A SEMINAR REPORT
submitted to
SARATH BABY
BACHELOR OF TECHNOLOGY
in
MECHANICAL ENGINEERING
NOVEMBER 2018
1
HEURISTIC APPROACHES TO DETERMINE BASE-
STOCK LEVELS IN A SERIAL SUPPLY CHAIN WITH A
SINGLE OBJECTIVE AND WITH MULTIPLE
OBJECTIVES
A SEMINAR REPORT
submitted to
by
SARATH BABY
BACHELOR OF TECHNOLOGY
in
MECHANICAL ENGINEERING
NOVEMBER 2018
2
ACKNOWLEDGEMENT
First and foremost, I thank Almighty God for bestowing me with his profound grace,
without which this seminar could not have been a success.
I sincerely thank the staff in charge of seminar for most valuable advices and support.
The encouragement given by my seminar guide helped me through the course of seminar.
Finally I express my heartfelt gratitude to my parents and my friends for the mental
support they had offered me and the immense encouragement to the successful completion
of the seminar.
3
CERTIFICATE
4
ABSTRACT
This paper deals with the problem of determination of installation base-stock levels
in a serial supply chain. The problem is treated first as a single-objective inventory-cost
optimization problem, and subsequently as a multi-objective optimization problem by
considering two cost components, namely, holding costs and shortage costs. Variants of
genetic algorithms are proposed to determine the best base-stock levels in the single-
objective case. All variants, especially random-key gene-wise genetic algorithm
(RKGGA), show an excellent performance, in terms of convergence to the best base-stock
levels across a variety of supply chain settings, with minimum computational effort.
Heuristics to obtain base-stock levels are proposed, and heuristic solutions are introduced
in the initial population of the RKGGA to expedite the convergence of the genetic search
process. To deal with the multi-objective supply-chain inventory optimization problem, a
simple multi-objective genetic algorithm is proposed to obtain a set of non-dominated
solutions.
5
CONTENTS
Title Page
ACKNOWLEDGEMENTS............................................................................................i
ABSTRACT....................................................................................................................ii
LIST OF TABLES..........................................................................................................iii
LIST OF FIGURES....................................................................................................... iv
ABBREVIATIONS..........................................................................................................v
NOMENCLATURE .......................................................................................................vi
CHAPTER 1 : INTRODUCTION…………………………………………………... 11
CHAPTER 6 : SUMMARY………………………………………………………….. 25
REFERENCE…………………………………………………………………………. 26
6
LIST OF TABLES
1 Different supply chain cost rate and lead time settings ......................................................
18
7
LIST OF FIGURES
4 Chromosome representation............................................................................... 14
5 Crossover operator.............................................................................................. 15
8
NOMENCLATURE
English Symbols
j Stage index
j,t
B Installation backlog at stage j , at the end of time t
j+1,j
TR Transportation cost between j+1 and j
Fk Fitness constant
u Random number
9
ABBREVIATIONS
GA Genetic algorithm
MR Mutation rate
CE Computational effort
CR Crossover rate
10
CHAPTER 1
INTRODUCTION
It is observed from the literature that the solution approaches developed to optimize
base-stock levels in supply chain include heuristic algorithms and iterative search
procedures. Determination of base-stock levels in a supply chain for every stage and the
minimization of the total supply chain cost are quite complex and computationally tedious.
Motivated by such observations and findings, the current study aims at developing
efficient simulation-based algorithms (such as genetic algorithms, gas) to determine the
best local base-stock levels in a serial supply chain.
11
CHAPTER 2
The supply chain model envisaged in the present study is a single-product, four-stage
serial supply chain consisting of a retailer, distributor, manufacturer and supplier. All the
members in the supply chain operate under a periodic-review base-stock policy, where the
review period is unit time (or a day). Order cost is assumed to be zero or negligible. The
retailer faces random customer demand. Lead time for information (or order processing) is
zero or negligible. Processing lead time and transportation lead time are combined
accordingly at each stage and considered together as one component, called installation or
local replenishment lead time for that stage. Base-stock level at every member or
installation in the supply chain takes discrete integer values. All installations have infinite
capacity. The source of supply of raw materials to the most upstream member, namely the
supplier, is assumed to have infinite raw material availability. Even though the
replenishment lead time is assumed to be deterministic, there exists supply uncertainty in
the sense that the customer or any member will be replenished, provided that the
immediate upstream member has sufficient on-hand inventory to meet the replenishment
order quantity; otherwise, a replenishment order is passed to the next upstream member
who in turn will satisfy the demand depending on the on-hand inventory. The
replenishment lead time for any member can vary between the respective replenishment
lead time and the maximum replenishment lead time that corresponds to the sum of all
replenishment lead times from that member up to all its upstream members in the supply
chain.
The supply chain model envisaged in this work is simulated to evaluate the
performance of the supply chain with respect to every base-stock policy (generated
through the various solution approaches considered in this study), with the total supply
chain cost (TSCC) as the measure of performance in the case of single objective supply
chain inventory problem. The supply chain is simulated for random customer demands
12
that are generated from different demand distributions (i.e., Uniform, Normal and Poisson)
for a specified long run length over which the statistic, TSCC, is collected.
TSCC across all members in the supply chain over T (a long run) is given as follows:
Figure 1
The objective is to obtain optimal installation base-stock policy, that minimizes the TSCC.
It is to be noted that in the current study, we consider the installation base-stock levels and
hence we define the total cost function accordingly. Since we allow backlog of unsatisfied
demand at every stage, the last two terms do not influence sj
Figure 2
We term the above TSCC as TSCC-I in this study. For the case where b j exists at every
stage, we have the TSCC, called TSCC-II in this study, as follows:
Figure 3
13
CHAPTER 3
Figure 4
A set of installation base-stock levels sj are randomly generated within the range [sjUL ;
sJLL] for j = 1 to n, to form a chromosome. Thus n chromosomes are generated, constituting
the parent population, called par_pop. . Chromosomes in par_pop are evaluated through
simulation, and their respective objective function values (TSCC-1 or TSCC-2) are
14
obtained. For the sake of generality, let TSCC k correspond to objective-function value of
chromosome k. Fitness value fk is computed for the kth chromosome by setting, fk = 1/(1 +
TSCCk). Based on fk values, the chromosomes are selected probabilistically (i.e., roulette-
wheel selection) to the mating pool. The size of the mating pool is same as par_pop. A
single-point crossover operator is deployed.
Figure 5
The first two chromosomes in the mating pool are selected for crossover with a crossover
rate (i.e., probability of crossover) CR, by sampling a uniform random number u. These
chromosomes are directly placed in int_pop (an intermediate population), if u is >CR; else
the two chromosomes are subjected to crossover.
Offspring are placed in int_pop. Repeat the possible crossover operation on the rest of
the chromosomes in the mating pool by considering two consecutive chromosomes at a
time.
The mutation operator used in this study is a gene-wise mutation operator, with a gene
representing an installation base-stock level. Genes of offspring in the int_pop are mutated
with a probability of MR.
15
Thus the generated offspring inherits four genes from the chromosomes in the
par_pop, based on chromosomes fitness values. Generate offspring in this manner until n
offspring are generated. It is to be noted that in this variant of GA with gene-wise
crossover operator, we do not generate a mating pool nor have we used the crossover rate.
We generate n offspring by constructing each offspring directly from par_pop through
gene-wise crossover operator. But for this mode of generation of offspring, all other steps
of the BGA hold for Variant 2
The genes for chromosomes are represented as random keys (or uniform random
numbers), instead of phenotypes, in RKGA. Random-key representation as a solution
representation was introduced by Bean (1994). A chromosome is represented as random
keys as follows:
Figure 6
where uj is the uniform random number or the random key that stands for the scaled base-
stock level at the jth member in a supply chain. The initial population is created by
generating uniform random numbers in the interval (0, 1), and filling up all the
chromosomes in par_pop. To evaluate the goodness of the chromosomes, chromosomes
are decoded to obtain the installation base-stock levels.
The crossover and mutation operations used in Variant 3 are the same as in the BGA,
but these operators operate on random keys. Every resultant chromosome in int_pop is
then decoded to get the base-stock levels at different stages (or installations). All other
steps in Variant 3 are the same as in BGA.
16
changes, this variant is same the BGA. Note that Variant 4, just like Variant 2, does not
create a mating pool.
CHAPTER 4
A supply-chain setting consists of a cost-rate setting and a lead time setting (see Table
1). For example, the cost-rate setting CS1 and the lead time setting LT1 constitute one
supply chain setting. Two measures of performance, namely TSCC-1 and TSCC-2, are
considered separately. Accordingly, six supply chain settings are studied with respect to
TSCC-1, and six more supply chain settings are studied with the consideration of TSCC-2.
We have considered different holding cost-rates, shortage cost-rates and lead times in
order to check for the robustness of the proposed GAs in terms of their consistent and
good performance across various supply chain settings and scenario.
The run length for the simulation experiment is 1400 days, and each simulation
experiment is replicated 30 times by using 15 pairs of random numbers and antithetic
random numbers to sample customer demand. This method of antithetic sampling is a
commonly used procedure for generating negatively correlated pairs of samples (Deo,
1999). The customer demand is uniformly distributed in the range [20, 60] units per unit
time (i.e., per day). Customer demand is sampled from a uniform distribution because,
given the minimum and maximum values that a random variable can take, the variance is
maximum when the random variable follows the uniform distribution.
17
Table 1
The measure of performance is observed over a period of 1200 within the run length of
1400 days, by leaving out the first 100 and the last 100 days. At the end of 30 replications,
the mean and standard deviation of the measures of performance are computed. The
supply chain is simulated by initializing the on-hand inventory at the members equal to
their respective base-stock levels. A pilot study for determining the best parameter settings
in the proposed GAs has been carried out, and the best {CR,MR} settings are obtained for
the variants of GA with respect to TSCC-1 and TSCC-2. Having chosen the best
{CR,MR} settings for GAs (in the neighborhood of 0.85 and 0.15 respectively)
18
installation base-stock levels at different installations with respect to TSCC-1 and TSCC-2
for every supply chain setting (termed CEn.).
Table 2
19
variants of GAs have performed very well, and are capable of generating near-optimal
base-stock levels in all the supply chain settings, with the average percentage deviation of
TSCC from the respective optimal solution being very small. It is seen from Table 2 that
the percentage THC for CS1 + LT1 is 77.85%, whereas for CS2 + LT1 it is 86.69%, and
for setting CS3 + LT1 it is 88.14%. This holding-cost increase from one cost-rate setting to
another is mainly due to the increase in the shortage cost-rate across different supply chain
settings (see Table 1).
To get more insight from the data presented in Tables 3, the computational effort
values are normalized by dividing the computational effort required by every solution
procedure for a supply chain test problem by the minimum computational effort required
for that supply chain test problem. An average normalized computational effort (ANCE) is
calculated for each of the solution procedures. It is evident that the proposed GAs,
especially with the random-key representation and the gene-wise crossover operator,
require less computational effort.
20
parameters (mean and standard deviation) of the normal distribution are assumed to be 40
units per unit time and 7 units per unit time respectively.
Table 3
Following the earlier supply chain settings and simulation-based analyses, the
performance of RKGGA is found to be very good in this case as well, in terms of
obtaining near optimal base-stock levels. Next, by assuming the demand to follow Poisson
distribution, with the mean demand rate assumed to be 40 units per unit time, the
performance of RKGGA is evaluated. The performance of RKGGA is found to be very
good in this case as well.
21
The performance of RKGGA is analyzed with an alternative form of the problem by
assuming random replenishment lead times. In this phase of work we obtain the optimal
base-stock policy in the class of installation base-stock policies by using RKGGA with the
consideration of random replenishment lead times, although the base-stock policy need not
be necessarily be optimal. It is found that the RKGGA is able to converge to the best base-
stock policy in the class of base stock policies with minimal computational effort.
The performance of RKGGA (in these analyses) reinforces the fact that the RKGGA is
very robust, and is not sensitive to distributions, demand variance and alternative forms of
supply chain problems. In all the analyses RKGGA is able to unearth solutions with very
good quality, with the percentage deviation from the optimal solution being very marginal.
22
CHAPTER 5
Most real-life optimization problems involve multiple and conflicting objectives, and
different solutions may produce trade-offs among different objectives. There exist a set of
solutions which are Pareto-optimal, and at least one such solution dominates a solution not
contained in this set of non-dominated solutions. Since a number of solutions can be
Pareto-optimal, the task of multi-objective optimization is to find as many such (trade-off)
non-dominated solutions as possible, and this task is quite complex. The single-objective
problem considered so far (in terms of minimizing the TSCC) aims at finding the best
base-stock levels at various stages in a serial supply chain.
This research setting has served as the motivation to develop a MOGA-SCIP for
generating a set of non-dominated solutions (involving trade-off between total holding
cost and total shortage cost in the supply chain). Many multi-objective GAs have been
developed for optimizing continuous systems, and most existing multi-objective GAs are
quite complicated and they make use of many parameters.
Step 1: Generate 199 solutions with their installation base-stock levels set randomly
between the GAs base-stock level and the upper bound on the respective base-stock level
at every stage; the solution generated by the GA is also considered, and these 200
solutions together constitute the initial set of solutions (i.e., initial population).
Step 2: Evaluate all the chromosomes in the initial population via simulation and obtain
the values of their respective THC and TSC.
23
Step 5: Select one chromosome at random from the non-dominated set of solutions. As for
the second chromosome, sample u. If u 6 P, then choose a chromosome at random from the
non-dominated front; else generate randomly a solution within the respective bounds on
base-stock levels at different stages and compute its THC and TSC.
Step 6: Subject the two chromosomes obtained in Step 5 to gene-wise crossover operation.
Form one offspring by making use of fitness values computed with the consideration of
THC, and form another offspring by making use of the fitness values with the
consideration of TSC.
Step 7: Generate five neighborhood solutions from each of the two offspring.
Step 8: Evaluate the two offspring and the 10 neighborhood solutions by using simulation
and obtain the values of THC and TSC.
Step 9: Form a set of non-dominated solutions, out of the two offspring and 10
neighborhood solutions.
Step 10: Combine the non-dominated set obtained from Step 9 and the current set of non-
dominated solutions to form the net or updated non-dominated set.
Step 11: If T = 100, then go to Step 12; else increment T = T + 1 and return to Step 5.
Step 12: If P = 1.0, go to Step 13; else increment P by 0.10, set T = 1, and return to Step 5.
Step 13: Stop. The final non-dominated set of solutions is the solution set.
24
CHAPTER 6
SUMMARY
The present work has considered the problem of determination of base-stock levels in
a serial supply chain so as to minimize the TSCC (i.e., TSCC-1 or TSCC-2). Three new
variants of GAs are proposed, namely, GGA, RKGA and RKGGA. The proposed GAs are
evaluated with a variety of supply chain settings, and it is found that the proposed
RKGGA outperforms all other solution procedures. To expedite the convergence of the
genetic search process, problem-specific heuristic solutions are introduced in the initial
population. The performance analysis shows that the convergence of the genetic search
process has improved considerably after introducing the proposed heuristic solutions in the
initial population. More importantly, a multi-objective genetic algorithm has been
proposed to obtain non-dominated solutions in respect of total holding costs and total
shortage costs in a serial supply chain.
25
REFERENCES
Bean, J.C., 1994. Genetic algorithms and random keys for sequencing and optimization.
ORSA Journal on Computing 6, 154–160.
Clark, A.J., 1960. The use of simulation to evaluate a multiechelon dynamic inventory
model. Naval Research Logistics Quart 7, 429– 446. Clark, A., Scarf, H., 1960. Optimal
policies for a multi-echelon inventory problem. Management Science 6, 474–490.
Daniel, J.S.R., Rajendran, C., 2005. A simulation-based genetic algorithm for inventory
optimization in a serial supply chain. International Transactions in Operational Research
12, 101–127 (also, see Erratum 12, 479).
Deb, K., Agrawal, S., Pratap, A., Meyarivan, T., 2000. A fast elitist non-dominated
sorting genetic algorithm for multi-objective optimization: NSGA-II. In: Proceedings of
the Parallel Problem Solving from Nature VI (PPSN-VI), pp. 849–858.
Deo, N., 1999. System Simulation with Digital Computer. Prentice-Hall of India Private
Limited, New Delhi.
Ettl, M., Feign, G.E., Lin, G.Y., Yao, D.D., 2000. A supply network model with base-
stock control and service requirements. Operations Research 48, 216–232.
Federgruen, A., Zipkin, P., 1984. Computational issues in an infinite horizon, multi-
echelon inventory model. Operations Research 32, 818–836.
Gallego, G., Zipkin, P., 1999. Stock positioning and performance estimation in serial
production-transportation systems. Manufacturing & Service Operations Management 1,
77–88.
Glasserman, P., Tayur, S., 1995. A sensitivity analysis for base-stock levels in multi-
echelon production-inventory systems. Management Science 41, 263–281.
26
Lee, H.L., Billington, C., 1993. Material management in decentralized supply chains.
Operations Research 41, 835–847.
Minner, S., 1997. Dynamic programming algorithms for multi-stage safety stock
optimization. OR Spektrum 19, 261–271.
Papoulis, A., 1985. Probability, Random Variables and Stochastic Processes. McGraw-
Hill, Singapore. Petrovic, D., Roy, R.,
Rao, U., Scheller-Wolf, A., Tayur, S., 2000. Development of a rapid-response supply
chain at Caterpillar. Operations Research 48, 189–204.
27
28