hence by gap may be obtained if the sequence of retailer locations
dividing the total retailer population into an increasingly and sizes (multipliers) can be assumed to arise from a
large set of subregions {Aj : j 5 1, 2, . . . , m}. On the specific probabilistic pattern. A basic probabilistic model
other hand, the larger the subregions {Aj: j 5 1, 2, . . . , assumes that the sequences of retailer locations { x1,
m} are, the more efficient solutions can be obtained for x2, . . .} and retailer multipliers {k1, k2, . . .} are both inde-
the bin packing problem to be solved in the second step pendent and identically distributed, and independent of
of the procedure, thus reducing the values of {bj}. each other.
102 / CHAN, FEDERGRUEN, AND SIMCHI-LEVI
In this probabilistic model we have for all subregions j Rhee completely characterizes the class of distributions
with limn3` uNju 5 `, i.e., for all j [ J1, that the bin- which allow for perfect packing, see also Coffman and
packing problem to be solved in the second of the two-step Lueker.
construction procedure for the FP policy, deals with a se- More generally, in many distribution problems we en-
quence of retailer sizes {ki} which is i.i.d. and with the counter significant correlations between the retailer loca-
same common multiplier distribution in all subregions. It is tions and their sales volumes, e.g., retailers in urban areas
well known from the theory of subadditive processes by or in specific states or regions may have larger demands.
Kingman (1976) (see also Rhee and Talagrand (1987)) We may thus wish to generalize the above probabilistic
that for all j [ J1, model to one in which the sequence of pairs {( xi, ki)} is
i.i.d with a common joint distribution, characterized by the
lim b~N j !/uN j u 5 g (a.s.),
n3` marginal distribution of the locations m[ and the condi-
while by the strong law of large numbers, tional multiplier distributions (kux 5 x0). Let g( x) denote
the asymptotic average number of bins (routes) required
lim
n3`
1
uN j u
O kb# 5 E~k!
i[N j
i

b#
(a.s.).
for retailers with multipliers independently distributed as
(kux) and let a( x) 5 g( x)b# /E(kux). Following, once again,
the proof of Theorem 2 we obtain:
We conclude that for all j 5 J1,
Corollary 2. Consider a sequence of pairs of retailer loca-
lim b j 5 a ; g b# /E~k!, (3) tions and multipliers {( x1, k1,); ( x2, k2,); . . .} which are
n3`
independent and identically distributed with a common
and obtain the following corollary from Theorem 2.
joint distribution, characterized by m[, the marginal cdf of
Corollary 1. Consider a sequence of retailer locations { x1, the retailer locations, and the conditional distributions
x2, . . .} and retailer multipliers of {k1, k2, . . .} which are (kux 5 x0). Let Z*(n), Z*zi(n), ZFPP(n) be defined as in
both i.i.d and independent of each other. Let Z*(n), Z*zi(n) Theorem 2. Then, almost surely
and ZFPP(n) be defined as in Theorem 2. Then Z *zi ~n! Z FPP~n!

~a! lim
Z *zi ~n!
< lim
Z ~n! FPP
< Îa (a.s.).
~a! lim
n3` Z*~n!
< lim
n3` Z*~n!
< # Îa ~ x! d m ~ x!.
n3` Z*~n! n3` Z*~n!
D@2E~kd! 1 cE~k!# hE~k!q Z*~n!
~b! 1 < lim
DE~k!@2E~d! 1 c# hE~k!q Z*~n! Q 2 n3` n
~b! 1 < lim

# Îa ~ x! d m ~ x! F G
Q 2 n3` n
D@2E~kd! 1 cE~k!# hE~k!q
< Îa F DE~k!@2E~d! 1 c# hE~k!q
Q
1
2
G (a.s.).
<
Q
1
2
.

(c) If all conditional multiplier distributions (kux) allow

(c) If a 5 1, i.e., if the distribution of retailer multipli- for perfect packing:
ers allows for perfect packing:
Z *zi ~n! Z FPP~n!
Z *zi ~n! Z FPP~n! lim 5 lim 51 (a.s.), and
lim 5 lim 51 (a.s.), n3` Z*~n! n3` Z*~n!
n3` Z*~n! n3` Z*~n!
Z*~n! D@2E~kd! 1 cE~k!# hE~k!q
i.e., Fixed Partition policies are (a.s.) asymptotically opti- lim 5 1 (a.s.).
n3` n Q 2
mal, and
We observe that the above probabilistic analysis is based
Z*~n! DE~k!@2E~d! 1 c# hE~k!q on the sequence of retailer multipliers satisfying two ele-
lim 5 1 .
n3` n Q 2 mentary limit results:
Proof. Parts (a) and (c) are immediate from Lemma 2,
(2), and (3); the existence of limn3` Z*(n)/n follows from (i) in each subregion j, the sequence of multipliers sat-
Z*(n1 1 n2) ¶ Z*(n1) 1 Z*(n2) and the fact that the isfies the strong law of large numbers;
minimum cost to cover a group of retailers { xn11, . . . , (ii) in each subregion, the bin-packing problem associ-
xn1n} is identically distributed for all n Ä 1, see Kingman ated with the sequence of multipliers, has an asymptotic
(1973). The upper and lower bound for limn3` Z*(n)/n almost sure average value, i.e., for all j [ J1, limn3`
follow again from Lemma 2, (2), and (3). □ b(Nj)/uNju 5 gj(a.s.) for some gj.

Remarks. Part (c) of the corollary deals with the case Both limit results can be established under conditions far
where the distribution of retailer multipliers allows for more general than those of Corollaries 1 and 2. Both limit
perfect packing, i.e., the wasted space in the bins is asymp- results apply, e.g., when the sequence of retailer attributes
totically insignificant. Karmarkar (1982) first proved that { xi, ki)} is stationary, i.e., the joint distribution of any
any nonincreasing probability density function (with some m-tuple {( xn, kn); ( xn11, kn11), . . . , ( xn1m, kn1m)} is in-
mild regularity conditions) allows for perfect packing. dependent of n; see Kingman (1976).
 CHAN, FEDERGRUEN, AND SIMCHI-LEVI / 103
The above probabilistic models are suitable to charac- The CCLP can be formulated as the following integer
terize the relative performance of FP and ZIO policies for linear program. Let
sequences of progressively expanding retailer chains, with
each additional retailer acquiring an incremental clientele. yj 5 $ 1,0, if a concentrator is located at site j,
otherwise,
A different model is needed when a fixed customer market
is covered by a progressively larger and denser retailer and let
chain, e.g., one with a given customer demand rate density
w[ in the plane. If n retailers with i.i.d. locations offer
x ij 5 $ 1,0, if terminal i is connected to concentrator j,
otherwise.
identical service and merchandise, it is reasonable to as-
OOc Ovy
n m m

sume that any particular retailer attracts the customers in Problem P: Min ij x ij 1 j j
i51 j51 j51
the region of those locations for which this retailer is the
Ox
m
closest among all of the n available retailers. Alternatively, s.t. 51 ; i, (4)
ij
in an expanding market with n i.i.d. customer locations j51

Owx
{ y1, . . . , yn} and associated i.i.d. demand rates each re- n
tailer i attracts the subset of customers to which it is clos- i ij <C ; j, (5)
i51
est. The problem in analyzing this model is that the retailer
size (multiplier) ki of any given retailer i (1 ¶ i ¶ n) has a x ij < y j ; i, j, (6)
distribution which varies as n is increased, i.e., the retailer x ij [ $0, 1% ; i, j, (7)
sizes need to be described by a tableau of random variables
y j [ $0, 1% ; j. (8)
{kin; 1 ¶ i ¶ n} rather than a single sequence. The asymp-
totic behavior of the packing constants {bj} appears to be Constraints (4) ensure that each terminal is connected
unknown under this type of probabilistic model. to exactly one concentrator, and constraints (5) ensure that
the concentrator’s capacity constraint is not violated. Con-
straints (6) guarantee that if a terminal is connected to site
5. AN EFFICIENT ALGORITHM FOR INVENTORY
j, then a concentrator is located at that site. Constraints
ROUTING MODELS
(7) and (8) ensure the integrality of the variables.
The effectiveness of the Fixed Partition Policies suggests a
new algorithm for general inventory-routing problems sim- 5.2. Formulation
ilar to the one developed by Bramel and Simchi-Levi for To formulate the region partitioning part of the inventory-
the Capacitated Vehicle Routing Problem with Unsplit routing problem as an instance of the Capacitated Concen-
Demands. The algorithm is based on formulating the trator Location Problem, we refer to each retailer as a
inventory-routing model as a Capacitated Concentrator terminal whose weight is ki, i.e., we set wi 5 ki. Each
Location Problem (CCLP), for the purpose of generating a retailer is also a possible site for a concentrator with ca-
partition of regions. Each of these regions is assigned a vehi- pacity b# , i.e., C 5 b# . Thus, in our formulation of the
cle which visits all retailers in the region at equidistant inventory-routing problem as a CCLP, m 5 n. The set-up
epoches. The CCLP is subsequently solved, and its solu- cost for installing a concentrator at site j, (where site j
tion provides a policy whose cost is, asymptotically, no corresponds to retailer j) is 2dj. Similar to Bramel and
larger than ZFPP. Simchi-Levi, we have used two possible connection costs,
cij:
5.1. The Capacitated Concentrator Location
Problem direct cost: c ij 5 2d ij ,
The Capacitated Concentrator Location Problem (CCLP) nearest insertion cost: c ij 5 d i 1 d ij 2 d j .
can be described as follows: given m possible sites for concen- The solution to the CCLP provides the grouping of the
trators of fixed capacity C, we would like to locate con- retailers into subsets. Each such subset is served together.
centrators at a subset of these m sites and connect n We have implemented two versions of the algorithm
terminals, where terminal i uses wi units of a concentra- corresponding to this pair of connection cost specifications
tor’s capacity, in such a way that each terminal is con- which we refer respectively as the Star-Connection Heuris-
nected to exactly one concentrator, the concentrator tic (ST) and the Nearest Insertion Heuristic (NI).
capacity is not exceeded and the total cost is minimized. A The next Theorem shows that asymptotically, the cost of
site-dependent cost is incurred for locating each concen- the solution produced by the ST heuristic approaches the
trator; that is, if a concentrator is located at site j, the value of ZFPP.
set-up cost is vj, for j 5 1, 2, . . . , m. The cost of connect-
ing terminal i to concentrator j is cij (the connection cost), Theorem 3. Under the assumptions of Corollary 1 and for
for i 5 1, 2, . . . , n and j 5 1, 2, . . . , m. No assumptions any distribution F of the retailer multipliers which allows
need to be made on the costs {cij} and {vj}. We assume for perfect packing, we have
that there is enough capacity so that a feasible solution
lim Z ST/Z FPP 5 1(a.s.).
exists. n3`
104 / CHAN, FEDERGRUEN, AND SIMCHI-LEVI
Table I same combination of retailer and warehouse locations and
List of Parameter Values for Each Instance the same retailer multipliers and value of b# . In each set, the
Parameter
retailer and warehouse locations are independently and
Set Category q Q c h D uniformly located in a square of size [100, 100]. Across
[1] I 5 74 2 6 10 each set, the number of retailers varies from 30 to 200. In
[2] I 5 74 10 6 10 all cases b# 5 14, and the retailer multipliers are uniformly
[3] I 5 74 100 6 10 distributed on the integers {1, 2, . . . , 14}. For every set,
[4] II 5 74 2 1 10 we have generated 10 different problem instances differing
[5] II 5 74 2 10 10
[6] II 5 74 2 100 10 in the values of the parameters q, Q, c, h, and D only. The
[7] III 5 74 2 6 5 characteristics of each problem are reported in Table I.
[8] III 5 74 2 6 20 The instances in each set can be subdivided into three
[9] III 5 74 2 6 50 categories. In the first category (instances 1–3), we investi-
[10] III 5 74 2 6 500
gate the impact of an increase in c, the fixed set-up cost. In
the second category (instances 4 – 6), we investigate the
Proof. We omit the details of the proof since it is similar impact of increasing the holding cost h. In the third cate-
to the proof of the upper bound from Section 3. See Chan gory (instances 7–10), we investigate the impact of increas-
(1995) for details. □ ing the demand rate D.
Table II reports the ratios between the heuristic cost
6. A NUMERICAL STUDY and the lower bound on the cost of every policy (i.e.,
In this section we report our computational experience ZH/B*).
with the Location Based Heuristic using randomly gener- We observe that the algorithm produces solutions rela-
ated problems. Clearly, computing the optimal cost for tively close to the lower bound; the Optimality gap with
even small size problems is intractable. We therefore re- respect to the lower bound B* is always less than 16% and
port the heuristic’s performance relative to the two lower in most cases no more than 10%. The results also show
bounds developed in the previous sections: B*zi and B*. that increasing the fixed set-up cost, c, tends to improve
We have generated 80 problem instances, partitioned the performance of the algorithm; in category I the rela-
into eight different sets. All the instances in a set share the tive error decreases as c increases. A similar behavior is

Table II
Heuristic Cost Over Lower Bound on the Cost of Every Policy
Parameter
Set 30 50 80 100 120 150 180 200
[1] 1.140 1.111 1.110 1.100 1.096 1.094 1.086 1.090
[2] 1.137 1.098 1.109 1.099 1.096 1.094 1.085 1.089
[3] 1.103 1.089 1.086 1.079 1.079 1.076 1.067 1.070
[4] 1.153 1.111 1.115 1.120 1.105 1.100 1.090 1.093
[5] 1.112 1.089 1.088 1.080 1.078 1.075 1.070 1.072
[6] 1.017 1.013 1.013 1.012 1.011 1.011 1.010 1.010
[7] 1.100 1.080 1.078 1.071 1.069 1.067 1.062 1.064
[8] 1.150 1.112 1.118 1.114 1.106 1.103 1.092 1.096
[9] 1.153 1.112 1.116 1.119 1.105 1.100 1.091 1.093
[10] 1.154 1.104 1.111 1.122 1.101 1.096 1.088 1.089

Table III
Heuristic Cost Over Lower Bound on the Cost of Every Zero-Inventory Ordering Poicy
Parameter
Set 30 50 80 100 120 150 180 200
[1] 1.113 1.085 1.085 1.075 1.071 1.070 1.061 1.065
[2] 1.110 1.084 1.083 1.074 1.070 1.069 1.060 1.064
[3] 1.081 1.068 1.064 1.057 1.057 1.054 1.045 1.049
[4] 1.143 1.101 1.105 1.110 1.095 1.090 1.081 1.083
[5] 1.090 1.068 1.067 1.059 1.057 1.055 1.049 1.051
[6] 1.014 1.010 1.010 1.009 1.008 1.008 1.007 1.008
[7] 1.080 1.060 1.060 1.052 1.050 1.049 1.043 1.045
[8] 1.129 1.096 1.097 1.093 1.085 1.082 1.072 1.075
[9] 1.140 1.101 1.104 1.108 1.094 1.089 1.080 1.082
[10] 1.153 1.103 1.109 1.120 1.100 1.094 1.086 1.088
 CHAN, FEDERGRUEN, AND SIMCHI-LEVI / 105
Table IV
Computational Result for Nonperfect Packing Instances
30 50 80 100 120 150 180 200
ZH/B*zi 1.151 1.126 1.115 1.107 1.100 1.091 1.094 1.091
ZH/B* 1.188 1.161 1.149 1.142 1.134 1.125 1.128 1.125

observed in category II; increasing the holding cost h tends Sections 2– 4 is the assumption that the common quantity
to decrease the relative error. D used to measure the retailer demand rates {di 5 kiD:
We have also compared the lower bound B* which ap- i 5 1, . . . , n} is chosen small enough that Q, the vehicle
plies to all strategies and the bound B*zi which applies to capacity, is an integer multiple of D/f. Such a choice con-
the ZIO policies only. B*zi is extremely close to B*; the siderably simplifies the analysis. However, tighter bounds
ratio between these two values is no more than Q/b# q. In and tighter characterizations of optimality gaps may be
our problem sets B*zi/B* is no more than 1.025. Table III obtained under a larger (and often more natural) common
reports the ratios 2H/B*zi for all problem sets. quantity D; see Chan.
The next set of experiments is designed to estimate the We now extend our model to include discounted costs.
effect of multipliers distributions that do not allow for per- For this purpose observe that the model introduced in the
fect packing on the performance of the location based introduction, considers the average cost criterion, capturing
heuristic. For this purpose, Table IV reports the results of the capital costs associated with system-wide inventories as
the eight instances when b# 5 10 and the retailer multipli- part of the holding costs. An alternative model, perhaps
ers are uniformly distributed on the integers {3, 4, 5, 6, 7, more directly reflecting the company’s cash flows, consid-
8}. The values of the parameters are q 5 5, Q 5 54, c 5 2, ers the total, continuously discounted, value of all out-of-
h 5 6 and D 5 10. pocket expenses (i.e., the routing costs and inventory
Finally, we investigate the impact of the frequency con- carrying charges beyond the cost of capital, if any).
straints on our algorithm. Since the frequency constraint The model with the discounted cost criterion is signifi-
plays a role only through q, we have changed this parame- cantly more complex to analyze. This applies even for the
ter, and hence the value of b# in the last set of parameters. simple EOQ model which corresponds with the special
The next two tables reports our computational experience
case of a single retailer (n 5 1), and even when the con-
with this set of problems. Observe, that the results are
straints on vehicle capacities and delivery frequencies are
similar to the previous ones; the error decreases as the
ignored. For this basic, discounted EOQ-model, it is possi-
number of retailer increases and for problems with at least
ble to show, see, e.g., Jesse et al. (1983), Porteus (1985), or
80 retailers the optimality gap between the solution pro-
Lee and Nahmias (1993), that analogous to the classical
duced by our algorithm and the lower bound B* is no
EOQ-model with the average cost criterion, a stationary
more than 19% and the gap with respect to B*zi is no more
ZIO policy is optimal with a constant delivery quantity q.
than 15%.
The optimal value of q is the (unique) cost of a nonlinear
equation, which cannot be obtained in closed form. How-
7. FURTHER RESULTS AND ALTERNATIVE ever, by ignoring third and higher degree terms in the
MODELS Taylor series expansion of the nonlinear components of
The analysis performed in this paper can be carried over the cost expression, we obtain an approximation for the
to more general versions of our model. For instance, a optimal value of q which is identical to the well-known
somewhat restrictive assumption in the model analyzed in (Harris-Wilson) EOQ-formula in the average cost case,

Table V
ZH/B*zi for Different Values of q
b# 30 50 80 100 120 150 180 200
8 1.104 1.073 1.078 1.070 1.063 1.062 1.066 1.067
10 1.151 1.126 1.115 1.107 1.100 1.091 1.094 1.091
12 1.140 1.132 1.119 1.116 1.117 1.108 1.099 1.096

Table VI
ZH/B* for Different Values of q
b# 30 50 80 100 120 150 180 200
8 1.142 1.108 1.113 1.105 1.098 1.096 1.101 1.102
10 1.188 1.161 1.149 1.142 1.134 1.125 1.128 1.125
12 1.174 1.165 1.151 1.149 1.150 1.141 1.131 1.129
106 / CHAN, FEDERGRUEN, AND SIMCHI-LEVI
with the capital cost component in the unit carrying cost FEDERGRUEN, A., AND Y. S. ZHENG. 1992. The Joint Replen-
rate h, replaced by the product of the (continuous) dis- ishment Problem with General Joint Cost Structures:
count factor and the item’s unit dollar value. Moreover, it General Solution Methods and Performance Bounds.
has been substantiated that this approximation comes very Opns. Res. 40, 384 – 404.
close to being optimal. FEDERGRUEN, A., AND D. SIMCHI-LEVI. 1992a. Routing and
To our knowledge, all continuous time models for multi- Inventory Control Models for General Distribution Net-
item or multi-location systems have confined themselves to works. Working Paper, Columbia University.
the average cost case, and it has not been possible to adapt FEDERGRUEN, A., AND D. SIMCHI-LEVI. (1995), Analytical
to the discounted cost model, any of the numerous charac- Analysis of Vehicle Routing and Inventory Routing
Problems. Handbooks in Operations Research and Man-
terizations of optimality and accuracy gaps for heuristics
agement Science, Network Routing. M. Ball, T. Magnanti,
and bounds, respectively. This applies even for the special
C. Monma and G. Nemhauser, Eds., North-Holland,
case of our model in which the cost of a vehicle route does
Amsterdam, 297–373.
not depend on the number of miles driven, i.e., where all
GALLEGO, G., AND D. SIMCHI-LEVI. 1990. On the Effectiveness
vehicle routes have an identical cost value, c. On the other of Direct Shipping Strategy for the One Warehouse
hand, the analysis in this paper suggests a natural strategy Multi-Retailer R-Systems. Mgmt. Sci. 36, 240 –243.
for the discounted cost model: compute a (close-to) opti- HALL, R. W. 1991. Comments on “One-Warehouse Multiple
mal FPP for the average cost model (e.g., via the methods Retailer Systems with Vehicle Routing Costs.” Mgmt. Sci.
described in Section 5): then, for each of the routes gener- 37, 1496 –1497.
ated by the FPP, determine the (constant) delivery quan- HERER, Y., AND R. ROUNDY. 1997. Heuristics for a One-
tity q as the value achieving the minimum in (1) with the Warehouse Multiretailer Distribution Problem with Per-
average cost expression replaced by its discounted cost formance Bounds. Opns. Res. 45, 102–115.
analogue. JACKSON, P. L., W. L. MAXWELL, AND J. A. MUCKSTADT. 1985.
The Joint Replenishment Problem with Powers of Two
Restriction. AIIE Trans. 17, 25–32.
ACKNOWLEDGMENT JESSE, R. R., A. MITRA, AND J. F. COX. 1983. EOQ formula: Is
The research was supported in part by ONR Contract it Valid Under Inflationary Conditions. Decision Sci. 14,
N00014-90-J-1649, N00014-95-1-0232, NSF Contracts 370 –374.
KARMARKAR, N. 1982. Probabilistic Analysis of Some Bin-
DDM-8922712 and DDM-9322828.
Packing Algorithms. Proc. 23rd Annual Symposium.
Foundations Comp. Sci. 107–111.
REFERENCES KINGMAN, J. F. C. 1976. Subadditive Processes. Lecture Notes
in Math. 539, Springer-Verlag, Berlin, 168 –222.
ANILY, S., AND A. FEDERGRUEN. 1990. One Warehouse Multi- LEE, H. L., AND S. NAHMIAS. 1993. Single Product, Single
ple Retailer Systems with Vehicle Routing Costs. Mgmt. Location Models. Handbooks in Operations Research and
Sci. 36, 92–114. Management Science, the volume on Logistics of Produc-
ANILY, S., AND A. FEDERGRUEN. 1991. Rejoinder to “Com- tion and Inventory, S. C. Graves, A. H. G. Rinnooy Kan
ments on One Warehouse Multiple Retailer Systems with and P. H. Zipkin (Eds.), 3–55.
Vehicle Routing Costs.” Mgmt. Sci. 37, 1497–1499.
PORTEUS, E. L. 1985. Undiscounted Approximations of Dis-
BRAMEL, J., AND D. SIMCHI-LEVI. 1995. A Location Based
counted Regenerative Models. O. R. Lett. 3, 293–300.
Heuristic for General Routing Problems. Opns. Res. 43,
RHEE, W. T., AND M. TALAGRAND. 1987. Martingale Inequali-
649 – 660.
ties and NP-Complete Problems. Math. O. R. 12,
CARR, C. R., AND C. W. HOWE. 1962. Optimal Service Policies
and Finite Time Horizons. Mgmt. Sci. 9, 126 –140. 177–181.
CHAN, L. M. A. 1995. Probabilistic Analysis of Inventory- RHEE, W. T. 1988. Optimal Bin Packing with Items of Ran-
Routing Models. Ph.D. Thesis, Columbia University. dom Sizes. Math. O. R. 13, 140 –151.
COFFMAN, E. G. JR., AND G. S. LUEKER. 1991. Probabilistic STALK, G., P. EVANS, AND L. E. SHULMAN. 1992. Competing on
Analysis of Packing and Partitioning Algorithms. John Capabilities: The New Rule of Corporate Strategy. Har-
Wiley & Sons Ltd., New York. vard Business Review, 57– 69.
FEDERGRUEN, A., M. QUEYRANNE, AND Y. S. ZHENG. 1992. VISWANATHAN, S., AND K. MATHUR. 1997. Integrating Routing
Simple Power of Two Policies Are Close to Optimal in a and Inventory Decisions in One Warehouse Multiretailer
General Class of Production/Distribution Networks with Multiproduct Distribution Systems. Mgmt. Sci. 43, 294 –
General Setup Costs. Math. O. R. 17, 951–963. 312.