TRENDY EKONOMIKY A MANAGEMENTU

TRENDS ECONOMICS AND MANAGEMENT

ISSN 1802-8527 (Print) / ISSN 2336-6508 (Online)

2016 25(1): 48–55

DOI: http://dx.doi.org/10.13164/trends.2016.25.48

A Suitable Artificial Intelligence Model for Inventory

Level Optimization
Tereza Šustrová

Abstract

Purpose of the article: To examine suitable methods of artificial neural networks and their application in
business operations, specifically to the supply chain management. The article discusses construction of an
artificial neural networks model that can be used to facilitate optimization of inventory level and thus improve
the ordering system and inventory management. For the data analysis from the area of wholesale trade with
connecting material is used.
Methodology/methods: Methods used in the paper consists especially of artificial neural networks and
ANN-based modelling. For data analysis and preprocessing, MS Office Excel software is used. As an
instrument for neural network forecasting MathWorks MATLAB Neural Network Tool was used. Deductive
quantitative methods for research are also used.
Scientific aim: The effort is directed at finding whether the method of prediction using artificial neural
networks is suitable as a tool for enhancing the ordering system of an enterprise. The research also focuses on
finding what architecture of the artificial neural networks model is the most suitable for subsequent prediction.
Findings: Artificial neural networks models can be used for inventory management and lot-sizing problem
successfully. A network with the TRAINGDX training function and TANSIG transfer function and
6-8-1 architecture can be considered the most suitable for artificial neural network, as it shows the best results
for subsequent prediction.
Conclusions: It can be concluded that the created model of artificial neural network can be successfully used
for predicting order size and therefore for improving the order cycle of an enterprise. Conclusions resulting
from the paper are beneficial for further research.

Keywords: lot-sizing problem, inventory management, artificial neural network

JEL Classification: C53, C45, L81

48
 Tereza Šustrová: A Suitable Artificial Intelligence Model for Inventory Level Optimization

Introduction become a single-item problem. The demand for that

item is assumed to be stationary, i.e. demand occurs
In today’s highly competitive environment characte- continuously with a constant rate. The EOQ model
rized by high consumer’s requirements for products is a continuous time model with an infinite planning
with high quality, low profit margins and short de- horizon. The optimal solution is easy to derive. The
livery times, enterprises are forced to take advan- purpose of the model is to find the optimum size of a
tage of every opportunity to optimize their business manufacturing supply which is the most economical
processes. for a manufacture. This means balancing of acquisi-
The issue of inventory management is considered tion and storage costs.
to be one of the most important functions of Since the assumptions appear to be very
manufacturing and business enterprises, which often restrictive and do not correspond to the practice in
have a big impact on their overall performance. a vast majority of cases, more sophisticated models
Usually it is necessary to consider a trade-off were developed in the latter half of the 20th century.
between high holding costs on one hand and poor First to mention is the economic lot scheduling
service and high shortage costs resulting from low problem (ELSP) where capacity restrictions come
inventory levels on the other. The required solution in (Rogers, 1958 and Elmaghraby, 1978). Because
lies in a suitable inventory management which scarce resources are usually shared in common by
ensures a satisfactory service level without keeping several items, the ELSP is a single-level, multi-
unnecessarily large stocks causing high expenses. item problem. However, the ELSP still assumes
The paper presents the effectivity inquiry of stationary demand. It is a continuous time model too
artificial neural network modelling and forecasting and the planning horizon is infinite again. Solving
in  the issues of inventory management, especially the ELSP optimally is NP-hard. Hence, heuristics
in the lot-sizing problem. Several types of neural dominate the area (Zipkin, 1991).
networks are created and tested within the research, A quite different step was made from the EOQ
searching for the most efficient neural network model assumptions towards dynamic demand
architecture. MS Office Excel was used for data conditions. Wagner, Whitin (1958) created a so
analysis, preprocessing and standardization; called Wagner-Whitin model which assumes a finite
MathWorks MATLAB Neural Network Tool was planning horizon which is subdivided into several
used for neural network forecasting. discrete periods. Demand is given per period and
may vary over time. However capacity limits are
1. I nventory management models and not considered which means that the single-level
artificial neural networks WW problem is a single-item problem. The solution
is simple (Aggarwal et al., 1993). Next model
The purpose of the inventory management models generations combined models with different capacity
is to provide an appropriate method for inventory limitations and different dynamic approaches.
regulation, mostly with regard to costs. So far no The delivery size specification was extended with
universal model for inventory management has been purchase planning (Drexl et al., 1997).
developed and a specific solution must be sought in A separate, later developed model type is the
each individual situation on the basis of the existing model with deteriorating goods. Deteriorating refers
models. to the damage, spoilage, dryness, vaporization, etc.
EOQ model is frequently modified for various of the products. These models are also based on the
types of the required input data entry, such as in classical EOQ model or the Wagner-Whitin model,
the case when rebate (a discount depending on subsequently modified for the required conditions.
the volume of the order for delivery) needs to be The basic typology is again derived from the demand
respected. nature and includes models with deterministic demand,
The first mathematical model for delivery size time-varying demand, stock-dependent demand or
specification was the model developed by Ford W. price-dependent demand and also with stochastic
Harris called Economic Order Quantity (Harris, demand. Special models are developed for the case of
1913). This model is one of the single-product deter- permissible delay in payment, including delay interest
ministic dynamic models and is the simplest model payment, purchase price discount and models including
type – a model of periodically replenished inventory inflation and time value of money or inventory model
with constant delivery size. The assumption for the with two warehouses (Goyal, Giri, 2001).
EOQ model is a single-level production process with The study of various model situations, focus on
no capacity constraints, which makes the problem simulating conditions, searching for outcomes,

49
Tereza Šustrová: A Suitable Artificial Intelligence Model for Inventory Level Optimization

optimal solutions, and so on are among the most comes with a lower prediction error than other
important current trends. For example, McNelis methods.
in his paper (McNelis, 2003) applies the neural Research introduced by Hachicha (2011) also
network methodology to inflation forecasting in uses artificial neural network and deals with the
the Euro-area and the USA. There are many other lot-sizing problem in supply chain by application of
research methods besides modelling, such as metamodelling simulation. The supply chain which
laboratory experiment (Gammisch, Balina, 2015), is a subject of research operates in make-to-order
case study (Rudiene, 2014) etc. environment (no possibility of stock keeping and
The modeling of phenomena which arise from limited production capacity) and is characterized
economic reality and are described by statistical data by multi-product, multi stage, multi-location
is possible thanks to methods based predominantly production planning with capacity constraints and
on branches of mathematics such as statistics, stochastic parameters such as lot arrivals order,
numerical methods, operational research, linear and transit time, set-up time, processing time, etc. The
dynamic programming, optimization, etc. (see, for results confirm the effectiveness, flexibility and
example, Novotná (2012) or Novotná (2014)). usability of the artificial neural network method in
The study of various model situations, focus on practical applications.
simulating conditions, searching for outcomes, Another artificial neural network-based model
optimal solutions, and so on are among the most was developed by Paul, Azaeem (2011) which
important current trends. The paper Ioana et al. determines the optimum level of finished goods
(2010) presents a new concept for fuzzy logic in inventory as a function of product demand, setup,
economic processes. In his paper Zhang (2012), holding, and material costs. The model was tested
investigated the sensitivity of estimated technical with a manufacturing industry data and the results
efficiency scores by different methods including the indicated that the model can be used to forecast
stochastic distance function frontier. finished goods inventory level in response to the
Artificial neural network models have been used model parameters. Overall, the model can be applied
to successfully solve the demand forecasting and for optimization of finished goods inventory for any
production scheduling problem. In their paper, manufacturing enterprise.
Gaafar, Choueiki (2000) applied a neural network A fast convergent BP neural network model for
model to a lot-sizing problem which is part of predicting inventory level is described in a paper by
Material Requirements Planning (MRP) for the case He (2013). The paper also applies the improved BP
of deterministic time-varying demand over a fixed neural network model to predict the inventory level
planning horizon. of an automotive parts company. The results show
Paper presented by Megala, Jawahar (2006) that the improved algorithm not only significantly
addresses a dynamic lot sizing problem with exceeds the standard algorithm but also outperforms
capacity constraint and discount price structure. some other improved BP algorithms both on
Two meta-heuristics, genetic algorithm and Hopfield convergence rate and prediction accuracy.
neural network are designed for dynamic lot sizing We can summarize that soft computing methods
problem. A computational study in paper shows that can be successfully used in the area of inventory
genetic algorithm is capable of producing adequate management and optimization. Developed models
result whereas Hopfield neural network produces often showed robustness, high performance and gave
satisfactory results only for small size problems. a valid solution to the presented problem. Research
Aburto, Weber (2007) in their research presented a of artificial neural networks confirmed their usability
hybrid intelligent system combining Autoregressive and effectiveness in practical application especially
Integrated Moving Average models and neural for solving the lot-sizing problem in  production
networks for demand forecasting. Study showed planning systems. Now we can concentrate on
improvements in forecasting accuracy and proposed development of a soft computing methods-based
a replenishment system for a Chilean supermarket, model for solving the lot-sizing problem in business
which leads simultaneously to fewer sales failures venture.
and lower inventory levels.
Artificial neural network structure was also used
and compared with traditional statistical methods by 2. Artificial neural networks
Hamzaçebi (2008). The results from the modeled
artificial neural network utilized for time-series Artificial neural networks are a branch of artificial
forecasting proved that the proposed ANN model intelligence. Multi-layer perceptron form one type

50
 Tereza Šustrová: A Suitable Artificial Intelligence Model for Inventory Level Optimization

of artificial neural network. We know also Single-la- ANN model which can be used to optimize inven-
yer perceptron ANN, Radial basis function networks tory level and thus improve inventory management
or the whole group of Recurrent/feedback networks and order system of an enterprise. The variables of
(Gardner et al., 1998). current demand, demand in the next 3 months, de-
Artificial neural networks represent an analogy mand in 3 months following after the 3-month or-
to human thinking, a simplified imitation of the der cycle (3-month delay), current inventory level,
principle of human brain. They are described as the purchase prices and transport costs are used as input
black box with unknown inner structure and known parameters. Output data is the ordered quantity.
outputs. Artificial neural networks usually work in All input, output and sample data were provided
two stages. In the first stage the network acts as an by an existing company, a wholesale dealer with
inexperienced man, learning to set its parameters connecting materials. Being part of a supply chain,
to correspond to the required network topology. In the company buys from an Asian supplier and has
the second stage the network already independently to allow for a replenishment lead time of up to
transforms inputs to outputs on the basis of 60 days from the date of order. The data represents
knowledge obtained in the first stage. Each neural the information about the highest-turnover type of
network consists of defined layers, input layers, goods on the company’s assortment, as monitored
hidden layers and output layers, and the learning from January 2009 to December 2014.
method and the knowledge acquisition process. This study develops the model of neural network
Neural networks are appropriate in situations to determine the optimum amount of ordered goods
when chance plays a significant role in the process to optimize the current inventory amount. The
and deterministic relations are so complex and selection of input and output parameters is crucial
interwoven that they cannot be analysed and for the model construction. The input variables
identified. They are therefore used for modelling of for this model are the 6 most significant factors
complex and irreversible strategic decisions where affecting decisions on ordered amounts. These
a network with more hidden layers is appropriate are the current demand (monthly), demand in the
(Dostál, 2008, p. 42). next 3 months, demand in the following 3 months,
Some authors have recently focused a great part of current warehouse value, purchase price per unit and
their attention on the so called Neural-fuzzy systems transport costs per unit. The output parameter of the
combining the techniques of fuzzy modelling and model is optimal amount of purchased goods. Inputs
neural networks. Neural-fuzzy models combine and outputs are presented in Figure 1.
the advantages of fuzzy logic and neural networks, The neural network model has been constructed
namely adaptability, quick convergence and high using the MathWorks MATLAB programme. The
accuracy (Shie-Jue, Chen-Sen, 2003). following steps need to be taken when constructing
the model of neural network: collection of input,
output and sample dataset, designing, training and
3. Proposed Artificial neural network model validation of neural network.
Before using collected dataset with neural
The aim of the research is the use of artificial neu- network, data must be processed into input-output
ral network methods in designing an order cycle of patterns. Each pattern is formed with an input vector
an enterprise. This paper deals with developing an and the corresponding target vector . Input and

Figure 1. Diagram of the inputs and output of constructed neural network. Source: Author’s own study.

51
Tereza Šustrová: A Suitable Artificial Intelligence Model for Inventory Level Optimization

Table 1. Example of standardized input-output dataset.

IDact 0.44 0.41 0.67 0.67 0.66 0.65 0.32

ID3 0.27 0.09 0.45 0.50 0.89 0.31 0.48
ID33 0.13 0.65 0.06 0.61 0.61 0.72 0.64
II 0.15 0.01 0.01 0.08 0.10 0.09 0.21
IP 0.72 0.30 0.41 0.59 0.43 0.63 0.60
IT 0.24 0.75 0.43 0.63 0.24 0.64 0.62
OQ 0.043 0.171 0.336 0.426 0.674 0.997 0.342
Source: Author’s own study.

output data must be transformed into standardized visualize and simulate neural networks; it can be
form within the range of 0 to 1 in order to achieve used for applications such as data fitting, pattern
consistent result while using ANN. recognition, clustering, time-series prediction, and
The standardization is done using the following dynamic system modelling and control.
equation: A feed-forward backpropagation network model
xi − min ( xi ) with one hidden layer and output layer is used,
Xi = , (1) while the number of hidden neurons can be defined.
max ( xi ) − min ( xi ) As training function was used both TRAINGDX
where is the ith element of the input or target vector and TRAINSCG, adaptation learning function was
and Xi is corresponding standardized value within used LEARNGDM, also TANSIG transfer function
0  and 1. An example of standardised input-output is used. The maximum number of training epochs
vectors is shown in Table 1. is set to be 1  000 and the maximum number of
A total of 31 input-output dataset pairs were validation checks is set to 1 000. The architecture of
collected. This dataset was divided randomly into proposed ANN model is shown in Figure 2.
two categories – training dataset which consists The neural network training is performed using
of 77% of the data and test dataset with 23% of 24 data samples containing both input and output
the data. There are 24 training patterns, which are data. The coefficient of determination value (R2)
used further to model the artificial neural network. for training phase is calculated to evaluate the
After completion of the network training the model performance of network during the training. Values
is tested with 6 data samples and the network of the R2 indicator are between 0 and 1, the objective
performance is evaluated. The sample of input and being to determine the accuracy of the predicted
output dataset is shown in Table 1. data. Therefore the closer the R2 coefficient is to
1, the more accurate the neural network can be
considered.
3.1 Neural Network Architecture After reaching the best possible R2 coefficient
MathWorks MATLAB Neural Network Toolbox value the testing of newly emerged trained neural
was used for construction of the inventory manage- network is performed. The sample of 7 data vectors
ment model which provides functions and apps for containing both input and output data is used for
modeling complex nonlinear systems that are not testing; however, for the neural network only the
easily modeled with a closed form equation. With input vector is used. The neural network testing
neural network toolbox it is possible to design, train, output is later compared to the real output value

Figure 2. ANN architecture of proposed model. Source: Author’s own study.

52
 Tereza Šustrová: A Suitable Artificial Intelligence Model for Inventory Level Optimization

– i.e. with amount ordered in reality. The testing riable too. The evaluation of optimal neural network
performance is evaluated by an error measure – was governed by the MSE value. Neural networks
Mean squared error (MSE). MSE is the average with lower MSE value and at the same time with
squared difference between outputs and targets. R2 value closer to 1 were considered more suitable.
Lower values are better, zero means no error. The Table 2 shows R2 and MSE values of the
constructed neural network models. It shows the
results of network training and testing with the
4. Results and Discussion TANSIG and LOGSIG transfer function and with
TRAINGDX and TRAINSCG training functions
To evaluate a suitable artificial neural network for for the respective values of hidden neuron numbers
inventory level optimization and order cycle man- used for ANN. The overview indicates that neural
agement. To find the best constructed artificial ne- networks using the TRAINGDX training function
ural network with optimal architecture, 16 neural and the TANSIG transfer function generally
networks with the TRAINGDX or TRAINSCG performed better; with higher number of neurons
learning function and with the TANSIG or LOGSIG the network performance tends to decrease. The
transfer function were constructed, trained and tes- lowest MSE values (0.029) and at the same time
ted. The number of neurons in hidden layer was va- the highest R2 values (0.99) are demonstrated by

Table 2. R2 and MSE values of NN training and testing.

TRAINGDX TRAINSCG
TANSIG
Neurons 8 10 12 15 8 10 12 15
R2 0.990 0.920 0.970 0.910 0.980 0.950 0.920 0.950
MSE 0.023 0.030 0.029 0.075 0.050 0.040 0.032 0.035
LOGSIG
R2 0.900 0.960 0.940 0.900 0.840 0.970 0.990 0.890
MSE 0.033 0.030 0.044 0.036 0.053 0.032 0.034 0.096
Source: Author’s own study.

Table 3. Summary of optimal artificial neural network model.

Input data 6 – Actual month demand

– Demand in next 3 months
– Demand in 3 months following next 3 months
– Actual stock level
– Unit product costs
– Unit transport costs
Output data 1 – Quantity
Network type Feed-forward backpropagation
Training function TRAINGDX
Adaptation learning function LEARNGDM
Performance fuction MSE
Number of hidden layers 1
Neurons in hidden layers 8
Transfer function TANSIG

Training Parameters
Maximum number of training epochs (iterations) 1,000
Maximum training time Infinite
Maximum performance value 0
Minimum gradient magnitude 0.00001
Learning rate 0.01
Source: Author’s own study.

53
Tereza Šustrová: A Suitable Artificial Intelligence Model for Inventory Level Optimization

the neural network with 8 hidden neurons; on the demand in the next 3 months, demand in 3 months
contrary the largest error (0.075) and the lowest R2 following after the 3-month order cycle (3-month
value (0.91) is demonstrated by the neural network delay), current warehouse values, purchase price
with 15 hidden neurons. and transport costs.
The absolutely best performance was shown The architecture with feed-forward  back
by the neural network with 8 hidden neurons, propagation learning algorithm, TRAINGDX
TRAINGDX training function and TANSIG transfer training function and TANSIG transfer function
function. Therefore, the neural network with the and 6-8-1 architecture was assessed as the most
6-8-1 architecture has been chosen as the optimal suitable neural network. A performance of neural
model for subsequent prediction and for designing network model was evaluated by Coefficient of
the ordering system of a company. Determination (R2) and Mean Squared Error.
Table 3 presents a summary of the proposed The constructed model of artificial neural network
network architecture, consisting of  6 input neuron can be used for further order cycle optimization.
(inputs), 8 neurons in hidden layer and one output The future order amount can be planned based
neuron (target). The R2 and MSE indicators were on predicated demand and thus the inventory
used to find the optimal network architecture; these management can be improved as a part of supply
indicators were used for evaluation of each neural chain management. The article presents artificial
network. neural networks as very useful and bringing many
opportunities for further research.

5. Conclusion
Acknowledgment
The article investigates the options of using artificial
neural networks and their applications in a compa- This paper was supported by grant FP-S-15-2787
ny ordering system. Within the research a model of “Effective Use of ICT and Quantitative Methods for
artificial neural network was developed to optimize Business Processes Optimization” from the Internal
the order amount as a function of current demand, Grant Agency at Brno University of Technology.

References

Aburto, L., Weber, R. (2007). Improved supply chain Gardner, M. W., Dorling S. R. (1998) Artificial neural
management based on hybrid demand forecasts. Applied networks (the multilayer perceptron) – a review of
Soft Computing, (7)1, pp. 136–144. applications in the atmospheric sciences.  Atmospheric
Aggarwal, A., Park, J. K., Krarup, J. (1993). Improved environment, (32)14, pp. 2627–2636.
Algorithms for Economic Lot Size Problems: An 0 (mn)- Goyal, S. K., Giri, B. C. (2001). Recent trends in
Algorithm for Structured Problems. Operations Research, modeling of deteriorating inventory. European Journal of
41(3), pp. 155–180. Operational Research, 134(1), pp. 1–16.
Dostál, P. (2008). Pokročilé metody analýz a modelování Hachicha, W. (2011). A simulation metamodelling based
v podnikatelství a veřejné správě. Brno: Akademické neural networks for lot-sizing problem in MTO sector.
nakladatelství CERM. 342pp. International Journal of Simulation Modelling, 10(4), pp.
Drexl, A., Kimms, A., Seeanner, F. (1997). Lot sizing and 191–203.
scheduling — Survey and extensions. European Journal of Hamzaçebi, C. (2008). Improving artificial neural
Operational Research, 99(2), pp. 143–154. networks’ performance in seasonal time series
Elmaghraby, S. (1978). The Economic Lot Scheduling forecasting.  Information Sciences,  (178)23, pp.  4550–
Problem (ELSP): Review and Extensions. Management 4559.
Science, 24(6), pp. 587–598. Harris, F. W. (1913). How Many Parts to Make at Once.
Gaafar, L., Choueiki, M. H. (2000). A neural network In: Factory, The Magazine of Management, 10(2), pp.
model for solving the lot-sizing problem. Omega, 28(2), 135–152.
pp. 175–184. He, W. (2013). An inventory controlled supply chain model
Gammisch, M., Balina, S. (2015). Support of Risk Management based on improved BP neural network. Discrete Dynamics
through Careful Selection of Anti-corruption Measures–a Critical in Nature and Society. 2013(1), pp. 1–7.
Evaluation of Compliance Management Activities.  Trends Ioana, A., Mirea, V., Bălescu C. (2010). Economic
Economics and Management, 9(22), pp. 9–18. processes study through Fuzzy Logic. Journal of Economic

54
 Tereza Šustrová: A Suitable Artificial Intelligence Model for Inventory Level Optimization

Computation And Economic Cybernetics Studies And Computations, (2)2, pp. 431–438.
Research, 44(2), pp. 129–137. Rogers, J. (1958). A Computational Approach to the
McNelis, P. D. (2003). Nonlinear phillips curves in the Economic Lot Scheduling Problem. Management Science,
Euro area and USA? Evidence from linear and neural 4(3), pp. 264–291.
network models. Computational Intelligence for Financial Rudiene, E. (2014). Lithuanian Retail Internationalization:
Engineering, 2003(1), pp. 145–149. Features and Prospects.  Trends Economics and
Megala, N., Jawahar, N. (2006). Genetic algorithm and Management, 8(20), pp. 58–67.
Hopfield neural network for a dynamic lot sizing problem. Shie-Jue, L., Chen-Sen, O. (2003). A neuro-fuzzy system
The International Journal of Advanced Manufacturing modeling with self-constructing rule generation and hybrid
Technology, 27(11–12), pp. 1178–1191. SVD-based learning. IEEE Transactions on Fuzzy Systems,
Novotná, V. (2012). The impact of the inflation and 11(3), pp. 341–353.
unemployment values from the previous period on the Wagner, H. M., Whitin, T. M. (1958). Dynamic Version of
Phillips curve. 7th International Scientific Conference the Economic Lot Size Model. Management Science, 5(1),
Business and Management, 2012(1), pp. 147–158. pp. 89–96.
Novotná, V. (2014). Feedback Influence on the Behavior of Zhang, T. (2012). A Monte Carlo Analysis for Stochastic
Financial System. Economic Computation and Economic Distance Function Frontier. Inzinerine Ekonomika-
Cybernetics Studies and Research, 2014(1), pp. 217–231. Engineering Economics, 23(3), pp. 250–255.
Paul, S. K., Azaeem, A. (2011). An artificial neural Zipkin, P. H. (1991). Computing Optimal Lot Sizes in the
network model for optimization of finished goods Economic Lot Scheduling Problem. Operations Research,
inventory. International Journal of Industrial Engineering 39(1), pp. 56–63.

Received: 18. 1. 2016 Ing. Tereza Šustrová

Reviewed: 2. 2. 2016 Brno University of Technology
Accepted: 20. 5. 2016 Faculty of Business and Management
Department of Informatics
Kolejní 2906/4, 612 00 Brno
Czech Republic
Tel.: +420 54114 3709
E-mail: [email protected]

55