Available via license: CC BY 4.0
Content may be subject to copyright.
Engineering, Technology & Applied Science Research
Vol. 14, No. 5, 2024, 16317-16325
16317
www.etasr.com Ait Ben Hamou et al.: Design of a Machine Learning-based Decision Support System for Product …
Design of a Machine Learning-based Decision
Support System for Product Scheduling on Non
Identical Parallel Machines
Khalid Ait Ben Hamou
Computer Sciences Engineering Laboratory, Faculty of Sciences, Cadi Ayyad University, Marrakech,
Morocco
khalid.aitbenhamou@ced.uca.ma (corresponding author)
Zahi Jarir
Computer Sciences Engineering Laboratory, Faculty of Sciences, Cadi Ayyad University, Marrakech,
Morocco
jarir@uca.ac.ma
Selwa Elfirdoussi
Emines - University Mohammed VI Polytechnic, Benguerir, Morocco
selwa.elfirdoussi@emines.um6p.ma
Received: 25 May 2024 | Revised: 1 July 2024 and 7 July 2024 | Accepted: 8 July 2024
Licensed under a CC-BY 4.0 license | Copyright (c) by the authors | DOI: https://doi.org/10.48084/etasr.7934
ABSTRACT
Production planning in supply chain management faces considerable challenges due to the dynamics and
unpredictability of the production environment. Decision support systems based on the evolution of
artificial intelligence can provide innovative solutions. In this paper, an approach based on machine
learning techniques to solve the problem of scheduling the production of N products on M non-identical
parallel machines is proposed. Using regression and classification models, our approach aims to predict
overall production costs and assign products to the right machines. Some experiments carried out on
simulated data sets demonstrate the relevance of the proposed approach. In particular, the XGBoost model
stands out for its superior performance compared with the other tested ML algorithms. The proposed
approach makes a significant contribution to the optimization of production scheduling, offering
significant potential for improvement in Supply Chain Management.
Keywords-decision support system; machine learning; scheduling problem; supply chain management
I. INTRODUCTION
Decision-making is a major challenge, especially in
complex or conflictual situations. The use of computers to
assist decision-makers has become indispensable when the
amount of data and the number of constraints to be managed
increase significantly. The concept of Decision Support
Systems (DSS) was developed in the 1970s [1]. The work
presented in [2] defined as computer-based technological
solutions supporting complex decision-making and problem-
solving. Since then, DSS applications have evolved
considerably. Once based on limited user interfaces and
restricted databases, they were primarily aimed at individual
decision-makers. With technological progress and increasing
digitization, modern DSS offer advanced functionalities for
groups of decision-makers, and even for virtual teams, enabling
inter-organizational collaboration [2]. With the rise of Artificial
Intelligence (AI), which is now ubiquitous, DSSs have
benefited from various areas of AI, becoming smarter and able
to process unstructured data and model complex relationships.
In Supply Chain Management (SCM), the adoption of AI-based
methods, particularly Machine Learning (ML), is providing
innovative solutions to complex decision-making problems,
particularly in the face of uncertainty and multiple constraints
[1]. However, information systems dealing with production
scheduling face considerable difficulties [2] due to the
dynamics and unpredictability of the production environment,
and the complexity of combinatorial problems. To meet these
challenges, it is essential to support production planning with
intelligent algorithms capable of analyzing and exploiting
various types of data.
Recent literature, circumscribed between 2018 and 2023,
highlights the growing use of ML to refine scheduling
processes at the heart of SCM. It shows that most research
Engineering, Technology & Applied Science Research
Vol. 14, No. 5, 2024, 16317-16325
16318
www.etasr.com Ait Ben Hamou et al.: Design of a Machine Learning-based Decision Support System for Product …
works are characterized by their heterogeneity both in terms of
methods and targeted industrial applications.
Authors in [3] conceptualized a genetic programming-based
heuristic for optimizing project scheduling with limited
resources. Their distinctive approach relies on the creation of
priority rules through genetic mutation and recombination.
Following this innovative impulse, authors in [4] explored the
use of the Deep Q Network algorithm for automated scheduling
of production tasks, eliminating the need for human
intervention. Authors in [5] took a turn towards real-time
scheduling in a smart factory, using Reinforcement Learning to
overcome the constraints of the MDRs method, demonstrating
increased flexibility in the face of variations in the production
environment. In 2019 authors in [6] presented a predictive-
reactive framework that combines ML and simulation-based
optimization to refine production planning. Authors in [7]
focused on a multi-objective flow shop scheduling problem,
developing a mathematical model to minimize various
performance indicators using opposing learning strategies and
cluster analysis. Reinforcement learning has been successfully
applied in the context of chemical production by [8] and in the
integration of genetic algorithms and ant colonies by the
authors in [9] who sought to accelerate the convergence of
scheduling solutions. Authors in [10] presented a reordering
framework, considering balancing schedule adjustments with
the accumulation of delays. Authors in [11] developed a hybrid
control solution combining Big Data techniques and ML for
reactive management of predictive production planning
systems. Authors in [12] took a turn towards automating the
solution of job-shop scheduling problems, with the use of a
graphical neural network to improve the generalization of
solutions. Authors in [13] dealt with customer order scheduling
using a two-stage optimization method, incorporating ML to
refine production schedules. Authors in [14] recommended the
use of decision trees to select priority rules, associating
performance indicators with specific rules. The application of
Neuro Evolution of Augmenting Topologies (NEAT) [15] in
complex scheduling problems has demonstrated its superiority
over traditional methods, offering a new perspective for hybrid
two-stage workshops. Authors in [16] considered the use of
ML for estimating processing times, thus optimizing the
scheduling of parallel machines. Authors in [17] integrated
predictive maintenance into production scheduling, employing
deep learning models to predict equipment service life.
Significant advances were made by the authors in [17, 18] who
used the graphical neural network for applications in job-shop
scheduling and flexible workshop planning. In addition,
authors in [20] applied Genetic Programming to flexible shop
scheduling, while authors in [19] combined ML with reasoning
about domain problems to develop efficient dispatching rules.
More recently, authors in [21] adopted a deep Q-learning
approach to solve job-shop scheduling problems, and authors in
[22] presented a Markov Decision Process model for
reinforcement learning in scheduling. Authors in [23] improved
the solution of job-shop scheduling problems by exploiting ML
techniques to predict an initial solution. Finally, authors in [24,
25] have pushed the boundaries of Deep Reinforcement
Learning, combining it with graphical neural networks to
address flexible job shop scheduling.
This review highlights the digital transformation driven by
ML, enabling more responsive, efficient, and flexible SCM,
testifying to a burgeoning field of research and promising
prospects for the future. Although previous research has
significantly contributed to the integration of ML techniques
into scheduling within supply chain management, certain gaps
remain and serve as the basis for our study. Firstly, except for
[22], the reviewed works focus on specific production
environments and do not discuss the ability to generalize
scheduling solutions to different industries or diversified
production systems. Secondly, although powerful, the models
used can present significant complexity in their configuration
and optimization, requiring advanced technical expertise.
Thirdly, the integration of predictive maintenance into
scheduling is little addressed, with only [10, 16] having
explored this aspect. These studies do not address how
scheduling and predictive maintenance can be jointly optimized
to improve both equipment durability and efficiency.
Our research aims to fill these gaps by developing a
scheduling method that dynamically adapts to variations in
production conditions and integrates predictive maintenance,
thereby increasing the accuracy and practical applicability of
scheduling solutions.
II. MATERIALS AND METHODS
A. Scheduling Poblem Analysis
Production scheduling is defined as the allocation of
available production resources over time to meet some set of
performance criteria [26]. As mentioned in [27], production
scheduling is concerned with the efficient allocation of
resources over time for the manufacture of goods. In general, it
aims to satisfy cost minimization or production maximization
objectives and to improve the efficiency of a supply chain in
terms of delivery times and production costs.
This paper deals with a scheduling problem where the
resources are non-identical parallel machines, and the tasks are
an order book containing several items to be produced. The
problem is complex in view of the number of constraints to be
satisfied. The constraints considered in this work are:
1. Delivery time constraint: An order book has a delivery
start and end date. This duration is divided into a unit of
time called a period (an hour in our case).
2. Constraint of heterogeneous machines: Not all machines
are identical; there are special machines that produce only
a few products.
3. Production cost constraint: An item can be produced on
several machines, with different costs, and in a different
number of periods. The quantity produced per hour differs
from one machine to another. The choice of machine for a
product must therefore take this constraint into account.
4. Maintenance schedule constraints: The various machines
are serviced at preventive maintenance periods known in
advance. This schedule is set by the maintenance managers
prior to the planning of product scheduling on the
Engineering, Technology & Applied Science Research
Vol. 14, No. 5, 2024, 16317-16325
16319
www.etasr.com Ait Ben Hamou et al.: Design of a Machine Learning-based Decision Support System for Product …
machines. No production will take place on these lines
during these periods.
5. Delivery window constraint for an order: The end of
production of a product must respect the delivery window.
In fact, the end of production must be before the final date
set by the customer, and must not be before the delivery
start date, to avoid stockpiling the produced items. This
storage can be costly.
The complexity of this problem lies in the heterogeneity of
resources and the non-authorization of preemption, as well as
the delivery windows imposed by customers. This classifies it
as an NP-Hard problem. The vast majority of scheduling
problems are NP-Hard and therefore cannot be formulated as
linear programs [28]. Solving this scheduling problem requires
accurate, representative data. In the context of this study, we
have opted for the use of simulation data, generated using
Python programs specifically developed for this purpose. The
following section details these data sets, covering both
structured data, data from the scheduling problems studied and
the optimal solutions generated.
B. Dataset
In the absence of real data suited to our research objectives,
we proceeded in three phases to create a dataset representative
of scheduling problems for N products on M non-identical
machines.
The first phase is devoted to generating the structural data
essential for modeling production instances. This process
includes determining whether it is possible to produce item i on
machine j, and quantifying the operational parameters if
production is feasible. For each item-machine pair where
production is possible, the program establishes the production
rate, expressed in quantity of finished products per hour, and
the hourly cost associated with this production for each
machine concerned.
In the second phase, we generated scheduling problem
instances. Each instance is defined by a specific list of
products, identified by Pi, which must be produced. For each
product, the program determines not only the quantity required
for production, but also the delivery period. In addition, the
program assigns maintenance time slots to each machine,
delimited by a start and end date, during which the machines
are not available for production.
Finally, the third phase consists of calculating the set of
possible solutions and selecting the optimum solution in terms
of minimum production cost.
Algorithm 1 illustrates the method used to calculate the
optimal solution to each instance of the scheduling problem. It
starts by establishing a scheduling table, considering machine
maintenance periods (line 2). For each possible permutation of
products (line 4), the algorithm evaluates the cost and
allocation of products to machines, rejecting unfeasible
permutations (line 10) and selecting the machine with the
lowest production cost for each product (line 12). The
permutation that results in the lowest total cost is recorded as
the optimum solution (line 19) and stored in the database (line
22).
Although the algorithm guarantees the discovery of the
optimal solution, it is mainly applicable to small and medium-
sized instances due to the increasing computational complexity
of large datasets.
Algorithm 1:
1 FUNCTION generate solutions(id_problem)
2 INIT solution with maintenance periods
3 RETRIEVE problem details for id_problem
4 FOR EACH permutation of products
5 INITIALIZE cost for permutation
6 FOR EACH product in permutation
7 FIND manufacturable machines
8 GET list of possible machines
9 IF no possible machines
10 REJECT permutation
11 ELSE
12 SELECT least cost machine
13 ASSIGN product to machine
14 UPDATE solution
15 UPDATE cost for permutation
16 END FOR
17 IF current cost < optimal cost
18 SET optimal cost to current cost
19 RECORD optimal solution &
20 RECORD permutation
21 END FOR
22 STORE optimal solution in database
23 END FUNCTION
24
25 FOR EACH problem instance
26 CALL generate_solutions (instance id)
24 END FOR
Our dataset comprises a total of 25710 instances of the
problem of scheduling 6 products on 3 machines over a 12-
period horizon, with their optimal solutions. This dataset is
stored and managed using MongoDB, a NoSQL database
management system renowned for its flexibility and
performance with large amounts of data.
C. Data Pre-Processing and Standardization
Data pre-processing is an important step in our
methodology, not least because of the heterogeneous nature of
our feature types, which come in the form of vectors, matrices
and dictionaries. To make these complex data suitable for ML
models, specific pre-processing was necessary, as summarized
in TABLE I for a scheduling problem involving p products on
m machines. The total number of features is: 4p+2m+3pm+1 in
our case, where p = 6 and m = 3, the total value is 85. To
standardize continuous characteristics, such as product
quantities and production rates, we used StandardAero from the
scikit-learn library. This tool adjusts each feature so that the
mean is zero and the standard deviation is one, making
predictive models less sensitive to the varying scales of
different features and contributing to faster convergence during
learning.
Engineering, Technology & Applied Science Research
Vol. 14, No. 5, 2024, 16317-16325
16320
www.etasr.com Ait Ben Hamou et al.: Design of a Machine Learning-based Decision Support System for Product …
TABLE I. DATA PRE-PROCESSING
Features Description Count
quantity_p{i} Quantity of product requested P{i} p
delivery_start_hour_p{i} Delivery start time of P{i} p
delivery_end_hour_p{i} P{i} delivery end time p
P{i}_producible_m{j} Is P{i} productive by M{j} p * m
P{i}_production_cost_m{j} Cost of one period of production of P{i} by machine M{j} p * m
P{i}_production_qty_m{j} Quantity of P{i} that can be produced by M{j} per period p * m
maint_start_hour_m{k} Machine maintenance start period M{k} m
maint_end_hour_m{k} Machine maintenance end time M{k} m
P{i} = k The product P{i} will be assigned to the machine M{k}. p
optimal_cost
Optimum production
cost
1
D. Proposed Methodology
Our approach to solving the production scheduling problem
in a supply chain that integrates ML into a structured DSS.
Figure 1 shows an overview of the proposed DSS. Upstream,
structural, and problem data feed the ML model. The latter then
performs classification and regression predictions, which are
then integrated into the ordination program to generate the final
scheduling solution. The diagram of our methodology details
the following components:
1) Composition of the Dataset
The dataset is composed of two categories of data:
Structural data: This data is fixed and does not vary
between different scheduling problems. It includes
information such as production rates and costs associated
with producing a product on a given machine.
Problem data: Specific to each problem instance, this data
includes quantities of each product to be produced, time
windows for product delivery, and scheduled maintenance
periods for each machine.
2) Output Characteristics
There are two types of output characteristics:
Categorical characteristics (P{i}): These represent the
assignment of a product to a machine, where each product
takes as its value the number of the machine assigned to it.
Continuous characteristic (Optimal cost): A single
characteristic reflecting the optimal production cost for the
entire scheduling problem.
3) ML Models
Regression for optimal cost: We use regression models to
predict the continuous characteristic of optimal cost, thus
optimizing the selection of production parameters to
minimize expenditure.
Classification for product assignment: Classification
models are used to determine the assignment of products to
machines, with the creation of a separate model for each
product to predict the specific machine to which it will be
assigned.
Post-prediction ordering: Having predicted the assignment
of products to machines, a third Python program is used to
order the products assigned to each machine, respecting the
delivery time constraints associated with each product.
This methodology provides a comprehensive framework for
integrating ML into production scheduling problems,
addressing both production assignment forecasting and
production cost estimation. The aim is to provide optimal
scheduling that respects all the operational and economic
constraints inherent in SCM.
Fig. 1. The proposed methodology.
4) ML Algorithms Used
To address the two aspects of our production scheduling
problem, optimal cost prediction and product allocation to
machines, we have selected a suite of ML algorithms [29]
based on their performance in regression and classification. ML
algorithms used for regression (optimal cost prediction):
Linear regression: The foundation of predictive methods for
its interpretability and speed of calculation, used here as a
basic point of comparison.
KNN Regression (k=11): A decisive hyperparameter, the
number of nearest neighbors, was chosen after an
exhaustive search. k=11 was determined as offering the best
performance from a range of values from 1 to 50. Figure 2
shows the score obtained as a function of the value of k.
Regression Decision Tree (max_depth = 9): In a similar
way, the maximum depth of the tree has been optimized to
avoid overlearning while capturing the complexity of the
data as explained in Figure 3.
Engineering, Technology & Applied Science Research
Vol. 14, No. 5, 2024, 16317-16325
16321
www.etasr.com Ait Ben Hamou et al.: Design of a Machine Learning-based Decision Support System for Product …
Fig. 2. Performance of the KNN model as a function of the number of
neighbors.
Fig. 3. Decision Tree regression performance as a function of maximum
tree depth.
Extreme Gradient Boosting (XGBoost) Regression: the
model proposed by [30] renowned for its efficiency and
accuracy, is an advanced implementation of the gradient
boosting algorithm.
5) ML Algorithms used for Classification (Product
Assignment):
Logistic regression: Despite its simplicity, this robust model
is the benchmark for binary and multiclass classification.
KNN Classification (k=11): The same k selection principle
was applied here, ensuring consistency between our
regression and classification models.
Decision Tree Classification.
XGBoost Classification: Employs the same powerful
algorithm as for regression but adapted for categorical
results.
Each categorical output characteristic (P{i}) was addressed
with a separate model, enabling more accurate prediction and
fine-grained analysis of individual machine performance.
E. Metrics Used
To evaluate the effectiveness of the ML models, we chose
recognized metrics that reflect the objectives of our study.
For regression models, performance was measured by
RMSE (Root Mean Square Error), R2 score, MAE (Mean
Absolute Error) and Explained Variance Score:
RMSE: This metric is particularly informative as it
amplifies and penalizes larger errors, providing a strict
assessment of performance.
(1)
R² score (coefficient of determination): R² is a statistical
measure that tells us how much of the variance in the data is
explained by the model.
(2)
MAE: Gives an idea of the mean error without squaring it,
thus offering a more direct perspective of the mean absolute
error.
(3)
Explained Variance Score: To see how well our model
explains the variation in the data. This metric measures the
proportion of response variance that can be predicted from
the explanatory variables.
!"#$%&'(#)$#%*&
+,-./
0,-
(4)
where y
i
is the true values of y,
i
is the predicted values of y
and 1 is the average value of y.
For classification models, we evaluated performance using
metrics that capture how effectively our models correctly
identify target classes. In the following TP represents True
Positives, FP represents False Positives, TN represents True
Negatives, and FN represents False Negatives.
Accuracy: This metric calculates the percentage of correct
predictions relative to the total number of predictions made,
providing an overall measure of model performance.
**2)#*1
34536
3789:;<=>?@7AB@?CDE8D7;F
(5)
Precision: Reflects the proportion of actual positive
identifications among the items labeled as positive by the
model. It is an indicator of the reliability of the model's
positive predictions [31].
G)&*$H$I%
34
345J4
(6)
Recall (or Sensitivity): This metric measures the model's
ability to identify all real instances of a particular class.
K&*#""
34
345J6
(7)
F1-Score: The harmonic mean of Precision and Recall,
provides a single score that balances these two metrics,
particularly useful when classes are unbalanced.
L M*I)& NO
4@?EDFD7;OP?E9::
4@?EDFD7;5P?E9::
(8)
Engineering, Technology & Applied Science Research
Vol. 14, No. 5, 2024, 16317-16325
16322
www.etasr.com Ait Ben Hamou et al.: Design of a Machine Learning-based Decision Support System for Product …
III. EXPERIMENTAL RESULTS
A. Regression Results
Table II summarizes the performance of each model
according to RMSE, MAE, Explained Variance, and R² Score,
allowing a direct comparison of their effectiveness in
predicting production costs. The results indicate that XGBoost
showed the best performance with an RMSE of 1090.16 and an
R² Score of 0.96, suggesting a strong fit with the test data,
which implies a superior ability of this model to model the
complexity of production costs compared to the other models
tested.
TABLE II. PERFORMANCE OF ML MODELS ACCORDING
TO RMSE, MAE, EXPLAINED VARIANCE, AND R2 SCORE
Model ML RMSE MAE Explained
Variance
R2
Score
KNN regression (K=11) 2900.80 2242.57 0.74 0.74
Decision Tree 1676.67 1099.49 0.91 0.91
Linear regression
1415.74
880.53
0.94
0.94
XGBoost regression 1090.16 707.51 0.96 0.96
Figure 4 shows a comparative RMSE plot for each model,
visually illustrating the relative effectiveness of each algorithm.
XGBoost's clear superiority on this graph reflects its robustness
under the varied conditions of our dataset.
Fig. 4. Graphical comparison of RMSE values.
The poor performance of the KNN model is perhaps
explained by the sensitivity to scale of the input features. Some
features have much wider value ranges than others, and will
dominate the distance between points, which can lead to poor
performance. Before the standardization of input features, the
score was much lower. Our data have non-linear relationships
that KNN has difficulty capturing with its proximity-based
prediction method, so the XGBoost model, which can capture
non-linear relationships and interactions between features, was
able to perform better. To visualize how close the predictions
are to the actual values, Figure 5 shows scatter plots for the ML
regression models used. Scatter plots show the relationship
between model-predicted values (y-axis) and actual values (x-
axis). Ideally, predictions should be as close as possible to the
black diagonal line, which represents a perfect prediction
where predicted values are equal to actual values.
Fig. 5. Comparison of predicted vs. actual values for different regression
models.
For the Decision Tree algorithm, the points appear to be
close to the ideal line, indicating that the model's predictions
are relatively accurate. For the KNN, the predictions are a little
more scattered in relation to the ideal line, which could indicate
less precision in the predictions compared to the Decision Tree.
The Linear Regression plot shows that predictions are tightly
aligned around the ideal line, suggesting that this model has
performed well and that predictions are consistent with actual
values. Finally, the XGBoost Regression shows a distribution
of points very close to the ideal line, suggesting that this model
has superior predictive ability compared to other models, with
predictions very much aligned with actual values.
B. Classification Results
The metrics used to evaluate these models were Accuracy,
Precision, Recall, and F1-Score.Table III presents these metrics
for each product feature (P1-P6), highlighting the models'
performance in correctly assigning products to machines. For
example, for P1, XGBoost outperformed the other models with
an Accuracy of 0.96 and an F1-Score of 0.84, demonstrating its
ability to correctly classify products with great accuracy and
efficiency. For P3, all models achieve perfection with a score
of 1.00 on all metrics, suggesting that P3 has highly distinctive
features that are easily learned by all models because the
structured data shows that P3 can only be produced by the m3
machine, and the value of P3 is the same for all dataset
instances. Figure 6 shows the ROC curves for feature P1
generated by different ML algorithms. These curves are a
visual indicator of the discriminatory performance of the
models, where an Area Under the Curve (AUC) close to 1.0
testifies to excellent classification ability. For P1, the XGBoost
model clearly stands out with an AUC of 0.97, denoting near-
perfect accuracy in class distinction. Next are the logistic
Engineering, Technology & Applied Science Research
Vol. 14, No. 5, 2024, 16317-16325
16323
www.etasr.com Ait Ben Hamou et al.: Design of a Machine Learning-based Decision Support System for Product …
regression model and the decision tree, with respective AUCs
of 0.90 and 0.77, indicating significant classification
competence, albeit slightly inferior to XGBoost. The KNN
model, despite an AUC of 0.69, shows a more moderate
performance, suggesting a less clear distinction between classes
for this specific model. These ROC curves confirm the
robustness of the XGBoost approach in our classification
context, reaffirming its potential as a predominant tool for the
accurate allocation of products to machines in DSSs.
TABLE III. RESULTS OF ML CLASSIFICATION MODEL
ML Model Accuracy
Precision
Recall
F1-Score
P1
KNN
0.92
0.96
0.50
0.48
Decision Tree
0.91
0.72
0.77
0.74
Logistic Regression
0.92 0.77 0.60 0.63
XGBoost 0.96 0.89 0.80 0.84
P2
KNN 0.88 0.63 0.33 0.31
Decision Tree 0.87 0.57 0.55 0.56
Logistic Regression
0.89
0.64
0.45
0.49
XGBoost 0.92 0.75 0.61 0.67
P3
KNN 1.00 1.00 1.00 1.00
Decision Tree 1.00 1.00 1.00 1.00
Logistic Regression
1.00 1.00 1.00 1.00
XGBoost 1.00 1.00 1.00 1.00
P4
KNN 0.87 0.29 0.33 0.31
Decision Tree 0.76 0.44 0.49 0.46
Logistic Regression
0.88 0.64 0.46 0.49
XGBoost 0.92 0.76 0.62 0.67
P5
KNN 0.96 0.48 0.50 0.49
Decision Tree 0.97 0.80 0.79 0.79
Logistic Regression
0.96 0.75 0.52 0.53
XGBoost 0.98 0.94 0.81 0.86
P6
KNN 0.97 0.48 0.50 0.49
Decision Tree 0.98 0.80 0.85 0.83
Logistic Regression
0.97 0.68 0.52 0.53
XGBoost 0.99 0.98 0.79 0.86
Fig. 6. Receiver Operating Characteristic (ROC) curves for P1.
The following section will explore the implications of these
results, analyzing the performance of the different models and
the potential reasons behind the observed trends. We will also
discuss the models' limitations and prospects for future
research.
Fig. 7. Confusion Matrices for P1.
IV. DISCUSSION
The experimental results presented highlight the robustness
of the XGBoost approach, which consistently outperformed
other classification models in our study. With near-perfect
ROC curve AUCs for features P1, P5 and P6, XGBoost proved
its ability to operate with high accuracy, even when
classification conditions were less distinct and more complex.
The previous section illustrated XGBoost's superiority with
metrics such as Accuracy, Precision and F1-Score, which, in
correlation with the ROC curves, strengthen the case for its
adoption for similar classification tasks in production contexts.
Notably, for features where other models faltered, such as P5
and P6, XGBoost demonstrated exceptional performance,
suggesting its ability to handle the variability and complexity
inherent in production data. These results are consistent with
the findings of other research in the field of ML on other
problems [31-36]. However, it's important to recognize that,
despite these successes, there are inherent limitations to the
application of ML models in real-world environments. The
models, while accurate, are subject to the quality of the input
data and can be sensitive to use cases that deviate from training
conditions.
Furthermore, the evaluation of regression models showed
that the XGBoost model, with the lowest RMSE and the
highest R² Score, illustrated an exceptional ability to anticipate
costs accurately and consistently. Scatter plots of predictions
versus actual values confirmed XGBoost's superiority,
reflecting a dense concentration of points along the ideal line,
symbolizing near-perfect predictions. The linear regression and
Decision Tree models also performed well, but with greater
variance, indicating slightly lower accuracy in cost modeling.
The KNN model, despite its simplicity and speed of execution,
showed fewer convincing results in terms of RMSE and R²,
Engineering, Technology & Applied Science Research
Vol. 14, No. 5, 2024, 16317-16325
16324
www.etasr.com Ait Ben Hamou et al.: Design of a Machine Learning-based Decision Support System for Product …
which can be attributed to its difficulty in capturing the
complexity and non-linearity of the data.
Although our results highlight the power of the XGBoost
approach for classifying production data, a notable constraint of
our study is the scalability of exhaustive solution computation.
This method, while optimal for small instances, becomes
impractical for large datasets due to its exponential
computational cost. This reality points to the urgency of
adopting approximation strategies or heuristics that could
deliver near-optimal solutions faster, thus better aligning with
the dynamic requirements of modern production environments.
In addition, this research paves the way for future
investigations, where the exploration of more advanced
techniques, such as deep learning or hyperparameter
optimization, could be employed to further improve model
performance. Future research could also examine the impact of
integrating real-time feedback and dynamic adjustments into
models, aiming to make production systems even more
responsive and economically efficient.
V. CONCLUSION
This research has tackled the complex challenge of
production scheduling by combining the capabilities of ML
with a structured DSS for SCM. Although previous research
has significantly contributed to integrating ML techniques into
scheduling within SCM, several issues remain open.
Specifically, (a) many contributions deal with specific
production environments and do not discuss the ability to
generalize scheduling solutions to different industries or
diversified production systems, (b) the proposed models
present significant complexity in their configuration and
optimization, requiring advanced technical expertise, and (c)
the integration of predictive maintenance into scheduling is
rarely addressed.
To address this issue, we propose a methodology that offers
a comprehensive approach integrating structural and problem-
specific data to feed an effective predictive model, having the
advantage of optimizing costs and product allocation to
machines. The strength of our approach lies in its ability to
dynamically adapt scheduling solutions to variations in
production data, while taking operational constraints into
account. Through the elaboration of a rich and well-structured
dataset, we were able to demonstrate the relevance of
regression models for estimating optimal costs and
classification models for predicting product assignment. Our
approach has demonstrated accurate prediction of production
assignments while providing cost estimates in line with current
economic requirements. The proposed methodology
incorporates a post-prediction ordering stage, which enables the
products assigned to each machine to be logically organized,
guaranteeing on-time delivery and maximizing the efficiency
of the production process. This refinement is crucial for supply
chain practitioners who are looking for solutions that are not
only cost-optimized, but also achievable on time. However, we
recognize that exhaustive computation of optimal solutions for
all permutations has limitations in terms of scalability,
particularly in the case of large data sets. This limitation opens
up prospects for the integration of heuristic and optimization
methods that could speed up the search for solutions while
maintaining acceptable quality.
To summarize, our contribution provides a solid foundation
for the future integration of intelligent DSSs into production
scheduling. Future research could focus on improving
computational efficiency and exploring even more
sophisticated predictive models, propelling the production
industry into an era of digitization and automation.
REFERENCES
[1] D. Ni, Z. Xiao, and M. K. Lim, "A systematic review of the research
trends of machine learning in supply chain management," International
Journal of Machine Learning and Cybernetics, vol. 11, no. 7, pp. 1463–
1482, Jul. 2020, https://doi.org/10.1007/s13042-019-01050-0.
[2] G. Schmidt, "A Decision Support System for Production Scheduling,"
Journal of Decision Systems, vol. 1, no. 2–3, pp. 243–260, Jan. 1992,
https://doi.org/10.1080/12460125.1992.10511527.
[3] M. Đumić, D. Šišejković, R. Čorić, and D. Jakobović, "Evolving priority
rules for resource constrained project scheduling problem with genetic
programming," Future Generation Computer Systems, vol. 86, pp. 211–
221, Sep. 2018, https://doi.org/10.1016/j.future.2018.04.029.
[4] B. Waschneck et al., "Optimization of global production scheduling with
deep reinforcement learning," Procedia CIRP, vol. 72, pp. 1264–1269,
Jan. 2018, https://doi.org/10.1016/j.procir.2018.03.212.
[5] Y.-R. Shiue, K.-C. Lee, and C.-T. Su, "Real-time scheduling for a smart
factory using a reinforcement learning approach," Computers &
Industrial Engineering, vol. 125, pp. 604–614, Nov. 2018,
https://doi.org/10.1016/j.cie.2018.03.039.
[6] S. L. Takeda Berger, R. M. Zanella, and E. M. Frazzon, "Towards a
data-driven predictive-reactive production scheduling approach based on
inventory availability," IFAC-PapersOnLine, vol. 52, no. 13, pp. 1343–
1348, Jan. 2019, https://doi.org/10.1016/j.ifacol.2019.11.385.
[7] L. He, W. Li, Y. Zhang, and Y. Cao, "A discrete multi-objective
fireworks algorithm for flowshop scheduling with sequence-dependent
setup times," Swarm and Evolutionary Computation, vol. 51, pp.
100575, Dec. 2019, https://doi.org/10.1016/j.swevo.2019.100575.
[8] C. D. Hubbs, C. Li, N. V. Sahinidis, I. E. Grossmann, and J. M.
Wassick, "A deep reinforcement learning approach for chemical
production scheduling," Computers & Chemical Engineering, vol. 141,
pp. 106982, Oct. 2020, https://doi.org/10.1016/j.compchemeng.
2020.106982.
[9] K. Guo, M. Yang, and H. Zhu, "Application research of improved
genetic algorithm based on machine learning in production scheduling,"
Neural Computing and Applications, vol. 32, no. 7, pp. 1857–1868, Apr.
2020, https://doi.org/10.1007/s00521-019-04571-5.
[10] Y. Li, S. Carabelli, E. Fadda, D. Manerba, R. Tadei, and O. Terzo,
"Machine learning and optimization for production rescheduling in
Industry 4.0," The International Journal of Advanced Manufacturing
Technology, vol. 110, no. 9, pp. 2445–2463, Oct. 2020, https://doi.org/
10.1007/s00170-020-05850-5.
[11] C. Morariu, O. Morariu, S. Răileanu, and T. Borangiu, "Machine
learning for predictive scheduling and resource allocation in large scale
manufacturing systems," Computers in Industry, vol. 120, pp. 103244,
Sep. 2020, https://doi.org/10.1016/j.compind.2020.103244.
[12] C. Zhang, W. Song, Z. Cao, J. Zhang, P. S. Tan, and X. Chi, "Learning
to Dispatch for Job Shop Scheduling via Deep Reinforcement Learning,"
in Advances in Neural Information Processing Systems, 2020, vol. 33,
pp. 1621–1632.
[13] Z. Shi, H. Ma, M. Ren, T. Wu, and A. J. Yu, "A learning-based two-
stage optimization method for customer order scheduling," Computers &
Operations Research, vol. 136, pp. 105488, Dec. 2021,
https://doi.org/10.1016/j.cor.2021.105488.
[14] W. Guo, M. Vanhoucke, J. Coelho, and J. Luo, "Automatic detection of
the best performing priority rule for the resource-constrained project
scheduling problem," Expert Systems with Applications, vol. 167, pp.
114116, Apr. 2021, https://doi.org/10.1016/j.eswa.2020.114116.
Engineering, Technology & Applied Science Research
Vol. 14, No. 5, 2024, 16317-16325
16325
www.etasr.com Ait Ben Hamou et al.: Design of a Machine Learning-based Decision Support System for Product …
[15] S. Lang, T. Reggelin, J. Schmidt, M. Müller, and A. Nahhas,
"NeuroEvolution of augmenting topologies for solving a two-stage
hybrid flow shop scheduling problem: A comparison of different
solution strategies," Expert Systems with Applications, vol. 172, pp.
114666, Jun. 2021, https://doi.org/10.1016/j.eswa.2021.114666.
[16] H. Yamashiro and H. Nonaka, "Estimation of processing time using
machine learning and real factory data for optimization of parallel
machine scheduling problem," Operations Research Perspectives, vol.
8,pp. 100196, Jan. 2021, https://doi.org/10.1016/j.orp.2021.100196.
[17] S. Zhai, B. Gehring, and G. Reinhart, "Enabling predictive maintenance
integrated production scheduling by operation-specific health
prognostics with generative deep learning," Journal of Manufacturing
Systems, vol. 61, pp. 830–855, Oct. 2021, https://doi.org/10.1016/j.jmsy.
2021.02.006.
[18] J. Juros, M. Brcic, M. Koncic, and M. Kovac, "Exact solving scheduling
problems accelerated by graph neural networks," in 2022 45th Jubilee
International Convention on Information, Communication and
Electronic Technology (MIPRO), Feb. 2022, pp. 865–870,
https://doi.org/10.23919/MIPRO55190.2022.9803345.
[19] K. Lei et al., "A multi-action deep reinforcement learning framework for
flexible Job-shop scheduling problem," Expert Systems with
Applications, vol. 205, pp. 117796, Nov. 2022, https://doi.org/
10.1016/j.eswa.2022.117796.
[20] R. Braune, F. Benda, K. F. Doerner, and R. F. Hartl, "A genetic
programming learning approach to generate dispatching rules for flexible
shop scheduling problems," International Journal of Production
Economics, vol. 243, pp. 108342, Jan. 2022, https://doi.org/10.1016/
j.ijpe.2021.108342.
[21] S. Jungbluth, N. Gafur, J. Popper, V. Yfantis, and M. Ruskowski,
"Reinforcement Learning-based Scheduling of a Job-Shop Process with
Distributedly Controlled Robotic Manipulators for Transport
Operations," IFAC-PapersOnLine, vol. 55, no. 2, pp. 156–162, Jan.
2022, https://doi.org/10.1016/j.ifacol.2022.04.186.
[22] A. Rinciog and A. Meyer, "Towards Standardising Reinforcement
Learning Approaches for Production Scheduling Problems," Procedia
CIRP, vol. 107, pp. 1112–1119, Jan. 2022, https://doi.org/10.1016/
j.procir.2022.05.117.
[23] E. Morinaga, X. Tang, K. Iwamura, and N. Hirabayashi, "An improved
method of job shop scheduling using machine learning and mathematical
optimization," Procedia Computer Science, vol. 217, pp. 1479–1486,
Jan. 2023, https://doi.org/10.1016/j.procs.2022.12.347.
[24] C.-L. Liu, C.-J. Tseng, T.-H. Huang, and J.-W. Wang, "Dynamic Parallel
Machine Scheduling With Deep Q-Network," IEEE Transactions on
Systems, Man, and Cybernetics: Systems, vol. 53, no. 11, pp. 6792–6804,
Aug. 2023, https://doi.org/10.1109/TSMC.2023.3289322.
[25] W. Song, X. Chen, Q. Li, and Z. Cao, "Flexible Job-Shop Scheduling via
Graph Neural Network and Deep Reinforcement Learning," IEEE
Transactions on Industrial Informatics, vol. 19, no. 2, pp. 1600–1610,
Oct. 2023, https://doi.org/10.1109/TII.2022.3189725.
[26] S. C. Graves, "A Review of Production Scheduling," Operations
Research, vol. 29, no. 4, pp. 646–675, May 1981, https://doi.org/
10.1287/opre.29.4.646.
[27] F. A. Rodammer and K. P. White, "A recent survey of production
scheduling," IEEE Transactions on Systems, Man, and Cybernetics, vol.
18, no. 6, pp. 841–851, Aug. 1988, https://doi.org/10.1109/21.23085.
[28] M. Pinedo and K. Hadavi, "Scheduling: Theory, Algorithms and
Systems Development," in Operations Research Proceedings 1991,
Berlin, Heidelberg, 1992, pp. 35–42, https://doi.org/10.1007/978-3-642-
46773-8_5.
[29] S. Angra and S. Ahuja, "Machine learning and its applications: A
review," in 2017 International Conference on Big Data Analytics and
Computational Intelligence (ICBDAC), Mar. 2017, pp. 57–60,
https://doi.org/10.1109/ICBDACI.2017.8070809.
[30] T. Chen and C. Guestrin, "XGBoost: A Scalable Tree Boosting System,"
in Proceedings of the 22nd ACM SIGKDD International Conference on
Knowledge Discovery and Data Mining, New York, NY, USA, May
2016, pp. 785–794, https://doi.org/10.1145/2939672.2939785.
[31] A. Saini, K. Guleria, and S. Sharma, "An Efficient Deep Learning Model
for Eye Disease Classification," Jun. 2023, pp. 1–6, https://doi.org/
10.1109/SCSE59836.2023.10215000.
[32] P. Hajek, M. Z. Abedin, and U. Sivarajah, "Fraud Detection in Mobile
Payment Systems using an XGBoost-based Framework," Information
Systems Frontiers, vol. 25, no. 5, pp. 1985–2003, Oct. 2023,
https://doi.org/10.1007/s10796-022-10346-6.
[33] S. Ramaneswaran, K. Srinivasan, P. M. D. R. Vincent, and C.-Y. Chang,
"Hybrid Inception v3 XGBoost Model for Acute Lymphoblastic
Leukemia Classification," Computational and Mathematical Methods in
Medicine, vol. 2021, no. 1, pp. 2577375, 2021, https://doi.org/
10.1155/2021/2577375.
[34] R. Szczepanek, "Daily Streamflow Forecasting in Mountainous
Catchment Using XGBoost, LightGBM and CatBoost," Hydrology, vol.
9, no. 12, pp. 226, Dec. 2022, https://doi.org/10.3390/hydrology
9120226.
[35] W. Ismaiel, A. Alhalangy, A. O. Y. Mohamed, and A. I. A. Musa, "Deep
Learning, Ensemble and Supervised Machine Learning for Arabic
Speech Emotion Recognition," Engineering, Technology & Applied
Science Research, vol. 14, no. 2, pp. 13757–13764, Apr. 2024,
https://doi.org/10.48084/etasr.7134.
[36] C. Matara, S. Osano, A. O. Yusuf, and E. O. Aketch, "Prediction of
Vehicle-induced Air Pollution based on Advanced Machine Learning
Models," Engineering, Technology & Applied Science Research, vol. 14,
no. 1, pp. 12837–12843, Feb. 2024, https://doi.org/10.48084/etasr.6678.
[37] H. Al-Dossari, F. A. Nughaymish, Z. Al-Qahtani, M. Alkahlifah, and A.
Alqahtani, "A Machine Learning Approach to Career Path Choice for
Information Technology Graduates," Engineering, Technology &
Applied Science Research, vol. 10, no. 6, pp. 6589–6596, Dec. 2020,
https://doi.org/10.48084/etasr.3821.
