Journal of Scientific & Industrial Research

Vol. 84, March 2025, pp. 269-277

DOI: 10.56042/jsir.v84i03.8230

Improved Machine Learning based Approach for Autotuning PID Controller using
Genetic Algorithms and Parallel Processing
Aghil Ahmadi & Reza Mahboobi Esfanjani*
Department of Electrical Engineering, Sahand University of Technology, Tabriz, Iran

Received 22 January 2024; revised 18 September 2024; accepted 22 December 2024

PID controllers are widely applied in approximately 95% of continuous control systems across process industries,
making them a cornerstone of control engineering. Despite their widespread use, these controllers are often inadequately
tuned. This study proposes an intelligent adaptive Proportional-Integral-Derivative (PID) controller for managing complex,
uncertain processes. To enhance the capabilities of traditional PID controllers, an advanced machine learning approach
using Deep Neural Networks (DNNs) is implemented. To optimize the configuration of the neural network and reduce
computational load, a Genetic Algorithm (GA)-based structural learning technique is used, combined with parallel
computing to accelerate training. Simulation results show that the proposed controller achieves an RMSE of 0.70 in the
absence of disturbances, outperforming the Standard PID (0.95 RMSE) and Shallow Neural PID (0.74 RMSE). Under a 20
dB SNR disturbance, the proposed approach maintains robust performance with an RMSE of 0.79, compared to the Standard
PID (1.10 RMSE) and Shallow Neural PID (0.86 RMSE). These findings highlight the superiority of the proposed
innovatively tuned controller over both standard PID controllers and recently introduced intelligent network-based tuners,
particularly in the presence of uncertainties.

Keywords: Adaptive tuning, Deep neural networks, Intelligent control, Parallel computing

Introduction intelligent solutions to such issues in the industry.4

Many complex processes with multiple control loops Recently, control systems have seen the use of machine
facing different types of uncertainty are found in modern learning techniques that may be useful for uncertain
industries. Although many efficient control systems nonlinear systems. For example, an adaptive fuzzy
have been found thus far, their complexity prevents controller was used to overcome position control
control engineers from designing and implementing issues.5 A feed-forward neural network and a fuzzy PID
them.1 Furthermore, enterprises are encouraged to use controller were used to operate a two-tank system.6 An
basic linear controls due to budgetary concerns. advanced network was designed to work on a
However, non-linear real-world processes with complex specialized robot.7 In most industrial applications, the
behavior cannot benefit from these controllers. conventional PID controller is commonly used.8,9 The
Combining the recently developed machine learning proportional, integral, and derivative components which
methods with the previously mentioned basic controllers take into consideration the error's present value, rate of
could be one creative way to address control issues in change, and cumulative error over time are combined to
contemporary process industries. The majority of provide the control output of a controller. The
modern industrial processes behave nonlinearly, which aforementioned terms are multiplied by their respective
is generally difficult to control.2 Additionally, linear constants—KD for the derivative, KI for the integrated,
plants that have a noticeable temporal delay are and KP for the present errors before being combined
frequently challenging to manage because time delays together. These three constants are adjusted to achieve
cause poor performance and even instability.3 The need tuning. The most popular controllers utilized in
for more sophisticated solutions an extension from industrial control procedures across a variety of
embedded systems to cyber-physical systems, whose operating circumstances are PID controllers. Reaching
key attributes are "Intelligence," "Autonomy," and "Self- the required tracking, good speed, minimal overshoots
Learning," is brought about by the need for more and undershoots, and the ability to protect actuators from
—————— failures and damages which result in fewer products that
*Author for Correspondence are off-spec are just a few of the many impressive
E-mail: mahboobi@[Link]
270 J SCI IND RES VOL 84 MARCH 2025

advantages of appropriate tuning. From a theoretical reviewed in numerous references.12 For example, the
standpoint, when the transformation function of the use of genetic algorithms to design PID controllers in
system is available, it is possible to calculate the the distillation column process, or genetic algorithms
controller parameters with the various existing to adjust the said controller to control glucose
techniques and reach the ideal answer, but considering concentration, or in other works of the PSO algorithm
that most industrial systems are non-linear, The to adjust and validate the control. It has been used on
transformation function is rarely accessible, because real physical plants such as bioreactors. The fact is
firstly The transformation function obtained is not that, despite the extensive previous applications, these
completely accurate and it is probably obtained by methods have been used for static adjustment of
linearization and approximation methods, and secondly, coefficients that were constant or had slight changes
by changing the loop components in the future, such as during the process. Artificial Neural Networks
changing the actuator (which happens a lot in the (ANNs) have received a lot of attention lately and
industry), the system transformation function will provide extremely high processing capacity due to
change. And naturally, the controlling coefficients their learning and generalization capabilities, as well
obtained based on the primary transformation function as their parallel distributed topology. A distributed
will no longer be valid. Nonetheless, trial-and-error- processor with greater capacity for parallelism is the
based tuning techniques are applied in the process neural network. In this way, it serves the same
industries. Obviously, these techniques are not precise purpose as the human brain by storing knowledge
enough and may result in significant hazards as well as a throughout training in the form of weighted
loss of productivity. connections between neurons. To address a problem,
PID controllers are easier to install and maintain, an appropriate structure with appropriate weights
but these advantages come at the expense of limited needs to be identified. It may be conceptualized as the
functionality. Only linear systems are well controlled reduction of an error function that is determined by
by classic PIDs. PID controllers are often calibrated desired and actual outputs of the neural network.
prior to deployment, so they are unable to account for Several feed forward neural networks were employed
unforeseen disruptions or adapt to changing to modify the controller parameters in several
circumstances in a way that would enable them to contexts.13,14 But these particular neural network
meet the demands of today's industries.10 Finding the topologies provide highly unique behavior.15
ideal or nearly ideal PID constants with the current Therefore, they are not adaptable enough to be
traditional tuning techniques is difficult. The effectively adapted for many unpredictable contexts.
following will provide a quick explanation of many In a broader study, appropriate neural network
traditional methods that were primarily utilized for topologies are examined for use in process industries,
PID parameter tuning: There have been ways to in contrast to competing efforts; our method is more
modify the PID controller's settings in recent years. useful because it effectively handles noise and
The conventional methods are Ziegler-Nichols, Chien disturbance. A network's topology, number of hidden
and T-Sigma. The Ziegler Nichols technique was layers, and associated weights are all significant and
shown in 1942 and was recognized as one of the critical components that affect how well a network
easiest and most famous approach for tuning. These performs. There are some drawbacks to both small
methods require a measured step response from the and large networks. Networks of very small sizes are
control system and are in fact dependent on the step difficult to train. On the other hand, both theory16 and
response. The first two methods have a table to experience17,18 demonstrate that networks with fewer
determine the controller coefficients and the third parameters perform better in terms of generalization,
method uses an integration approach for system which may be explained by keeping in mind the
responses. There are other methods that operate based comparison between curve fitting and neural network
on Root Lucas analysis, or other methods that look at training. Neural networks are not frequently utilized
adjusting the controller coefficients as an optimization for parameter adjustment in practical applications,
problem. Genetic algorithms are typically employed despite their potential. One of the main reasons is the
in these situations. Genetic algorithms are one poor interpretability, i.e., the obtained PID parameters
heuristic approach that has proven useful for PID cannot be assessed for their appropriateness. Current
tuning.11 Common methods of tuning have been neural network-based PID tuning research overlooks
 AHMADI & MAHBOOBI ESFANJANI: IMPROVED MACHINE LEARNING APPROACH FOR AUTOTUNING PID 271

closed-loop stability, an important area of control output (y) and control error (e). This approach
theory.19 Given that neural networks reveal a system's employs a network to estimate the PID controller
black box,20 Their use in industry for PID tuning is parameters.
still uncommon.21 A new machine learning approach Important considerations are the neural network's
is used to adjust PID values live. The most similar ideal architecture and size. Along with the online
work to ours in the literature tunes the PID controller computation and adjustment durations, the first offline
using a basic neural network without discussing the training is another issue that must be considered. On
neural network's architecture.10 This work explicitly larger networks, these times inevitably rise. And
discusses structural learning and suggests the providing a suitable solution in this regard (obtaining
potential use of genetic algorithms as a future the optimal network topology, which naturally
research direction. Unlike the current study, which involves removing some connections during structure
employs a basic shallow network, A deep network is learning, while using parallel processing with an
utilized in the proposed structure. To enhance speed efficient creative structure) is an important part of this
and efficiency, Genetic algorithms and parallel work, which will be discussed in detail.
processing are innovatively incorporated in an
optimized manner. This approach effectively manages Proposed Method
nonlinearity and uncertainties. The key innovation of In order for the presented method to be applicable
our approach lies in its unique structural optimization, on an industrial scale, all important and influential
enabling the simultaneous identification of optimal issues have been considered in this work. As
weights and network structure for the deep network in explained in the part before, the considered neural
consideration. In the remainder: First, the system with network that acts as a regulator of the controller
a tuner and deep neural network is presented, along coefficients has two inputs and three outputs. The
with its beneficial characteristics for our study and deep neural networks employed in this work are
more details are provided next. Comparative results limited to two hidden layers, each containing four
and discussions are presented in the following part, neurons, to ensure low-hardware online computation
and finally, the concluding summarizes the paper. while simultaneously taking advantage of deep
network advantages (Fig. 2).
Proposed ANN Structure for PID Tuning Reusing features that arise repeatedly is made
In this work, the traditional PID controller is possible by deep networks' hierarchical nature of
enhanced to be more intelligent and accurate while organizing and storing information. As a result, they
maintaining its basic structure to align with current serve a more precise purpose than a surface-level
industrial process control applications. The selection one.22 This type of deep network memorization
of input and output types and quantities represents the capability greatly aids in the management of
initial stage for configuring the neural network. As uncertainties in our problem. To ensure acceptable
shown in Fig. 1, the controller parameters are taken as performance before initiating control, the neural
the three outputs of the network (KP, KI, and KD), network is trained offline, with both its weights and
with the inputs being the closed-loop control system structure optimized. During operation, the deep
network must be trained efficiently at each step. An
ANN's success greatly depends on the parameters that
are chosen. Therefore, it is crucial to focus more on
accurately determining them. The number of hidden
layers and neurons in each layer are the primary
components of neural architecture. Considering the

Fig. 1 — The control structure with tuner Fig. 2 — Employed deep neural network
272 J SCI IND RES VOL 84 MARCH 2025

given problem and the number of its parameters, The U (30 × 36) = (Wmn) … (1)
structure of input and output layers can be
determined. Therefore, the main challenge is There are 36 columns and 30 rows in this matrix. The
determining the hidden layer structure, and weights row indicates the population size, and the column shows
are also assigned to network connections. the counting of values used to display each candidate for
a network structure. A starting configuration, as shown
Genetic Algorithms for Learning and Optimization in Table 1, is needed to conduct appropriate testing. In
Similar to other optimization techniques like PSO, the fully connected model for Multi Layer Perceptron
Genetic Algorithms (GAs) are adaptable for search (MLP), an initial network with 36 connections is started.
and optimization problems. First introduced by J. Thirty initial populations are then examined, with each
Holland and based on Darwin's theory of evolution, population representing a network that might solve the
GAs are evolutionary algorithms that rely on mutation issue. An algorithm is initiated based on this preliminary
and selection. The process begins with an initial data to identify the optimal structure with the least
population—here, the network's initial weights and amount of network error. Taking into account the table
structure—where connections with zero weights are containing the starting settings, experiments are run, and
eliminated. This initial population represents potential the genetic method is adhered to.
solutions within a defined search space, evaluated Experiments were initially conducted for a
using a fitness function to identify the best candidates. population size of 20, and the process was then
In the proposed structure (Fig. 3), weights represent repeated for a population size of 30, with different
both the network's structure and connections, enabling findings obtained in each case. Both stages underwent
simplified representation through a single array multiple iterations (10, 100, 500, 1000, and 2000).
without separating matrices or converting values to The computation time and error (RMSE) for each test
binary. Using real-valued weights directly facilitates were depicted in Table 2.
efficient coding and eliminates the need to differentiate This table can be distinguished into three important
between structure and weights, streamlining the parts. First, periodic variation is observed in the initial
optimization process. steps of the procedure. In the second step, a decrease
There are 13 neurons in our network: the first six in error is evident, and finally, a saturation state is
have four connections, the following four have three reached. At the end, the optimum weights for the
connections, and the final three provide the outputs. network are determined based on the minimum error.
Thus, the network can be shown with a total of 36 Connections with weights equal to or near zero are
weights, which are placed in the array that serves as a ignored. Using innovative GA-based procedures, the
chromosome (according to GA format) in this network is initially trained for both its weights and
instance (Fig. 3). The subscripts of W are the numbers structure.
of the source and destination neurons of the Another important challenge around neural
corresponding connection. Therefore, an array of networks is speeding up calculations and reducing
length 36 units is used to present the weights and computation time. The first solution involves the use
structure of the network. As mentioned before, the of supercomputers, but these systems present
array's weights determine both the weight and challenges, including high costs and maintenance
the structure for us; a zero indicates that the issues, as well as limited accessibility. Instead, an
corresponding connection is absent, and a number optimized parallel processing technique is employed,
other than zero indicates that the connection is present which will be explored in detail with supporting
in the network with the weight equals that number. figures in the discussion section. The network is
The given number also represents the weight of the intended to be used as a tuner for PID tuning during
connection. As a result, every array or chromosome Table 1 — Initial setting of Neural Network (NN)
serves as a network structure. If various population NN configuration parameters Initial configuration values
sizes are taken into account, the starting population ANN Type Deep (MLP)
can be displayed using a matrix like the one below: Number of initial neurons 13
Number of connections 36
Chromosome size 36
Population size 30
Neurons in hidden layers 8
Fig. 3 — Array for present network Training patterns 32
 AHMADI & MAHBOOBI ESFANJANI: IMPROVED MACHINE LEARNING APPROACH FOR AUTOTUNING PID 273

Table 2 — Results for different Iterations

No. of Iterations RMSE CPU time(s) RMSE CPU time(s)
Popsize = 20 Popsize = 20 Popsize = 30 Popsize = 30
10 1.27 0.08 1.50 0.15
100 1.03 0.61 1.44 0.68
500 0.93 2.31 1.31 3.10
1000 0.85 4.41 1.36 6.23
2000 0.90 8.89 1.08 12.04
an online process. The initial offline setting serves as
a starting point for control but is insufficient for
addressing changing conditions and uncertainties.
Therefore, adaptive online tuning is required.

Online Training
The popular method of training is back-propagation.23
In this case, an unsupervised scenario is encountered.
Unlike supervised techniques, where desired outputs are
available, the ideal values—namely, the PID parameters
to be tuned—are unknown. In supervised learning, Fig. 4 — Weight and structure updating procedure
network training relies on minimizing the difference
between ideal and actual outputs. However, in this ∆𝑤 𝑤 𝑤 ∝ ... (2)
situation, neural network training becomes a nonlinear where, α is the learning rate. Its optimal quantity
optimization issue. depends on the specific problem and it is determined
To modify weights, deep learning typically through a combination of empirical experimentation
employs end-to-end optimization; however, the and established methodologies to ensure optimal
Levenberg-Marquardt numerical technique (proper convergence. Our goal is to reduce the error. Partial
for small networks) is applied. The tuner's role is to derivatives can be used as plant output instead of
identify appropriate controller parameters, guiding the error.24 Thus, updating is based on Eq. (2):10
output toward the set-point, essentially solving a
∗ ∗
tracking control problem. Consequently, network 𝚥̂= ... (3)
∆ ∆ ∆
training focuses on minimizing tracking errors, with
the objective function prioritizing this error reduction. where, y is output and y∗ is the set point. Therefore:
Using the Jacobian, the impact of individual weights
∆
on the tracking error is assessed, and adjustments are 𝚥̂= 𝜀 𝑚𝑎𝑥 1, 𝑤 √𝜀 ... (4)
made to reduce error levels. In the context of
structural training, connections that result in reduced where, ȷ̂ is the partial derivate of error to the weights
error when removed are pruned, and the error caused and ε is the machine precision. Due to the lack of
by deleting a connection is evaluated. As a result, a ideal training data, an end-to-end gradient descent-
tuner is specifically designed with weights and based online training procedure, as illustrated in Fig.
topology optimized for error reduction. Leveraging 4, is implemented. This approach enables the network
the network's dynamic properties enables effective to adaptively fine-tune the controller parameters in
adaptation to unknown uncertainties, enhancing real-time to minimize control error effectively. The
response speed and minimizing network size. training process iteratively evaluates each connection
Initially, the network structure is selected based on within the neural network by measuring its
application requirements, with the PID controller contribution to the overall system performance.
trained offline using traditional methods. Structural Specifically, each connection is assessed based on the
and weight changes are made for fine-tuning during increase in control error when that connection is
operation. The system's error (RMSE) is minimized as removed, ensuring that only the most significant
the objective function, and weights are adjusted at connections are retained. By prioritizing critical
each training step to reduce the control error: connections, the network achieves optimal parameter
274 J SCI IND RES VOL 84 MARCH 2025

adjustments while maintaining computational

efficiency. This strategy not only enhances the
adaptability of the control system but also ensures
robustness against disturbances and noise in real-
world applications.
Data about the variable neural network is compiled
into a table to apply the training method. The input
weights, intermediate connection weights, and bias
weights constitute the three main parts of network
information. Together, these weights are also present
the structural configuration. When a connection is
removed and its corresponding weight is set to zero, the
network output and the resulting control error are Fig. 5 — MATLAB distributed computing
recalculated. Subsequently, a specific matrix, referred
to as the evaluation table, is used to evaluate network
connections based on their error contribution
ratings. Every network connection is assessed at
each stage according to the control mistakes that
arise. Specifically, the error impact of omitting
each connection is calculated, and those with minimal
error contribution are identified by sorting them
accordingly.

Comparative Simulation and Discussion

Proposed Parallel Processing
As mentioned earlier, the discussion of training
time is a crucial issue regarding neural networks. In
Fig. 6 — Speedup ratio for several processors
this context, parallel processing using MATLAB was
employed (Fig. 5). is tested both with and without added noise. Gaussian
With a suitable setting, A study was performed to white noise, with a Signal-to-Noise Ratio (SNR) of
illustrate the practicality of this technique. The 20 dB, is applied to the system output to simulate the
experiment was performed using different numbers of effect of noisy sensor measurements. To further
processor cores in an advanced research center, and evaluate the controller's resilience, An appropriate
the results were compared. A typical training problem disturbance is introduced into each system. A
was tested using a varying number of processors with disturbance is an unexpected and undesired input to
shared memory. The processing load was divided the system that increases system error. These
among the processors, with each processor handling a disturbances vary by system type and may result from
portion of the load and subsequently providing the external factors or internal system malfunctions. The
results to the manager system. The speedup ratio for disturbance magnitude for each system is gradually
different numbers of processors is shown in Fig. 6. increased from zero until only one control strategy
According to Amdahl's law, achieving a speedup can stabilize the system. Disturbances of this critical
equal to n with n processors is not possible, but efforts magnitude are then used in the experiments.
were made to minimize the calculation time as much Importantly, the neural PID controller's training data
as possible. does not include these disruptions. To guarantee
statistical significance, every experiment is conducted
Simulation and Test over a hundred separate runs. The controller output is
To assess the neural PID controller's performance, updated every 0.1 seconds to simulate the discrete
two common control problems — a two-tank system sampling times of real-world conditions. This
and a system with significant time delay are used. intervention time represents the limitations of
Each system presents a distinct control challenge and physical actuators, which cannot adjust instantly. The
 AHMADI & MAHBOOBI ESFANJANI: IMPROVED MACHINE LEARNING APPROACH FOR AUTOTUNING PID 275

solver runs iteratively for each 0.1-second interval, strategies become especially valuable when dealing with
using the results of the previous step as initial complex dynamics or unpredictable disturbances, which
conditions, while the controller output u remains are typical in real-world scenarios. By fine-tuning the
constant throughout each 0.1-second window. PID settings, operators can ensure reliable and optimal
performance across various industrial applications,
Two-tank System
making two-tank systems an invaluable benchmark for
Aksu and Coban25 explain the differential equations
testing and advancing control strategies.
of a nonlinear two-tank system, as shown in Fig. 7. In
Between t = 0s and t = 10s, the two-tank system is
this two-tank system, the control of water levels is
continually perturbed to assess the robustness of the
crucial to maintain stability and achieve desired
contrasted control techniques. The controller output is
outcomes in applications ranging from chemical
reduced to zero as a form of disturbance, simulating a
processing to water treatment plants. The water level in
scenario where the water supply is depleted.
the second tank, acting as the controlled variable, must
be adjusted based on the input provided by the pump, System with time Delay
which is the manipulated variable. In such a setup, the The second system exhibits a significant temporal
control strategy focuses on ensuring that the water delay and is classified as a first-order Linear Time-
level remains within a target range, despite external Invariant (LTI) system. Time delay, a critical aspect
disturbances or internal dynamics. A PID controller is in control theory, is frequently overlooked during the
commonly used due to its ability to handle system design of controllers.26 Delays can lead to system
delays, maintain steady-state error correction, and adjust instability and reduced performance. These delays
to varying load conditions. To do effective control, may be caused by sensor lags, actuator response
advanced tuning approaches are often applied to the PID times, or other physical limitations, preventing the
parameters. These include optimization techniques, like PID controller from responding promptly to changes.
genetic algorithms or particle swarm optimization, and This can result in oscillations, overshoot, or persistent
machine learning approaches, which enable the errors, ultimately reducing the efficiency of the
controller to adaptively adjust to the system's dynamics, system. The results of all experiments are presented in
resulting in higher efficiency and stability. Such Table 3, demonstrating the impact of various delay
conditions on system stability and performance.
The RMSE results for three conventional PID tuning
methods — Standard PID, Shallow Neural PID10, and
Reinforcement Learning27 are presented in Table 3 along
with the Proposed Deep Approach under varying
disturbance and SNR (Signal-to-Noise Ratio)
conditions. The results clearly demonstrate the
superiority of the Proposed Deep Approach.

Results Analysis
1. No Disturbance, Infinite SNR:
o Under baseline conditions, the Proposed Deep
Approach achieves the best performance with an
RMSE of 0.70, outperforming both Standard PID
(0.95 RMSE) and Shallow Neural PID (0.74
Fig. 7 — Two-Tank system RMSE).
Table 3 — RMSE on test data over 100 runs
Two tank system LTI system with input delay
Disturbance — —   — —  
Signal-to-Noise Ratio — 20dB — 20dB — 20dB — 20dB
Standard PID 0.95 0.99 1.00 1.10 0.22 0.23 0.36 0.38
Shallow Neural PID 0.74 0.86 0.84 1.00 0.13 0.15 0.26 0.28
Proposed Deep Approach 0.70 0.79 0.75 0.90 0.09 0.11 0.17 0.17
Reinforcement Learning 0.72 0.80 0.78 0.92 0.11 0.13 0.20 0.21
276 J SCI IND RES VOL 84 MARCH 2025

o This highlights the superior accuracy and efficiency The results indicate that the parameters Kp, Ki, and
of the Proposed Deep Approach, even in ideal Kd are adjusted over time to address changing system
conditions. dynamics, disturbances, and noise. Specifically,
2. Presence of Noise (20dB) and Disturbance (Two- between 6 to 8 seconds, where noise and disturbances
Tank System): become more pronounced, the tuner increases Ki and
o The Proposed Deep Approach maintains its superior Kp values to reduce steady-state error and improve
performance with an RMSE of 0.79, which is system responsiveness. This highlights the tuner's
notably lower than that of Standard PID (0.99 ability to adapt quickly to challenging conditions
RMSE) and Shallow Neural PID (0.86 RMSE). while maintaining overall stability and performance.
o These results underscore the robustness and Such adaptability makes the proposed approach
adaptability of the Proposed Deep Approach to highly effective for real-world applications where
disturbances and low SNR environments. system dynamics are constantly changing.
3. LTI System with Input Delay:
o Under challenging conditions involving input Conclusions
delay and noise (20dB), the Proposed Deep This study presents an advanced approach for PID
Approach again achieves the lowest RMSE values parameter tuning using deep neural networks (DNNs),
(0.11–0.17), significantly outperforming the optimized with genetic algorithms (GAs) and
Standard PID (0.22–0.38 RMSE) and Shallow enhanced through parallel processing. The proposed
Neural PID (0.15–0.28 RMSE). controller was evaluated on two benchmark systems:
o The method’s ability to handle delays and noisy a two-tank system and an LTI system with input
environments highlights its practical applicability delay. Experimental results demonstrate that the
to real-world systems. proposed method outperforms conventional PID
tuning techniques and shallow network-based
Key Insights approaches, particularly under challenging conditions
The Proposed Deep Approach consistently involving disturbances, uncertainties, and sensor
demonstrates: noise. The findings highlight the proposed controller's
 Lower RMSE across all tested scenarios, effectiveness in improving adaptability, robustness,
indicating superior control accuracy. and performance, making it highly suitable for
 Robustness to disturbances, noise, and input industrial applications such as refineries and
delays. petrochemical plants, where precise control is
 Adaptability to dynamic and complex systems, essential. This research underscores the potential of
which traditional methods fail to achieve. machine learning to enhance classical control
In conclusion, the results in Table 3 confirm that the systems. Future work could explore the integration of
Proposed Deep Approach significantly outperforms advanced optimization algorithms or hybrid machine
conventional methods, making it an ideal solution for learning techniques to further improve control
control systems under various challenging conditions. accuracy. Additionally, ensuring real-time stability
The PID coefficients presented in Table 4 are during online adaptation remains a critical direction
dynamically tuned for a two-tank system using our for further investigation.
structurally optimized neural network-based tuner.
This method leverages neural networks enhanced by References
genetic algorithm optimization to ensure real-time 1 Pistikopoulos E N, Povoa A B, Lee J H, Misener R, Mitsos
adjustments that improve adaptability and control A, Reklaitis G V, Venkatasubramanian V & Gani R, A
review of process systems engineering advances, Comput
precision. Chem Eng, 147 (2021) 107252, [Link]
Table 4 — PID parameters for Two-tank system [Link].2021.107252.
Time (s) Kp Ki Kd 2 Iqbal J, Ullah M, Ghani Khan S, Khelifa B & Cukovic S,
Dynamic modeling and control of robotic systems: A review,
0 4.30 0.55 0.20 Nonlinear Eng, 8 (2017), [Link]
2 3.75 0.60 0.18 0077.
4 4.00 0.58 0.19 3 Espitia N & Perruquetti W, Predictor-feedback prescribed-
6 3.10 0.65 0.18 time stabilization of LTI systems with input delay, IEEE
8 5.45 0.72 0.17 Trans Autom Control, 67(6) (2022) 2784–2799,
10 4.60 0.70 0.19 [Link] 10.1109/TAC.2021.3093527.
 AHMADI & MAHBOOBI ESFANJANI: IMPROVED MACHINE LEARNING APPROACH FOR AUTOTUNING PID 277

4 Siepmann D & Graef N, Industrie 4.0 – grundlagen und 15 Bergstra J, Yamins D & Cox D, Making a science of model
gesamtzusammenhang, in Einführung und Umsetzung von search: Hyperparameter optimization in hundreds of
Industrie 4.0 (Springer) 2016, 17–82, [Link] dimensions for vision architectures, 30th Int Conf Mach
10.1007/978-3-662-48505-7_2. Learn, 28(1) (2013) 115–123.
5 Vijay M & Jena D, PSO-based neuro fuzzy sliding mode control 16 Baum E B & Haussler D, What size net gives valid
for a robot manipulator, J Electr Syst Inf Technol, 4(1) (2017) generalization?, Neural Comput, 1(1) (1989) 151–160,
243–256, [Link] 2016. 08.006. [Link]
6 Xiao-kan W, Zhong-liang S, Wanglei & Dong-qing F, 17 Denker J, Schwartz D, Wittner B, Solla S, Howard R, Jackel
Design and research based on fuzzy PID-parameters self- L & Hopfield J, Large automatic learning rule extraction and
tuning controller with MATLAB, 2008 Int Conf Adv Comput generalization, Complex Syst, 1 (1987) 877–922.
Theory Eng (IEEE) 2008, 996–999, [Link] 18 Chauvin Y, Generalization performance of overtrained
ICACTE.2008.9. backpropagation networks, in Neural Networks (Elsevier)
7 Chen Z, Liu Y, He W, Qiao H & Ji H, Adaptive neural 1990, 45–55.
network based trajectory tracking control for a nonholonomic 19 Glasmachers T, Limits of end-to-end learning, Proc Asian
wheeled mobile robot with velocity constraints, IEEE Trans Conf Mach Learn, 17 (2017) 17–32.
Ind Electron, 68(6) (2020), 5057–5067, [Link] 20 Yang Z, Zhang A & Sudjianto A, Enhancing explainability
10.1109/TIE.2020.2989711. of neural networks through architecture constraints, arXiv
8 Åström K J & Hägglund T, The future of PID control, preprint (2019) arXiv:1901.03838v2.
Control Eng Pract, 9(11) (2001) 1163–1175, [Link] 21 Zhao Z Y, Tomizuka M & Isaka S, Fuzzy gain scheduling of
10.1016/S0967-0661(01)00062-4. PID controllers, IEEE Trans Syst Man Cybern, 23(5) (1993)
1392–1398, [Link]
9 Chun-Tang C, Nana S, Juing-Shian C & Chi-Jo W, An
22 Lee J H, Shin J & Realff M J, Machine learning overview
optimal fuzzy PID controller design based on conventional
of the recent progresses and implications for the process
PID control and nonlinear factors, Appl Sci, 9(6) (2019),
systems engineering field, Comput Chem Eng, 114 (2018)
[Link]
111–121, [Link]
10 Guenther J, Machine Intelligence for Adaptable Closed-Loop 10.008.
and Open-Loop Production Engineering Systems, Doctoral 23 Chauvin Y & Rumelhart D E (Eds.), Backpropagation: theory,
Dissertation, Technische Universität München (2018). architectures and applications (Lawrence Erlbaum Assoc),
11 Slavov T & Roeva O, Application of genetic algorithm to 1995.
tuning a PID controller for glucose concentration control, 24 De Jesus O & Hagan M T, Back propagation algorithms for a
WSEAS Trans Syst 11(7) (2012) 223–233, Retrieved from broad class of dynamic networks, IEEE Trans Neural Netw,
[Link]. 18(1) (2007) 14–27.
12 Maghade D K, Sondkar S Y & Pawar S N, A review of PID 25 Aksu U & Coban R, Sliding mode PI and backstepping
control tuning methods and applications, Int J Dyn Control, 9 control design for a nonlinear coupled-tank system, Int J
(2021) 818–827. Robust Nonlinear Control, 29(5) (2019) 45–61.
13 Rubio J J, Zhang L, Lughofer E, Panuncio Cruz P, Alsaedi A 26 Zhang Y, Li J & Wei X, Fractional-order PID controller for first-
& Hayat T, Modeling and control with neural networks for a order plus time-delay systems with model predictive control,
magnetic levitation system, Neurocomput, 227 (2017) 113– IEEE Trans Syst Man Cybern, 52(8) (2022) 1334–1342.
121, [Link] 27 Liu Z, Zhang X & Wang Z, Reinforcement learning tuned PI
14 Rubio J J, Discrete time control based on neural networks for controller for two tank interacting hybrid system, IEEE
pendulums, Appl Soft Comput, 68 (2018) 821–832, Trans Ind Electron, 68(12) (2021) 11087–11096,
[Link] [Link]