European Journal of Operational Research 129 (2001) 242±251

www.elsevier.com/locate/dsw

Theory and Methodology

A combinatorial approach to level of repair analysis

Lilian Barros *, Michael Riley
Business Engineering Group, Department of Civil and Environmental Engineering, University of Southampton, Southampton, UK
Received 6 March 2000; accepted 14 May 2000

Abstract

This paper presents an approach to optimise level of repair decisions taking into account submodular properties of
standard life cycle cost functions, which include ®xed and variable costs. It proposes an integer programming for-
mulation to solve level of repair problems for multi-echelon multi-indenture level systems. The method converges
quickly to the optimum solution relying on heuristics to obtain tight bounds for a subsequent branch-and-bound
procedure. A software package called level of repair optimisation model (LOROM) was developed to implement the
branch-and-bound method that does not rely on linear programming relaxations. This approach is rather generic and
can be applied to a wide class of problems with convex total cost functions such as plant location problems or
transportation problems with ®xed costs. Ó 2001 Elsevier Science B.V. All rights reserved.

Keywords: Combinatorial optimisation; Integer programming; Logistics; Submodular functions; Repair policy

1. Introduction analysis (LORA) that outperforms well-known

USA commercial packages whenever ®xed costs
This paper provides an optimisation tool to are signi®cant. In general, practical industrial and
minimise maintenance costs over the life cycle of military LORA problems can be very large and the
sophisticated hardware. Prospective users include solution of integer programming formulations can
engineering hardware designers, defence contrac- be time-consuming and/or exceed present compu-
tors, producers and buyers of expensive and long- tational capability.
lived hardware. The need for such a tool can be Therefore, an alternative branch-and-bound
appreciable given the large number of commercial method is proposed that does not rely on linear
life cycle cost models that do not use optimisation programming relaxations. The method includes a
methods. Barros (1998) proposes a pure integer very ecient heuristics and has three main ad-
programming formulation for level of repair vantages: (i) it can obtain tight upper and a lower
bounds for the life cycle cost estimates, (ii) the
*
Corresponding author. P.O. Box 212, Southampton SO14
greedy heuristic algorithm is generally able to cut
0HG, UK. Tel.: +44-23-80-512-654. many branches of a basic decision tree and thus
E-mail address: [email protected] (L. Barros). minimise the enumeration search procedure, and

0377-2217/01/$ - see front matter Ó 2001 Elsevier Science B.V. All rights reserved.
PII: S 0 3 7 7 - 2 2 1 7 ( 0 0 ) 0 0 2 2 1 - 6
 L. Barros, M. Riley / European Journal of Operational Research 129 (2001) 242±251 243

(iii) the algorithm only requires arithmetic sums The LORA problem can then be formulated as
and subtractions, thus being computer-time e- follows:
cient. XX XX
The problem is formulated as a multiple loca- ‰P1Š MINIMISE vcrx nrx fcri mri
tion problem with ®xed and variable costs, where r2R x2X r2R1 i2I

the objective is to make decisions about whether

and where to repair each hardware component as to subject to:
minimise total life-cycle maintenance costs. These X
decisions are subject to two sets of constraints, the nrx ˆ 1 for x 2 X ; …1†
r2R
®rst one specifying that a single repair option must
be selected for each component and the other a set mri P nrx for r 2 R; i 2 I; x 2 X ; …2†
of ``bundling'' constraints, which are the most nrt 6 nry for r 2 R2  R; 8y 2 Ct ;
innovative aspect of the formulation. These bun- 8t non-terminal; …3†
dling constraints ensure that repair decisions are
nrt P nry for r 2 R3  R;
coherent across indenture levels. For instance, if a
computer is discarded so are all enclosed cards 8y 2 Ct ; 8t non-terminal; …4†
as well as all chips enclosed in the cards. mri ; nrx 2 f0; 1g for r 2 R; i 2 I; x 2 X : …5†
Barros (1998) proposed the following formula-
tion for the LORA problem. Let x 2 X de®ne the set Constraint set (1) ensures that a single repair
of components in a system to be repaired, using option is selected for each system component.
r 2 R repair facilities as well as non-repair or dis- Constraint set (2) ensures that repair facilities are
card option. The system is broken down into i 2 I opened before system components can be assigned
indenture levels (there are modules enclosed into to be repaired at that site. Constraint sets (3) and
subsystems and piece parts enclosed into modules (4) are bundling constraints that ensure compati-
for a two-indenture level system). A subset of repair bility in the choice of repair options for father and
options R1  R requires specialised facilities, which son components in a system (for instance, if the
incur ®xed costs if they are opened. For systems father is discarded, so is the son etc.). These bun-
with more than one indenture level, the notation dling constraints are discussed in greater detail in
y 2 Ct means that component y is enclosed in Section 3 below. Constraint set (5) de®nes a pure
component t (y is the ``son'' of component t or t is integer problem.
the ``father'' of y). Two other subsets of repair op- Variable costs are de®ned for each component
tions R2  R and R3  R; R2 6ˆ R3 require bundling of the system while ®xed costs vary only by in-
constraints to impose compatibility between father denture level. Fixed costs may represent one-time
and son components. For example, if a father costs to open a repair facility such as large initial
component is discarded upon failure, so are all en- expenditures with support and test equipment and
closed son components. Life cycle costs are classi- other costs, which are not a€ected by the number
®ed in average and ®xed costs, where average costs of repairs taking place (assuming that facility
increase with failure rates (such as labour costs) and capacity is not exceeded). Positive ®xed costs are
®xed costs do not (such as support and test equip- an elementary concept in microeconomic theory
ment). Average variable costs are represented as (see Varian, 1984), which are responsible for
vcrx and ®xed costs as fcri . Two sets of Boolean diminishing returns to scale, a microeconomic
decision variables are de®ned below: concept that corresponds mathematically to the
 objective function in [P1] being submodular.
1 if component x is repaired in facility r; Submodularity shares practical and mathe-
nrx ˆ
0 otherwise; matical properties with convex functions, i.e. they
 have wide applications in ®elds such as economics
1 if facility r is open at indenture level i; and engineering, there are theoretical and practical
mri ˆ
0 otherwise: ways to minimise such functions and convexity is
244 L. Barros, M. Riley / European Journal of Operational Research 129 (2001) 242±251

preserved after many types of operations and above, a positive real-value marginal cost function
transformations (Lovasz, 1983). The maximisation …MC…x†† can be generically de®ned as follows:
of submodular functions (equivalent to the mini-
misation of a supermodular function) is NP-hard ‰MCŠ MC…x† ˆ c0 ‡ c1 x ‡ c2 x2 for x  X ‡ :
(Kubiak, 1995) but the minimisation of such Or more speci®cally, using [TC],
functions is much more amenable to mathematical
treatment, in particular where greedy algorithms MC…x† ˆ b1 ‡ 2b2 x ‡ 3b3 x2 for x  X ‡ :
can be used. Goldengorin et al. (1999) proposed an
algorithm for the NP-hard problem of minimising This function is represented graphically as a
a supermodular function that are applicable to a U-shaped curve (Varian, 1984) or a concave
wide class of problems, including transportation upward function (Nicholson, 1986), where c1 > 0
with ®xed costs and facility location problems. and c2 6ˆ 0.
Submodular functions satisfy what is known in These functions illustrate the process of starting
microeconomic theory as diminishing returns to production with little know-how so that TC in-
scale or the ``law of diminishing returns'' (Gold- creases as x increases but at decreasing rates until
engorin et al., 1999). the ®rst saddle point, where d…CT…x††=d…x† ˆ 0
i.e., the following inequalities hold in this region:
MC ˆ d…CT…x††=d…x† > 0; …6†
2. The law of diminishing returns and total cost
2
functions d2 …CT…x††=d…x† < 0: …7†

The law of diminishing returns state that ``when The second saddle point, which is the one in the
units of a variable factor (e.g. labour) are added to region of interest in a cost minimisation problem is
a ®xed factor (e.g. land), the output resulting . . . reached as TC starts to decrease. In this region, the
(®rst) increases at an increasing rate, then increases ®rm becomes ecient by adding more variable
at a decreasing rate, and ®nally decrease'' (Nich- costs to ®xed resources while operating within
olson, 1986, p. 210). It causes the typical total cost capacity constraints. However, the law of dimin-
function to be cubic as a function of the output. ishing returns will cause TC to decrease at de-
The same set of reasonable assumptions about creasing rates because ®xed resources are
diminishing returns generate a typically quadratic becoming gradually more stretched. In this region,
marginal cost function. A real-valued total cost the following inequalities hold:
function de®ned over a set X ‡ for x > 0 can be MC ˆ d…CT…x††=d…x† < 0; …8†
represented as shown below
2
d2 …CT…x††=d…x† < 0: …9†
‰TCŠ TC…x† ˆ b0 ‡ b1 x ‡ b2 x2 ‡ b3 x3
for x  X ‡ ; Finally, after the second saddle point total costs
start to increase because ®xed resources have been
where TC is total repair cost and x  X the total used to capacity and there is an e€ect similar to
quantity of output, b0 represents ®xed costs (which ``crowding'' of variable resources, as when too
do not vary with the output) and b3 6ˆ 0. Although many workers are ®tted within a limited size
no de®nite statements can be made about these physical plant. In this region
functions, they often have two turning points with
MC ˆ d…CT…x††=d…x† > 0; …10†
one segment concave and another segment convex
(Nicholson, 1986, p. 209). d2 …CT…x††=d…x†2 > 0: …11†
Another concept used in the discussion below is
marginal cost de®ned as the relative changes in The region of interest for cost minimisation is
TC…x† with respect to x or, in mathematical terms, when the ®rm is operating eciently within ca-
d…TC…x††=d…x†. Given the cubic TC function pacity constraints, i.e. inequalities (8) and (9) hold
 L. Barros, M. Riley / European Journal of Operational Research 129 (2001) 242±251 245

and TC is continuous and convex (see Varian, that minimises total costs for each type of repair
1984 for more complete mathematical treatment of facility (including non-repair or discard). This
these concepts and Alfredsson, 1997 for a de®ni- approach ®ts in a wide class of operational re-
tion of ecient points). What makes these state- search problems known collectively as the facility
ments particularly interesting is the fact that location problem.
inequality (8) corresponds to the de®nition of a Tcha and Lee (1984) proposes the following
submodular function assuming TC to be a con- formulation for the multi-level uncapacitated fa-
tinuous function in the domain of interest. Poly- cility location problem. Let ciw be the pro®t asso-
nomial algorithms exist for the minimisation of ciated with satisfying the demand by client i via
submodular functions such as the objective func- facility j located at level jr 2 Jr and xiw be the
tion in [P1]. This issue is discussed further in fraction of demand by client i 2 I satis®ed by path
Section 3 below. w 2 P . Let P …jr †  P be the set of all possible paths
For the particular case of LORA, the output x which include facility jr and fjr ‡ be the associated
is the quantity of repair actions to be performed ®xed costs with opening such a facility. Then for
during the hardware lifetime and TC is total repair each level r there is a Boolean variable yjr such that
costs, typically calculated as the present value of 
annual repair costs over the lifetime of the hard- 1 if facility j in level r is open;
yjr ˆ
ware. Total repair cost models contain ®xed and 0 otherwise:
variable costs, as discussed below. The multi-level uncapacitated facility location
problem can then be formulated as follows:
XX XX
3. LORA ‰LOCŠ Max ciw xiw ÿ fjr yjr
i2I w2P r2R jr 2Jr
Multi-echelon maintenance systems have at-
tracted much attention from researchers and subject to:
practitioners for the past 20 years. The ®rst broad X
xiw ˆ 1 for i 2 I; …12†
literature review, now a classic reference in the w2P
®eld, is Pierskalla and Voelker (1971); Blanchard
(1986) is another classic reference in the ®eld of yjr P xiw for i 2 I; w 2 P …jr †; jr 2 Jr ; r 2 R; …13†
Engineering Logistics as is Sherbrooke (1992) yjr 2 f0; 1g; jr 2 Jr ; r 2 R: …14†
concerning spares inventory models. More recent
publications include Barros (1998) and Diaz and Constraint set (12) ensures that each customer
Fu (1997). is served by a single facility and the constraint set
LORA is a relatively specialised domain in this (13) ensures that such facilities are open (and thus
broad ®eld and is a design tool to help research incur ®xed costs). Constraint set (15) de®nes
and development engineers evaluate the long-term [LOC] as a mixed-integer problem. It is easy to see
impact of early design decisions on life-cycle that constraints (1), (2) and (5) in formulation [P1]
maintenance costs. Spare parts, inventory holding correspond to constraints (12), (13) and (14) in
and ordering costs, labour charges, support and formulation [LOC].
technical equipment and transportation costs are There are however, two important di€erences
some important items in the evaluation of life- between these two formulations:
cycle costs (LCC). (a) The objective function in [P1] minimises total
Formulation [P1] above aggregates such costs costs (variable plus ®xed costs) while the maximi-
under two headings, ®xed and variable costs fol- sation of pro®t minus ®xed costs is the objective
lowing standard microeconomic theory as dis- function in [LOC]. The ®xed cost component in
cussed in Section 2. It deals with the decision of both formulations is exactly equivalent since
where to perform the repairs once TC (x ) has been min FC ˆ max…ÿFC†. The remaining terms are
calculated, where x is the number of repair actions discussed in more detail below.
246 L. Barros, M. Riley / European Journal of Operational Research 129 (2001) 242±251

(b) Constraint sets (3) and (4) in [P1] make that If an additional facility j is added to this set, then it
problem a special case of the multi-level un- may imply an increase in ®xed costs, i.e.
capacitated problem [LOC] when there are X X
bundled commodities or clients. In other words, TC…x† ˆ vcs xs ‡ fcs ‡ fcj …19†
s2S s2S
some clients must be served by the same ware-
house, while others must be served by di€erent and/or an increase in the volume of output for the
warehouses. It just happens that these prefer- same unit variable cost
ences are expressed across levels in [P1], for in- X X
stance if a client l is assigned to facility j in TC…x† ˆ vcs xs ‡ fcs …20†
level r then client m 6ˆ I in level t 6ˆ r must be as- s2S[fjg s2S

signed to the same facility (for instance, an

or both. If there are economies of scale, then a
equipment supplier wants to supply main hard-
larger set of facilities may have larger ®xed costs
ware and spare parts from the same warehouse).
and handle larger volumes but unit average costs
The issue of maximising pro®ts minus ®xed
will be lower, i.e., for S  Y  T :
costs used in [LOC] needs further analysis. In
X X X X
microeconomic theory, pro®ts P…x† are de®ned for vcs xs ‡ fcs < vcy xy ‡ fcy ; …21†
a set of possible positive outputs X as the di€er- s2S s2S y2Y y2Y
ence between total revenue TR…x† ± de®ned as the
product of price p and positive quantity x 2 X ± where xs < xy ; fcs 6 fcy and vcs > vcy . These
and total costs TC…x†. The mathematical repre- inequalities will be used to prove the following
sentation is as follows: theorem:

TR…x† ˆ px for x 2 X ; …15† Theorem. TC(x) is submodular in the presence of

P…x† ˆ TR…x†TC…x† for x 2 X : …16† economies of scale.

Replacing (16) and [TC] in (17), the resulting Proof. Submodularity means that for S  Y  T
equation becomes and j 2 T ÿ Y the following inequality holds (see
Benati, 1999):
P…x† ˆ px…b0 ‡ b1 x ‡ b2 x2 ‡ b3 x3 †; …17†
TC…S [ fjg†TC…S† P TC…Y [ fjg†TC…Y †; …22†
where b0 represents ®xed costs and
b1 x ‡ b2 x2 ‡ b3 x3 represent variable costs. The where TC(S) is de®ned as in (18). If there are only
representation of P…x† as the di€erence between changes in ®xed costs, then TC…S [ fjg† is de®ned
total revenue and ®xed costs as used in Tcha and as in (19). Consequently
Lee (1984) is rather unusual. On the other hand,
TC…S [ fjg† ÿ TC…S† ˆ TC…Y [ fjg†TC…Y † ˆ fcj :
formulation [P1] is in agreement with standard
microeconomic theory. In addition, the maximi- If there are economies of scale, then variable costs
sation of a submodular function is a NP-hard will be di€erent for the two sets of facilities and
problem, while this is not true for the minimisation vcs > vcy . Since
problem (Lovasz, 1983).
The submodularity of the objective function in TC…S [ fjg† ÿ TC…S† ˆ vcs xj ‡ fcj ;
[P1] can be proven using concepts such as econo-
mies of scale. Rewriting that objective function TC…Y [ fjg† ÿ TC…Y † ˆ vcy xj ‡ fcj ;
and specifying more clearly variable and ®xed
costs for a set of facilities S  Y  T and output the submodularity condition (23) holds.
x 2 X , the following equation results: The argument being put forward is that disre-
X X garding the bundling constraints for the moment,
TC…x† ˆ vcs xs ‡ fcs : …18† LORA problems are a particular application of
s2S s2S the more generic multi-level uncapacitated facility
 L. Barros, M. Riley / European Journal of Operational Research 129 (2001) 242±251 247

location problem with a few advantages, in par- Table 1

ticular the fact that it has for objective function the Feasible repair postures for two-echelon two-indenture level
systems
minimisation of a submodular function rather
than the NP-hard problem of maximising a sub- Subsystem repair option Module repair option
modular function. Therefore, LORA problems can D C L
bene®t from the growing facility location litera- Discard (D) 1 ± ±
ture, starting with the classic reference Erlenkotter Central repair (C) 2 3 ±
Local repair (L) 4 5 6
(1978) and more recently Jones et al. (1995),
Berman et al. (1995), Benati (1999) and Galvao
(1993). Conversely, the algorithm level of repair
optimisation model (LOROM) described in feasible repair postures for two-indenture level
Section 4 below can be applied to solve multi- systems (subsystems and modules) with two-eche-
level uncapacitated location problems as well as lons of maintenance (local and central repair as
multi-echelon multi-indenture LORA problems. well as discard). The other three options are either
logically inconsistent or uneconomical:
· If a subsystem is discarded upon failure, then
4. The LOROM algorithm enclosed modules must be equally discarded.
The reverse is not true.
The LOROM relies on the basic premise that · If a subsystem is sent to a central facility for re-
there is a ®nite and relatively small number of in- pair, then it is uneconomical to ship enclosed
denture levels and echelons of maintenance. UK modules back to the local facility for repair.
and USA military standards de®ne two indenture The reverse is not true.
levels (LRUs and SRUs, or subsystems and en- The LOROM algorithm can be described in four
closed modules) and two echelons of maintenance steps:
(base and depot). French military standards de®ne Step 1. Generate a relaxed formulation of the
®ve echelons of maintenance essentially by subdi- original problem [P10 ] by ignoring the sets of
viding base and depot echelons into two sub- bundling constraints (3) and (4). The new problem
groups. is then
Although LOROM is generic, problem size XX XX
‰P10 Š MINIMISE vcrx nrx ‡ fcri mri
grows exponentially with the number of indenture r2R x2X r2R1 i2I
levels I and echelons of maintenance R, given the
fact that the base case decision tree contains 2IR subject to:
branches. For I ˆ R ˆ 2 the tree contains 16 X
branches. The number of calculations in later steps nrx for x 2 X ; …1†
increases linearly with the number of components r2R

in the system, and these may run into several mri P nrx for r 2 R; i 2 I; x 2 X ; …2†
thousands for a reasonably large application. The
advantage of this non-LP based branch-and- mri ; nrx 2 f0; 1g for r 2 R; i 2 I; x 2 X : …3†
bound algorithm is that calculations are simple
algebraic sums and can be performed very quickly. Solve this problem by satisfying ®rst the ``eas-
As noted in Section 3, the bundling constraints iest'' constraint set (1). This is done by generating
are the most peculiar characteristic of formulation a base case tree with 2IR branches where all facili-
[P1]. These constraints force coherent repair deci- ties are initially open (mri ˆ 1 for rr 2 R and
sions across indenture levels, taking into account ri 2 I) and subsequently closed one by one until
the fact that some son components may be en- all facilities are closed (mri ˆ 0 for r 2 R and
closed into several higher indenture father com- ri 2 I). In general, not all 2IR branches need to be
ponents (a property called commonality among considered in Step 2 of the algorithm after a quick
hardware designers). Table 1 below shows six check against constraint sets (3) and (4) of the
248 L. Barros, M. Riley / European Journal of Operational Research 129 (2001) 242±251

original problem [P1]. In the case of the two- inated solutions and do not need to be examined.
indenture level two-echelon system illustrated in The recalculation of total repair costs after each
Table 1, there is no need to open any module adjustment is quite simple. Although there is no
repair facility if none are open at the subsystem guarantee that bounds will be improved during
level. Similarly, there is no need to open a local Step 3 of the algorithm, the de®nition of search
repair facility for modules if none are open for rules by infeasibility type can be rather useful in
subsystems. Step 4.
Step 2. Select least cost repair options for in- Step 4. Perform full search among the remain-
dividual system components under each base case ing base cases. Start from base case solutions,
thus respecting the ®xed cost constraints ± and which satisfy constraint sets (1) and (2) and elim-
obtain lower and upper bounds on the optimum inate infeasibilities against constraint set (3) by
solution for problem [P1]. If all facilities are open switching repair options for father±son component
at all indenture levels, then the selection of the combinations, ignoring dominated solutions. This
least cost repair option can be done by inspection. is equivalent to the depth search ®rst and then
If none are opened, then only discard is permitted back-tracking strategy described in Benati (1999).
this ± solution is necessarily feasible in [P1]. Other Real-life experience with LOROM has shown
base cases in the tree de®ne intermediate combi- that few cases are left after Steps 2 and 3 for full
natorial options between these two extremes. search and the optimum solution is quickly iden-
The overall least cost solution provides a lower ti®ed. The ratio between upper and lower bounds
bound on the optimum solution of the original and on the optimum solution tends to lie between 10%
more constrained problem [P1]. Evidently, if this and 20%, which is acceptable in many instances
lower bound is also feasible in [P1], then the op- during hardware design when cost and reliability
timum solution has to be identi®ed and there is no parameters are very imprecise. Experience has also
need to proceed to Step 3 of the algorithm. shown that the higher the ®xed costs relative to
For the upper bound, it is necessary to examine variable costs, the better the algorithm performs
base cases, which provide solutions with no pos- due to the fact that discard upon failure becomes a
sible incompatibility between father and son repair very cost-e€ective option.
options. These include all base cases where mri ˆ 0
for i P 1. The least cost solution among these
provides an upper bound on the optimum solution 5. An illustration: The I ˆ R ˆ 2 system
for the original problem [P1]. Evidently, if higher
and lower bounds are identical, then the optimum Fig. 1 illustrates Steps 1 and 2 of the LOROM
solution for [P1] has been identi®ed and there is no algorithm for a two-indenture two-echelon system.
need to proceed to Step 3 of the algorithm. If this The 16 branches are generated by setting mri to 1
is not the case, then the upper bound can be used and 0 at each indenture level (subsystems and
to fathom by value other base cases generated in modules). When mri ˆ 0, constraint set (2) implies
Step 1 of the algorithm. that nrx ˆ 0 for all subsystems t 2 X and all mod-
Step 3. Use a greedy heuristic algorithm to ules y 2 X . The base cases shown as INF provide
obtain tighter upper bounds on the optimum so- solutions necessarily infeasible in [P1] for any en-
lution for [P1] and cut further branches from the closed module y 2 Ct . On the other hand, base
base case decision tree which are fathomed by cases 4, 8, 10 and 11 provide solutions that are
value. The complicating bundling constraint set (3) necessarily feasible in [P1] since all modules in the
in problem [P1] are reintroduced at this step. In- system are discarded.
feasibility types can generally be de®ned by careful Step 3 requires the identi®cation of infeasibility
inspection of this constraint set and repair options types and adjustment rules for component repair
at each indenture level for sets of father and son option. In the case of I ˆ R ˆ 2 system, there are
components can be adjusted accordingly. In three possible types of infeasibilities, shown as
practice, a large number of options provide dom- blanks in Table 1:
 L. Barros, M. Riley / European Journal of Operational Research 129 (2001) 242±251 249

Fig. 1. Base case tree for the I ˆ R ˆ 2 system.

· central repair of enclosed modules when the only three cost-increasing repair cost adjust-
subsystem is discarded, ments to eliminate this infeasibility for all
· local repair of enclosed modules when the sub- modules enclosed in a given subsystem t where
system is discarded, and y 2 Ct :
· local repair of enclosed modules when the sub- · repair the subsystem at the local facility (posture
system is repaired at the central facility. 6);
Take for example a base case solution that · discard the module (posture 1);
is infeasible in the original problem [P1] because · repair both subsystem and module at the central
of the second type of infeasibility. There are facility (posture 3).
250 L. Barros, M. Riley / European Journal of Operational Research 129 (2001) 242±251

All other alternatives listed in Table 1 provide mality at this ®nal stage also bene®ts from the
dominated solutions: repair postures 4 and 5 are special structure of the original problem that
dominated by 6, and postures 2 and 4 are domi- eliminates dominated solutions from further con-
nated by 1. Depending on the base case where the sideration. A two-indenture level two-echelon of
infeasibility under scrutiny is located, further re- maintenance system is used as an illustration for
strictions may apply. In addition, not all three the LOROM algorithm. The illustration is rather
types of infeasibilities can occur in all base cases, realistic as it accurately portraits the recommen-
further simplifying the search procedure. For in- dations in USA and UK military standard hand-
stance, the second infeasibility type can only occur, books.
where m12 ˆ 1 in Fig. 1, i.e., base cases number 1, The rest of the paper argues that this type of
2, 5 and 6. problem ®ts into a more generic class of Opera-
Similar analysis for the other two types of in- tional Research problems, in particular the multi-
feasibilities can show that the ®rst type (local level uncapacitated location problem. LORA
module repair and subsystem discard) can only problems have the advantage of proposing to
occur when m11 ˆ 1 in Fig. 1 as in base cases 1±8 in minimise a submodular function rather than the
Fig. 1. The third type requires m21 ˆ m12 ˆ 1 and NP-hard pro®t maximisation formulation used in
can only occur in bases cases 1 and 2. In all cases, location theory. Finally, microeconomic mathe-
feasibility can be attained by examining a few cost- matical developments show theoretical links be-
increasing alternatives because other postures in tween submodularity and the law of diminishing
Table 1 provide dominated solutions or are not returns as well as economies of scale.
allowed in a particular base case.

Acknowledgements
6. Conclusions
The author wants to acknowledge the contri-
This paper presents an approach to solve multi-
bution of Dr. Mike Tso from UMIST in UK
indenture multi-echelon LORA problems based on
to this paper through useful discussions on
a branch-and-bound procedure that does not re-
submodularity and the provision of insightful
quire the use of LP relaxations. Instead, the LO-
published literature on maintenance planning.
ROM algorithm uses a relaxed version of the
original problem to create a tree with base cases,
which satisfy only the ®xed cost constraints in the
original problem. Among these cases, several so- References
lutions are necessarily feasible in the original
Alfredsson, A., 1997. Optimization of multi-echelon repairable
problem thus providing upper bounds on the value
item inventory systems with simultaneous location of
of the objective function. repair facilities. European Journal of Operations Research
The bundling constraints in the original prob- 99, 584±595.
lem, which link repair decisions at di€erent in- Barros, L., 1998. The optimisation of repair decisions using life-
denture levels are brought in at a later step in the cycle cost parameters. IMA 9 (4), 403±413.
Benati, S., 1999. The maximum capture problem with hetero-
algorithm. These impose that repair decisions
geneous customers. Computers and Operations Research
should be consistent for father and son compo- 26, 1351±1367.
nents in a hardware system. A relatively simple Berman, O., Bertsimans D., Larson, R.C., 1995. Locating
heuristics provides lower and upper bounds on the Discretionary Service Facilities, II Maximizing market size,
value of the optimum solution. Experience has minimizing inconvenience. Operations Research 43 (4).
Blanchard, B.S., 1986. Logistics Engineering and Management.
shown that the gap between upper and lower
Prentice-Hall, Englewood Cli€s, NJ.
bounds obtained in early steps of LOROM is Diaz, A., Fu, M., 1997. Multi-echelon inventory systems for
rather tight, and few branches of the original tree repairable items with limited repair facilities. European
are left for full enumeration. The search of opti- Journal of Operational Research 97, 480±492.
 L. Barros, M. Riley / European Journal of Operational Research 129 (2001) 242±251 251

Erlenkotter, D., 1978. A dual-based procedure for uncapaci- Lovasz, L., 1983. Submodular functions and convexity. In:
tated facility location. Operations Research 26 (6), Bachem, A., Grotschel, M., Korte, B. (Eds.), Mathematical
992±1009. Programming: The State of the Art, Springer, New York.
Galvao, R.D., 1993. The use of Lagrangian relaxation in the Nicholson, R.H., 1986. Mathematics for Business and Eco-
solution of uncapacitated facility location problem. Loca- nomics. McGraw-Hill, New York.
tion Science 1, 57±79. Pierskalla, M.P., Voelker, J.A., 1971. A survey of maintenance
Goldengorin, B., Sierksma, G., Tijssen, G.A., Tso, M., 1999. models: The control and surveillance of deteriorating
The data-correcting algorithm for the minimization of systems. Naval Research Logistics Quarterly 21 (2),
supermodular functions. Management Science 45 (11) 213±237.
1539±1551. Sherbrooke, C.C., 1992. Optimal Inventory Modeling of
Jones, P.C., Lowe, T.J., Mullere, G., Xu, N., Ye, Y., Zydiak, Systems: Multi-echelon Techniques. Wiley, Chichester.
J.L., 1995. Specially structured uncapacitated facility Tcha, D., Lee, B., 1984. A branch-and-bound algorithm for
location problems. Operations Research 43 (4), 661±669. the multi-level uncapacitated facility location problem.
Kubiak, W., 1995. New results on the completion time European Journal of Operational Research 18, 35±43.
variance minimization. Discrete Applied Mathematics 58, Varian, H.R., 1984. Microeconomic Analysis. W.W. Norton
157±168. and Company, New York.