Вы находитесь на странице: 1из 8

Metaheuristic based optimization for capacitated

disassembly lot sizing problem with lost sales


(presented at the 6th IESM Conference, October 2015, Seville, Spain) © I4e2 2015

Matthieu Godichaud, Lionel Amodeo, Mustapha Hrouga


Institut Charles Delaunay, LOSI, Université de Technologie de Troyes, UMR 6281, CNRS,
Troyes, France
{ matthieu.godichaud, lionel.amodeo,mustapha.hrouga}@utt.fr

Abstract—Disassembly operations are required for most of one product several component demands have to be satisfied.
manufactured products at the end of their life cycle. As economic Due to the yields (number of each component contained in one
activities and environmental pressures increase, the volume of product) and the demands between components that are not
product reverse flows are more and more important and costly. necessarily well balanced, high inventory surplus can be
In this context, we propose an optimization method to minimize generated throughout the planning horizon [3]. Disassembly lot
cost in disassembly planning with lot sizing and lost sales. sizing models with lost sales or (component) outsourcing are
Comparing to others lot sizing problems with lost sales, this thus interested to apply when it is possible in real industrial
problem has some specificities that required original cases.
optimization methods. To this end, we proposed a metaheuristic
based on genetic algorithm scheme that integrates some In this paper we consider multi-product disassembly lot
neighborhoods dedicated to this problem. The quality of the sizing problem with two-level product structures, setup and
solutions is compared with those obtained from a mathematical inventory cost, lost sales costs and limited capacity. Part
programming solver for small instances and different commonalities between products are not considered. To our
configurations of the algorithm are compared. The metaheuristic knowledge, this problem has not been studied in literature. In
allows finding good solutions in a reasonable computational time section II, a literature review is proposed on disassembly
for this tactical problem for all instance sizes. planning (without lost sales) and on production lost sizing with
lost sales. In section III, a mixed integer program (MIP) model
Keywords—Disassembly; Reverse logistic; lot sizing; Mixed
integer programming; metaheuristic
is presented. The MIP solver CPLEX do not give solutions in a
reasonable amount of time for large instances of the problem.
I. INTRODUCTION Our contribution is presented in section IV which is a
metaheuristic based approach with use several dedicated
Nowadays, manufacturing enterprise are responsive for the neighborhoods. An experimental study is also proposed in
products they produce until the end of their life cycle. Reverse section IV.
logistic aims at recovering end-of-life products to supply
recycling, remanufacturing or repair process or to dispose them II. LITERATURE REVIEW
in an environmentally conscious way. To this end, most of
manufactured products have to be disassembled in order to The disassembly planning problem for a single product
separate their materials and theirs components which have to without explicit cost function can be treated by a reversed
be assigned to different process. With current increasing Material Requirement Planning (MRP) approach [4]. This
environmental pressures, disassembly activities are caused to problem was further extended in [5] to include parts
develop and planning decisions are becoming more difficult. commonality for multiple product disassembly. The
capacitated problem can be modeled with an integer
In this paper, a disassembly planning defines how many programming model as in [6] for the case of single product
end-of-life products to disassemble knowing the demands of type. When set up costs and inventory costs are considered in
their components in each period of a finite and discrete the objective function, lost sizing decision have to be made. In
planning horizon. The objective is to satisfy customer demands this case, several lot sizing heuristic are studied in [7] to
while minimizing several costs. In disassembly lot sizing, improve solutions obtained from reverse MRP algorithm. MIP
inventory costs and disassembly costs are considered. As in models can be used for the problem with part commonalities
single item production lot sizing, disassembly (production) and relaxation based heuristic gives good solutions in
quantities of several periods are grouped into a single one to reasonable time [8] [9]. The capacitated problem is solved in
decrease disassembly costs (particularly fixed set-up cost) [10] without part commonalities with a Lagrangian relaxation
while increasing inventory costs. Optimization methods are that gives good lower and upper bounds. The approach also
used to find solutions with minimum costs [1]. However, integrates random demands and backorders. The relaxed
disassembly lot sizing problems have some specificities and problem is solved by using dynamic program based on the zero
need dedicated models and optimizations methods [2]. inventory property traditionally used in lot sizing [11]. This
property is not valid for the case with lost sales and the
One of the main specificity of disassembly lot sizing
previous approach cannot be used.
problems compared to others lot sizing problems is that with
The production lot sizing problem with lost sales has been contains in the product. The operation disassembly of root 1
studied by several authors. The single item case can be solved extracts 3 leaf items of type 2, 2 leaf items of type 3 and 2 leaf
by dynamic programming method in [12] for constant items of type 4.
parameters and in [13] for time varying parameter. They are
based on three main properties of optimal solutions that allow 2 3 4
decomposing the problem into subproblems : (1) demand is
fully satisfied in periods with strictly positive production (2)
(3) (2)
quantity, (2) a period with strictly positive production quantity
occurs only if the inventory level of the preceding period is
1
zero and (3) if any demand is lost in a period then the entire
demand of that period should be lost. The method is extended Fig. 1. Two-level product structure.
in [14] for problems with limited inventory by using similar
properties. For the case with production capacity, this approach The cost function to minimize is composed of setup cost for
could also be coupled with a lagrangian relaxation method each root item in each period, inventory cost and lost sales cost
[15]. The authors state that the capacitated lot sizing problem for each leaf item in each period and penalty cost for
with lost sales is NP-Hard and propose heuristic to solve the overloading disassembly capacity in each period.
problem. However, it can be shown on a simple instance (as in
section III) that these properties are not always valid for Assumptions for this model are summarized as follows:
optimal solutions of disassembly lot sizing problems with lost (a) demands for leaf items are given and deterministic;
sales. (b) lost sales is allowed, hence demand can be not satisfied;
(c) the disassembly process is perfect, all parts are in perfect
Given the previous references, metaheuristic algorithm can quality, no defective are considered;
be an approach to have good solutions in reasonable time for (d) disassembly operation times are given and deterministic;
disassembly lot sizing problems with lost sales. Different (e) the cost parameters are considered as constants.
metaheuristics had been proposed for lot sizing problems and a
review is presented in [16]. Among these metaheuristics The notations are summarized below.
genetic algorithm are widely used [17]. The authors present the
features and specifications to design genetic algorithm for lot Index
sizing problem. In [18], the authors proposed a genetic Index for root items, 1,2, … ,
algorithm for disassembly scheduling for multi-level product Index for leaf items, 1,2, … ,
structure. The solution are coding with the binary setup Index for periods, 1,2, … ,
variables only and the disassembly quantities in each period are
deduced from inventory variables and demands. This approach Parameters
is not possible for the lost sales case and a direct encoding of Setup cost of parent item .
solution based on disassembly quantities is prefered in this Capacity available, in time, in period .
paper.
Parent of leaf item .
To the best of the authors’ knowledge, no one has Disassembly operation time of root item .
addressed the optimization of capacitated disassembly lot Inventory holding cost of item .
sizing with lost sales and no method is available to provide Lost sales cost for in period .
solutions in reasonable time. We propose a metaheuristic Penalty cost disassembly time in period .
approach based on a genetic algorithm scheme with several Demand of item i in period .
neighborhoods to improve its efficiency. Number of unit of items obtained by disassembly of
III. MODEL STATEMENT one unit of its parent item .
Initial inventory of item .
In this section we propose a mixed integer programming
Large Number.
(MIP) model of the problem. Before formulating the
mathematical model, the disassembly process is described.
Variables
A parent (root) item can be disassembled to produce a Binary setup variable of root item in period .
specific number of child items. The demand is defined for the Disassembly quantity of root item through period .
leaf item (with no child items) only. Given a set of root items, Disassembly over-load in period .
the demand of each leaf items of all roots is given over a time Inventory level of leaf item at the end of period .
horizon. The total disassembly quantities is limited in each Lost sales for each leaf item in period .
period by a disassembly capacity, exceeding this capacity will
result a penalty cost. If the demand of a leaf item is not met in a
The disassembly lot sizing problem with lost sales (P) is
period it will be considered lost. The problem is to determine
the quantity and timing of disassembly of all root items to formulated as follow.
minimize a cost function. In this paper we consider only two
level product structure without part (leaf) commonalities. Fig. 1
displays an example of the disassembly structure for a product
with three child items. The yield is the number of child items
Subj ect t o: each period. In period 1, 2 and 3 there is one component for
which and the “zero inventory properties” (it is not
zero due to yield relationship) holds. However in period 2, we
(1) , , , f or 1,2, … , , have , , and period 3 is a setup
period. The properties do not hold in period 4.
1,2, … , and 1, … , ⁄ 0
We define a set of instance based on the setting in [10]
(2) f or 1,2, … , and 1,2, … , (described in the next section) to analyze the opportunity to use
CPLEX solver (version 12.6). The size of the instance is
(3) ∑ f or 1,2, … , defined in relation to the number of period, the number of root
and the number of child items per roots. CPLEX solver
(4) f or 1,2, … , and 1,2, … , provides optimal solutions in less than one hour only for small
instances with ten periods, ten roots and between one and ten
(5) , , 0 and ar e i nt eger f or 1,2, … , , child items per roots. For larger instances, we propose in the
,
next section a metaheuristic.
1,2, … , , 1,2, … ,
IV. A METAHEURISTIC BASED METHOD
(6) 0.1 f or 1,2, … , and 1,2, … , A metaheuristic method is presented in this section. It is a
genetic algorithm scheme improved by some neighborhoods.
The objective function is a minimization of the sum of The quality of the solution is evaluated compared to the
setup, inventory holding, lost sales, and overload cost. optimal solutions of small instances and then we compare
Constraint (1) represents the inventory flow conservation of several configurations of the metaheuristic.
leaf items at the end of each period. Constraint (2) guarantees
that a setup cost in a period is incurred if there is any A. Optimization scheme
disassembly operation at that period. Constraint (3) represents A genetic algorithm based approach is proposed to solve
the capacity constraint in each period. The total time required the problem. This type of algorithm is first characterized by the
to perform the disassembly operations in each period should be solution encoding and the solution evaluation method and,
less than or equal to the available disassembly capacity in that then, by an initialization step and search procedures that
period otherwise overtime is used with a proportional cost of . change the solution [19]. The initialization step allows
generating a set of solutions to form a population. The search
Constraint (4) ensures that any lost demand in period cannot
procedures, mutation and crossover, are iteratively repeated to
exceed the demand in that period. Constraint (5) assures that search for the best solution. This is the generic scheme of a
inventory level, lost sales and disassembly quantity are integer genetic algorithm which has to be adapted to the problem to
values and non-negatives. Setup variables are defined as binary solve. The adaptations for the disassembly lot sizing problem
in constraint (6). are presented in this section.
The example presented in Table 1 shows that the 1) Solution encoding
properties that allow decomposing a production lot sizing A disassembly planning over a time horizon decomposed
problem with lost sales into successive subproblems do not into periods 1, … , is given by the value of , variables.
hold for the disassembly case. It is a four periods problems All other variables of (P) can be computed based on these
with one products containing three leaf items. The demands are values and, therefore, the cost of a planning. A direct encoding
presented in Table 1, the cost parameters are 200, 50 of solution is used in the proposed algorithm with a matrix
for all leaf items, 100 for all leaf items and there is no form whose elements correspond to , . A disassembly period
capacity restriction. The yield are defined in Fig. 1. This is a for a root item , with a set-up ( , 1), correspond to a
simple example that can be solved with any MIP solver. The value , 0 and a period without disassembly correspond to
optimal solution is displayed in Table 1. a value , 0, no set-up ( , 0). The value of variables
TABLE 1. A FOUR PERIODS, THREE COMPONENTS EXAMPLE , and , are computed from the values of variables , by
rewriting the flow conservation constraint (1) : ,
1 2 3 4
max 0, , , , , and , max 0, ,
70 50 100 90
, , , . These are the best values when the unit
80 50 60 70 holding cost and unit lost sales cost are constants in all periods.
65 50 60 90 There is no advantages from a cost point of view to keep items
24 20 30 38 in stock in a given period if there are demands to be satisfied in
/ 2/0 12 / 0 2/0 26 / 0 this period and the demand are satisfied as soon as they can be.
/ 0 / 32 0 / 10 0/0 1/0 The capacity constraint is rewriting to compute the values of
/ 0 / 17 0 / 10 0/0 0 / 14 variables : max 0, ∑ , . The capacity overrun is
then possible and it corresponds to the overtime which is
penalized in the objective function.
There is a setup in each period and if a decomposition in
subproblem could be applied, the properties should be valid for
All integer values of , are then allowed and there is no The second method is divided into 3 steps for each root
need for constraint handling method in the algorithm. item to generate solutions satisfying (P1) and (P2).
However, we proposed in the following section maximum and According to these two properties, the value of a , 0 is
minimum values ( , and , ) when , 0 to limit the equal to the sum of the demand of one of its component on one
search space to potentially better solutions. or many successive periods between and the next
disassembly period ’ (i.e. ’ is defined compared to as
All the objective function components can be computed 0, 0 and 0 1 ; 1 ). The
, , ,
from the value of , variables. The evaluation method takes three steps are the following.
, values as input, computes all the variables and implements
the objective function. Step 1. Random choice of disassembly periods between 0 and
.
2) Initialization method Step 2. For each disassembly period , random choice of a
Several initialization methods are proposed to generate an target component of .
initial population of solutions i.e. generate integer value Step 3. For couple , , random choice of the number of
matrices. In lot sizing problem, disassembly period positioning period ∆ between and the next disassembly period of for
(periods with , 1 or , 0) is an important contributor which the demand of the target component will be totally
of cost variations between solutions and the choice between satisfied (∆ 0 ; 1 ).
, 0 and , 0 is as important as the choice of the value We deduce the value of , of a disassembly period
of , ( 0). A first initialization method consists in firstly given and ∆ according to equation (7).
randomly fix disassembly periods for each root item and then

generate a random value for , for each disassembly period. ∑ ,
, 7
A second method consists in generating solution with two ,
properties (P1) and (P2). For each disassembly period of a
root item , there is at least one component of with : 3) Crossover
Based on the solution encoding adopted, directly with the
• (P1) an inventory level less than , at period 1 values of , variable, a two point crossover method is used to
( , , ), generate two new solutions form two solutions of the preceding
generation. The two crossover points correspond to two periods
• (P2) if is the disassembly period preceding for root 1 and 2 : for a given period, all values of , for all root items
item , demands between and 1 are either are duplicated either in one new solution or the other based on
completely satisfied or totally lost ( , , , the position of (before 1 or after 2, between 1 and 2).
, , ; 1 ). Other crossover methods proposed in the literature have been
These two properties are adapted from the properties of tested but they do not provide any significant improvement for
optimal solution of single item lot sizing problem with lost this problem. The solutions are selected for crossover with a
sales. In the case of disassembly, an optimal solution do not binary tournament method based on the value of the objective
have necessarily these properties as presented in the example function of (1).
of Table 1. However, the gap between a neighbor solution 4) Mutation
satisfying (P1) and (P2) of an optimal solution not satisfying A first type of mutation consists in varying the value of one
(P1) and (P2) is small and on the basis of 100 instances of the
, 0 randomly respecting minimum value , and
problem with one root item and without capacity constraint
solving with CPLEX solver, 72% of the optimal solutions maximum value , according to equation (8) with a
satisfy (P1) and (P2). random number between 0 and 1. It is a generic mutation
method [20] non specific for this problem except for the
In the first method, the values of ( , 0) for disassembly calculation of , et , .
period are generated randomly according to uniform
distribution between a value , and a value , . The value (8) , , , , 2 if 0.5
, for a disassembly period and a root item corresponds 2 1 otherwise
, , , ,
to the maximum cumulated demands between and the next
disassembly period among the components of (set of with Two other types of mutation methods are proposed that are
, 0) : dedicated to the disassembly lot sizing problem. The first ones,
noted neighborhoods 1 et 2, consist in decreasing either
∑ , , inventory level either lost sales to zero for given periods and
, max components (only one variable is affected by this
/ , , ,
modification). The second ones consist in changing one
For the values of , , we consider that the minimum disassembly period by modifying one , from a value zero to
demand among the components of in period has to be a value greater than zero (neighborhood 3) or from a value
satisfied : greater than zero to a value equal to zero (neighborhood 4).
, Neighborhood 1 consists in decreasing a value of a
, min
/ , , , 0 in order to decreasing the inventory level of a
component of between and the next disassembly period .
For this purpose, for each component , the last period • inventory holding costs were generated from
; 1 with a strictly positive inventory level is detected U(1,10);
( , 0 and , 0 for all 1; 1 ). Among all
• : loss sales cost was generated from U(25,75);
the , detected, the minimum value is taken to decrease , .
In this manner, inventory periods of a component between • penalty costs for overtime were generated from
and (periods ; 1 with , 0) remain inventory U(5,15);
periods and, reversely, lost sales periods of a component
between and (periods ; 1 with , 0) • disassembly time was generated from U(1,3);
remain lost sales periods. According to this neighborhood, the • Initial inventory was generated from (20,100).
objective function varies linearly with a coefficient for all
the inventory period or with a coefficient for each lost sales • Available aggregate capacity in each period is set
periods for each component . to 600.

Neighborhood 2 consists in increasing a value of a Different tests with different parameter setting were made
0 in order to decreasing the inventory level of a to choose the most efficient parameter setting for the
,
component of between and the next disassembly period . algorithm. The metaheuristic is stochastic however the setting
In the same manner as neighborhood 1, for each component of the maximum number of generation and the population size
, the first period ; 1 with a strictly positive lost allow the algorithm to converge to same value in mean by
sales value is detected ( , 0 and , 0 for all making several start with the same instance. The following
1; 1 ) and then the minimum value is taken to increase parameters setting are used in our case study:
, . As the neighborhood 1, inventory (respectively lost • Maximum generation 1000.
sales) periods remain inventory (respectively lost sales) periods • Population size 200.
for each component of and the objective function varies • Mutation probability Coef 1/ .
linearly. • Crossover probability Coef 0,7.
Neighborhood 3 consists in adding a disassembly period To evaluate the quality of the metaheuristic solutions, we
for a product (i.e. a value , 0 is changed to a value have generated 100 small instances 10-S-10 (10 periods, low
, 0). If the disassembly period following is noted , a number of children, 10 roots). By testing 100 instances of the
target component of and ∆ , the number of periods where same size, we have a statistical significant image of the gap.
the demand for will be entirely satisfied between and , are The metaheuristics solutions are compared with the optimal
chosen. The choice of and ∆ can be made randomly or, with solution obtained from CPLEX solver. The gap between each
more computational time, by evaluating all possible couples solution is calculated as followed :
, ∆ and the one that gives the best solution is selected. The metaheuristic solution optimal solution
preceding period of is also be modified to have , 100
( is the target component chosen previously). If this is not metaheuristic solution
,
the case, , is decreased by a quantity , / , . The results from the 100 instances are :

Neighborhood 4 consists in withdrawing a disassembly - mean gap obtained : 0,615,


period for a root item . A value , 0 is changed to - maximum gap : 1,45,
, 0. If is the preceding period of , the quantity
withdrawn from , is added to , . - minimum gap : 0 ,
- standard deviation : 0,32.
B. Quality of the solutions
The algorithm tested in this paper was coded in Java and If we start 100 times one of the 100 instances, we have a
run on a personal computer with a processors operating at 2.50 probability of 99% to have a mean gap between 0,516 and
GHz clock speed. 0,708. We verify it by starting 100 times the first 10 instances.

For the test, instances are generated as in [10]. U(a,b) is the The computational time is less than 30s. The computational
discrete uniform distribution with a range of [a,b]. time of CPLEX with basic setting solver is varying widely and
it need more investigation to classify the small instances. We
• The size of the instances are characterized by three have generated 10 instances of each size and for all other
components : the number of root items (10,20,30), instances, no solution is found within one hour.
the number of children generated from a discrete
uniform distribution with a rang U(1,10) and C. Comparison of different metaheuristic configurations
U(10,100) for low (S), medium (M) and the number For instances larger than the 10-S-10s, we have compared
of periods (10,20,30). several configuration of the metaheuristic to quantify the added
value of the proposed neighborhood. The first configuration,
• setup cost for each root was generated from noted M1, corresponds to an “on the shelf” genetic algorithm
U(5000,10000); with the mutation method defined in equation (8). The second
• demand was generated from U(50,200); configuration, noted M2, corresponds to a genetic algorithm
improves with the four neighborhoods. For these two
configurations, we have noticed that the diversity of the 30 0,69/0,92/0,34/0,18
population could be lost while the best solution is a local 30 S 10 0,31/0,46/0,05/0,13
20 0,79/1,07/0,57/0,14
optimum. The third configuration, noted M3, aims at 30 1,32/1,69/0,93/0,24
overcoming this problem by performing a perturbation method 10 M 10 0/0/0/0
on several individuals of the population when the diversity is 20 0,2/0,2/0,1/0,0
lost (large number of indentical individuals). 30 0,4/0,5/0,3/0,1
20 M 10 0,2/0,2/0,1/0,0
The results are presented on Table 2. The configuration are 20 0,7/0,8/0,5/0,1
compared based on the mean gap on ten instances with the 30 1,2/1,4/1,1/0,1
same size. The best solution is always found with the 30 M 10 0,6/0,7/0,4/0,1
configuration M3 and configuration M2 is always better than 20 1,2/1,4/1,0/0,1
30 2,0/2,3/1,7/0,2
configuration M1. The improvement between M1 and M2 is
CPU Gap
larger than the improvement between M2 and M3. The M1-M3 M2-M3
difference between the configurations increases as the size of 10 S 20 -41,5 -31,8
the instances increases. The difference is worthwhile between 30 -59,8 -53,0
instances with low number of children per root and instances 20 S 10 -61,9 -57,4
with medium number of children per root. The CPU time 20 -61,7 -56,6
between between M1 and M2 increases. However the 30 -39,8 -31,0
30 S 10 -61,1 -56,6
perturbation method in M3 increases the CPU time particularly 20 -26,8 -17,0
for smaller instances. 30 -12,6 -0,8
TABLE 2. COMPARISON BETWEEN CONFIGURATIONS 10 M 10 -44,5 -42,9
20 -26,9 -22,9
Gap 30 -11,6 -6,1
R Child T M1-M3 20 M 10 -18,0 -14,5
mean /max/min/std 20 -5,0 -0,2
10 S 20 5,6/7,6/4,1/1,0 30 -5,6 -0,8
30 6,9/9,5/4,6/1,5 30 M 10 -4,1 -0,7
20 S 10 4,4/5,3/2,6/0,8 20 -4,2 -0,4
20 6,9/8,3/5,7/0,9 30 -4,0 -0,1
30 8,6/10,8/6,5/1,4
30 S 10 5,3/7,2/3,7/1,0 V. CONCLUSION
20 8,3/9,2/6,2/0,9
30 9,4/10,9/8,5/0,7 In this paper, we addressed the multi-products capacitated
10 M 10 1,9/2,5/1,4/0,4 disassembly planning with lost sales. To our knowledge, it is
20 4,6/5,2/3,9/0,4 the first time that disassembly planning problem with lost sales
30 6,6/7,6/5,3/0,8 is investigated. This problem has some specificities that do not
20 M 10 3,5/4,2/3,0/0,4 allow using method from others lot sizing problems with lost
20 8,4/10,2/7,2/1,0 sales. We proposed thus a metaheuristic that use genetic
30 11,1/12,6/9,7/0,9
30 M 10 5,6/6,7/5,0/0,5
algorithm scheme with some neighborhoods to improve its
20 11,0/13,4/9,8/1,2 efficiency. Indeed, for larger instance, the problem cannot be
30 14,1/16,1/12,4/1 solved in reasonable time with MIP solver. Furthermore, “on
M1-M2 the shelf” genetic algorithm is easily trap in local optimum.
mean /max/min/std The experimental study show that this algorithm gives good
10 S 20 5,4/6,7/3,8/0,9 solution in a reasonable time for this tactical problem.
30 6,8/9,4/4,5/1,5
20 S 10 4,2/5,3/2,6/0,8 The perspectives of the paper concern particularly the
20 6,6/7,8/5,4/0,8 improvement of the algorithm and the consideration of some
30 7,8/10,2/5,7/1,4
30 S 10 4,9/7,0/3,7/0,9
others specificities of disassembly process. The algorithm
20 7,4/8,3/5,5/0,8 could be improved by controlling its parameters during the
30 8,0/9,1/7,2/0,7 search. It could also be compared with others metaheuristic that
10 M 10 1,9/2,5/1,3/0,4 use the neighborhoods. Part commonalities and multi-level
20 4,4/5,0/3,8/0,4 product structure are two main features that are encountered in
30 6,3/7,3/5,0/0,7 real disassembly process. They necessitated new
20 M 10 3,3/4,0/2,8/0,4
neighborhoods. Uncertainties is also important in disassembly
20 7,7/9,4/6,6/1,0
30 9,7/11,0/8,6/0,7 process and new models have to be develop to manage it.
30 M 10 5,0/5,9/4,4/0,5
20 9,7/12,0/8,5/1,1 REFERENCES
30 11,8/13,6/10,2/0,9 [1] N. Brahimi, S. Dauzère-Pérès, N.M. Najid, A. Nordli. “Single item lot
M2-M3 sizing problems”. European Journal of Operational Research, 168(1), pp.
mean /max/min/std 1-16, 2006.
10 S 20 0,20/0,77/0/0,25 [2] H.-J. Kim, D.-H. Lee, and P. Xirouchakis, “Disassembly scheduling:
30 0,15/0,29/0,05/0,07 literature review and future research directions,” International Journal of
20 S 10 0,16/0,29/0,04/0,09 Production Research, vol. 45, no. 18–19, pp. 4465–4484, 2007.
20 0,26/0,48/0,08/0,13
[3] M. Hrouga, M. Godichaud, L. Amodeo.”Disassembly lot sizing problem [14] Liu, X., Yl. Tu. “Production planning with limited inventory capacity
whithout stockout”. Proceeding of the 5th International Conference on and allowed stockout.” International Journal of Production Economics
Information Systems and Supply Chain, ILS2014, August 2014, Breda, 111(1):180–191, 2008
The Netherlands. [15] N. Absi, B. Detienne, and S. Dauzère-Pérès, “Heuristics for the multi-
[4] S. M. Gupta and K. N. Taleb, “Scheduling disassembly,” International item capacitated lot-sizing problem with lost sales,” Computers &
Journal of Production Research, vol. 32, no. 8, pp. 1857–1866, 1994. Operations Research, vol. 40, no. 1, pp. 264–272, 2013.
[5] K. N. Taleb and S. M. Gupta, “Disassembly of multiple product [16] R. Jans , Z Degraeve. “Meta-heuristics for dynamic lot sizing: a review
structures,” Computers & Industrial Engineering, vol. 32, no. 4, pp. and comparison of solution approaches”. European Journal of
949–961, 1997. Operational Research 2007;177:1855–75.
[6] D.-H. Lee, P. Xirouchakis, and R. Zust, “Disassembly Scheduling with [17] H. G. Goren, Tunali S, R. Jan. “A review of applications of genetic
Capacity Constraints,” CIRP Annals - Manufacturing Technology, vol. algorithms in lot sizing”. J Intell Manuf, vol 21, pp575–590, 2010.
51, no. 1, pp. 387–390, 2002. [18] Gao N, and Weida C.“A Genetic Algorithm for Disassembly Scheduling
[7] Y. Barba-Gutiérrez, B. Adenso-Díaz, and S. M. Gupta, “Lot sizing in with Assembly Product Structure”. IEEE International Conference on
reverse MRP for scheduling disassembly,” International Journal of Service Operations and Logistics, and Informatics, IEEE/SOLI 2008,
Production Economics, vol. 111, no. 2, pp. 741–751, 2008. 2238–2243, 2008.
[8] D.-H. Lee and P. Xirouchakis, “A two-stage heuristic for disassembly [19] D. E. Goldberg. Genetic Algorithms in Search, Optimization, and
scheduling with assembly product structure,” J Oper Res Soc, vol. 55, Machine Learning. Addison-Wesley, 1989.
no. 3, pp. 287–297, 2004. [20] E. Talbi, “Metaheuristic. Form design to implementation”. Published by
[9] H.-J. Kim, D.-H. Lee, and P. Xirouchakis, “Two-phase heuristic for John Wiley & Sons, Inc., Hoboken, New Jersey, 2009.
disassembly scheduling with multiple product types and parts
commonality,” International Journal of Production Research, vol. 44, no.
1, pp. 195–212, 2006.
[10] H.-J. Kim and P. Xirouchakis, “Capacitated disassembly scheduling
with random demand,” International Journal of Production Research,
vol. 48, no. 23, pp. 7177–7194, 2010.
[11] H. M. Wagner, T. M. Whitin. “Dynamic Version of the Economic Lot
Size Model,” Management Science, vol. 50, no. 12 Supplement, pp.
1770–1774, 1958.
[12] R.A. Sandbothe, G.L. Thompson. A forward algorithm for the
capacitated lot size model with stockouts. Operations Research
1990;38(3):474–86.
[13] D.Aksen, K. Altınkemer., S. Chand. “The single-item lot-sizing problem
with immediate lost sales.” European Journal of Operational Research
147(3):558–566, 2003.

Вам также может понравиться