Annual Reviews in Control 52 (2021) 65–74

Contents lists available at ScienceDirect

Annual Reviews in Control

journal homepage: www.elsevier.com/locate/arcontrol

Review article

The dilemma of PID tuning✩

Oluwasegun Ayokunle Somefun a ,∗, Kayode Akingbade b , Folasade Dahunsi a
a
Department of Computer Engineering, Federal University of Technology Akure, Nigeria
b
Department of Electrical and Electronics Engineering, Federal University of Technology Akure, Nigeria

ARTICLE INFO ABSTRACT

Keywords: A lot of automatic feedback control and learning tasks carried out on many dynamical systems still funda-
PID control mentally rely on a form of proportional–integral–derivative (PID) control law. The PID law is often viewed as
PID tuning a simplistic computational control algorithm. However just like all non-convex optimization problems, tuning
Optimization
the PID algorithm for accurate and stable closed-loop control becomes a NP-Hard Problem. This leads to a
Dynamical system
dilemma, for both users and designers, most especially in practise. It is then no wonder that tuning software is
Settling-time identification
System identification
a big business in the industrial automation sector. In this review, we present and classify PID tuning methods
till date into three general areas. Finally, we then present a proposal to minimize the dilemma of complexity
and cost that has become associated with tuning the three main parameters of the PID control law. Hopefully,
continuous attempts at the minimization of this dilemma can lead to both a money-savings investment and a
significant improvement in the field of PID control design.

1. Introduction Proportional–Integral–Derivative (PID) control (Abramovitch, Hoen, &

Workman, 2008; Guo, 2020).
The controller in software is known as a control algorithm, and On this subject of PID control design, the recent book (Wang, 2020)
it is the brain of an automatic control system. Control system design is a recommended text for beginners. A discussion on PID tuning has
theory is the special theory that studies how control algorithms are been given in a recent review by Borase, Maghade, Sondkar, and
systematically designed and how to ensure their stable and accurate Pawar (2020), where the authors classify PID tuning into classical and
performance (Guo, 2020). One central concept in control theory is the intelligent methods. In contrast, and more suitably, in this review,
feedback principle. The presence of feedback realizes a closed loop. PID control design (tuning) methods are presented from the design
The main aim of feedback is to deal with the influence of uncer- perspective of being either plant-model based, plant-model free or an
tain internal system and external forcing factors on the stability and hybrid of both perspectives. Specifically, the focus of this paper is solely
accuracy performance of the measured output(s) of a system, com- to give a concise review of the complexity and cost dilemma in tuning
monly called the process variable. Consequently, a central motivation the PID control law, as observed in these approaches. This has led to
in controller design is the need for guaranteed stable and accurate
a continuous, unrelenting search for novel and better methods than
closed-loop performance of dynamical systems, especially in motor-
is currently available in theory for practical purposes. As emphasized
control tasks (See Figs. 1 and 2). Closed-loop performance in terms of
in Díaz-Rodríguez, Oliveira, and Bhattacharyya (2015) and Koelsch
both accuracy and stability is therefore critical. In this case, a well-
(2014), despite the boom of industrial tuning software, most real-
designed feedback controller then becomes a real money-savings and
time PID controlled loops in operation still behave poorly after some
life-savings investment (Yang, 2020). In other words, both optimality
time, and need retuning, often increasing the costs of regulatory-loop
and more importantly robustness are important (Keel, & Bhattacharyya,
maintenance. It is, then, no wonder that recently Guo (2020) argues the
2016) in the face of uncertainty. Using machine learning lingo, what
this means is that: compared to a controller overfitting on control need to review PID control design theory. Also, Aström, and Hägglund
tasks at a particular period of time, generalization of the controller (2018) has suggested the need for more extensive work on automatic
to control tasks at all time intervals is more important. Feedback is adaptive PID tuning algorithms.
therefore, important in automation and intelligent systems. In practise, In the next sections that follow, we will focus on the PID law, its
a lot of these automation tasks still fundamentally rely on a form of tuning dilemma (costs and complexity) and then succinctly discuss the

✩ This research received no specific funding.

∗ Corresponding author.
E-mail address: [email protected] (O.A. Somefun).

https://doi.org/10.1016/j.arcontrol.2021.05.002
Received 16 February 2021; Received in revised form 13 May 2021; Accepted 16 May 2021
Available online 29 June 2021
1367-5788/© 2021 Elsevier Ltd. All rights reserved.
O.A. Somefun et al. Annual Reviews in Control 52 (2021) 65–74

the three PID contributing terms (or kernels) 𝑢𝑝 , 𝑢𝑖 , 𝑢𝑑 ∈ R are related

to the error terms 𝑒𝑝 , 𝑒𝑖 , 𝑒𝑑 ∈ R respectively. 𝐾𝑝 ∈ R is the proportional
gain, 𝑇𝑖 ∈ R is the integral time-constant, 𝑇𝑑 ∈ R is the derivative
time-constant, 𝑒 ∈ R is the full error signal.
Modern PID Formalism. Considering that a defining factor on
the natural perception–action control principle for learning, decision
Fig. 1. Illustration of a Physical Dynamical System (𝑠), where 𝑑 and 𝑛 are the and prediction in living organisms is criticism. In a recent disserta-
respective input and output disturbance and uncertainty lumped into the system. tion (Somefun, 2021), the PID law (1) was redefined in a critic form
(6).
( )
𝑢 = (𝑟, 𝑠𝑎𝑡(𝑦)) = 𝐾𝑝 𝜆𝑝 𝑢𝑝 + 𝜆𝑖 𝑢𝑖 + 𝜆𝑑 𝑢𝑑 (6)

where 𝜆𝑝 , 𝜆𝑖 , 𝜆𝑑 ∈ R+ are the critic weights on each of the three PID

contributing terms related to the error terms respectively, such that
𝜆𝑝 ≡ 1, and 0 < 𝜆𝑖 ≤ 1, 0 < 𝜆𝑑 ≤ 1 represents the degree of belief
(between 0 and 1) on each of the three PID contributing terms. The
introduced critic weights serve as a belief-basis for discriminating the
output integral and derivative contributions. This critic formulation (6)
differs from the classic definition which assumes that the PID’s three
contributing terms equally dominate the output decision of the PID
function, that is 𝜆𝑝 , 𝜆𝑖 , 𝜆𝑑 ≡ 1.
Motivations and Successes. Specifically, it can be stated that any
Fig. 2. A PID controlled loop or closed-loop system.
controller or learning structure involving error feedback is a form of
the PID law. This PID control law (1) is, both in theory and practise,
specially regarded as ubiquitous (Aström & Hägglund, 1995; Johnson
& Moradi, 2005; Samad, 2017; Wang, 2020). A recurring motivation
three categorized perspective of PID tuning methods with respect to
for this is because, although an apparently simple, yet powerful mech-
the dependence of a plant model and the explicit use of numerical
anistic representation of the universal error feedback law (Silva, Datta,
optimization techniques. A major pre-requisite for systematic control & Bhattacharyya, 2005; Vagia, 2012), PID control applied to many
design is that some form of model is always needed for guaranteed con- practical problems continues to provide simple, least-cost, and at-least
trol (Campestrini, Eckhard, Bazanella, & Gevers, 2017; Gevers, 1996; satisfactory robust control performance (Bucz, & Kozáková, 2018; Li, &
Van Den Hof, & Schrama, 1995). Subsequently, we conclude this review Wang, 2016; Peretz, 2018; Wang, 2020; Yu, 2018).
by proposing a theory to streamline the underscored costs and com- More recently, the PID law has been applied as an optimizer in the
plexities of the time-consuming model identification and computational weight-training of deep neural networks (Wang et al., 2020). It is often
expensive optimization involved with PID tuning. In brief, we argue represented or designed as a continuous-time system but implemented
that the information of time, specifically the closed-loop input–output as a discrete-time system (that is in software) (Wang, Han, Liu, & He,
settling-time can be a sufficient system dynamics property for modelling 2020). In retrospect, the PID law is simply an algorithm or a computer,
feedback systems with a guaranteed finite settling-time. Therefore, that can be realized as software or as an hardware.
for designing the parameters of the PID control-law, the develop- The PID controller’s behaviour is such that it assumes the presence
ment of automatic and robust methods for this type of settling-time of measured or estimated output state variables (output and its deriva-
identification becomes an open problem. tives) for the control of one output in a system. Robust control theories
like Time-Delayed Control (TDC) (Roy, & Kar, 2020) and Uncertain
2. PID control law Disturbance Estimation (UDE) (Zhong, Kuperman, & Stobart, 2011)
have been shown to reduce to a specific form of PID control (Chang,
Seminally, the work of Minorsky (1922) realized the PID law as a & Jung, 2009; Nam, 2016). These robust control theories have further
computational mimicry of the natural perception–action control princi- shown that the PID control structure is generally an implicit closed-
ple deployed by skilled helmsmen. The PID law’s structure for control loop model reference structure for the control of systems with unknown
is to use the past, present and estimated future error information (Ben- (uncertain) nonlinear dynamics. It has also been noted in Aström and
nett, 2001). Today, one general computational representation of the Hägglund (1995) that PID control appears to follow an implicit model.
PID law is a two-degree-of-freedom (2DOF) control structure (Abdelaty, The use of such closed-loop adaptive model reference architectures
Ahmed, & Ouda, 2018; Araki, & Taguchi, 2003; Aström, & Hägglund, proposed to deal with parametric uncertainties for guaranteed transient
1995; Viteckova, & Vitecek, 2015; Wang, 2012) aimed at simultaneous properties has been treated in Gibson, Annaswamy, and Lavretsky
disturbance-rejection and set-point tracking. Such 2DOF structure with (2013) and Gibson, Qu, Annaswamy, and Lavretsky (2015).
respect to the proportional and derivative set-point weights 𝑏 and 𝑐 Although, historically, the PID controller was first mechanically
∈ [0, 1] ∈ R+ is implemented by means of gears (Bennett, 2001; Nicolai, 1922), it is
( ) not a souvenir of the past century of engineering (Aström & Häg-
𝑢 = (𝑟, sat(𝑦)) = 𝐾𝑝 𝑢𝑝 + 𝑢𝑖 + 𝑢𝑑 (1) glund, 2018). Many researchers and authorities have however expected
where control design advances like fuzzy logic, neural network, model pre-
dictive control (MPC), dynamic matrix theory, reinforcement learning,
𝑢𝑝 = 𝑒𝑝 , 𝑢𝑖 = 𝑇𝑖−1 𝑒𝑖 , 𝑢𝑑 = 𝑇𝑑 𝑒𝑑 (2)
dynamic programming, to render the PID obsolete (Aström & Hägglund,
∗
𝑒𝑝 = 𝑏 𝑟 − 𝑦 (3) 2018; Johnson & Moradi, 2005). Analysing the state of optimal and
( ) robust control, this point is stated clearly in Keel and Bhattacharyya
𝑒𝑖 = 𝑟 − 𝑦∗ 𝑑𝑡 = 𝑒 𝑑𝑡 ≡ 𝑠−1 𝑒 (4)
∫ ∫ (2016) that in the last six (6) decades, the control community’s success
( ) has been limited in the direction of obtaining robustness through
𝑒𝑑 = 𝑐 𝑟̇ − 𝑦̇∗ ≡ 𝑠 𝑐 𝑟 − 𝑦∗ (5)
quadratic optimization. One core reason given and proved in Keel, and
where 𝑢∗ = sat(𝑢) and 𝑦∗
= sat(𝑦) are the saturated input and output of Bhattacharyya (1997) and Keel and Bhattacharyya (2016) is that such
an illustrative single-input single-output system respectively, such that design methods lead to high order, fragile controllers.

66
O.A. Somefun et al. Annual Reviews in Control 52 (2021) 65–74

For reliability, the lowest level of all complex system architectures

is still a PID regulatory loop (Katebi, 2007). Industry assessments still
view the PID as a important core in loop regulation and PID tuning as
an important problem (Koelsch, 2014; Samad, 2017). One reason for
this, may be that the PID law is a representation of the natural law
of feedback. Despite the nonlinear and varying dynamic characteristics
of many physical processes, a well tuned PID gives a good tradeoff
between robustness and performance. The PID currently enjoys popular
Fig. 3. General classification chart of PID tuning methods.
use in the chemical, food, oil and gas processing, automotive, motion,
electronic industries and many other fields (Diaz-Rodriguez, Han, &
Bhattacharyya, 2019). It is widely respected in the industry and a good
percentage of researchers in the control-theory community have in to equivalent higher integer-order linear filter or transfer-function rep-
more than the last three decades contributed to research on PID control resentations (Dastjerdi et al., 2019). This then begs the question, of its
theory and industrial engineering practise (Datta, Ho, & Bhattacharyya, claimed advantage over integer-order representations.
2013). Furthermore, some questions remain as regards the comparative
performance analysis of tuned fractional PID representations to integer-
However, the success recorded by the PID law as a controller is over-
order representations, as shown in works like that of Koszewnik,
shadowed by certain real-time design and performance issues. One, a
Pawłuszewicz, and Ostaszewski (2021). Most tend to use a sub-optimal
large percent of PID controllers in operation are in manual. Two, from a
tuning method to find the PID parameters for the integer-order forms.
study carried out by ABB, most of the PID loops operating in automatic
Is it possible that the use of other tuning methods could have provided
behave poorly after some time than in open loop operation (Koelsch,
similar or better performance compared to the fractional forms, as
2014). Although, there is also some knowledge gap between industrial
reported in the literature, such as in Bingi, Ibrahim, Karsiti, Hassan,
practices and theoretical advances, these performance issues can either and Harindran (2018)? This question is important, as the search space
be traced to being a result of poorly chosen operating PID parameters for finding even the best three gain terms of integer PID forms is an
which are products of the selected PID tuning method or traced to NP-hard problem.
adversarial variations (uncertainty) in the dynamics of the system Again, it is possible to obtain better performance from PID con-
under control from the identified nominal models used for control trollers by considering the PID representation from a signal process-
design (Diaz-Rodriguez et al., 2019; Koelsch, 2014). ing perspective of filtering. The integral and derivative terms can then
Meanwhile, for context, before we discuss these problems on the represent integer-order filter systems of higher order integrals and
design of the PID law, we need to consider the mathematical language derivatives (Note that the benefit of higher order derivatives greater
of differential equations used to model the PID law as an artificial repre- than order 2 is not always clear in terms of control performance), a
sentation of natural human perception–action negative feedback-based notion close to that proposed in (Huba, 2015; Huba, & Vrančič, 2018).
intelligence. However, as fractional representations are always realized as linear
high dimensional integer-order filter transfer-functions, they have the
3. Representations: Integer order or fractional order advantage that the increasing amount of filter parameters is connected
to basically just two fractional parameters which can be tuned more
Differential equations (DEs) are used as the universal computational easily. To conclude, currently, control engineers are still faced with a
language for the formal description of dynamical systems (Li, Zheng, & dilemma in choosing fractional representations over integer-order rep-
Wang, 2019). The general form of the conventional integer-order DE is resentations of the PID. For real-time controller realization in software
the fractional-order DE. On this subject, there is an ever-growing body for fast algorithmic execution and safety-critical control purposes, the
of research works, particularly in the control literature on fractional- use of integer-order representations remains the dominant choice, as it
order PID (FOPID) controller representations as opposed to the integer- reduces the PID’s computational and tuning complexity.
It was Wiener (1950) who theorized that: nature and specifically,
order representation of the complex Laplace variable given in (4) and
humans use a negative-feedback, retuning (adjustment) process to gain
(5). As proposed by Podlubny (1999), in the fractional from, these
new knowledge and achieve goals. This way, feedback control theory
equations become
can also be viewed as learning theory or the science of intelligence
𝑒𝑖 = 𝐃−𝜇 𝑒(𝑡) ≡ 𝑠−𝜇 𝑒 (7) systems (Bhaya, & Kaszkurewicz, 2007; Hershberger, 1990; Poljak, &
( ) ( ) Tsypkin, 1973; Tsypkin, 1971). Arguably, the most successful applica-
𝑒𝑑 = 𝐃𝛿 𝑐 𝑟 − 𝑦∗ ≡ 𝑠𝛿 𝑐 𝑟 − 𝑦∗ (8) tion of Wiener’s cybernetic theory of artificial intelligence goes to the
where the fractional variables 𝜇, 𝛿 ∈ R+ . The growing interest in PID algorithm, used in different forms as gradient-descent optimization
fractional representations, has been attributed to the extra degree of algorithms (Hu, & Lessard, 2017; Lessard, Recht, & Packard, 2016;
tuning freedom, allowing for frequency response shaping, given by Recht, 2019; Wang et al., 2020; Zweiri, Seneviratne, & Althoefer,
the fractional variables (Dastjerdi, Vinagre, Chen, & HosseinNia, 2019; 2005) for learning (optimizing or controlling) the weights of artificial
Tepljakov et al., 2021; Vinagre, Monje, Calderón, & Suárez, 2007). neural networks, based on the reconstruction errors of the neural
network (LeCun, Bengio, & Hinton, 2015; Schmidhuber, 2015; Werbos,
Therefore it is often claimed that using FOPIDs give rise to superior
1975). However, as elegant as the representations of the PID are for
control loops. However, there are certain aspects left to be critically
mimicking intelligence, unfortunately, certain difficulties arise in its
addressed (Tepljakov et al., 2021). One, it is still difficult to concretely
design for control, also known as tuning. These problems has made it
validate the cost–benefit advantage of using fractional representations
that no one method can be said to be universal or best. To date, we are
for real-time PID control, as they are more computationally expensive
still lacking knowledge on the best way to achieve a universal optimal
in terms of memory requirements and speed of algorithm execution on
solution to the tuning design of the PID law, whether as a controller or
a computer chip (Petráš, 2012), hence subtly introducing delays in the an optimizer.
control loop. This complexity of both tuning and implementation has
made its adoption not widespread in the control engineering indus- 4. PID tuning problem
try (Dastjerdi, Saikumar, & HosseinNia, 2018; Dastjerdi et al., 2019;
Vinagre et al., 2007). It remains to see if in the future such bottlenecks Interestingly, one would have thought that the seminal work
can be removed. Two, in the literature, real-time implementation of in Ziegler, and Nichols (1942) on Optimum settings for automatic con-
fractional integrals and derivatives are always usually, approximated trollers would be the final and universal answer to PID tuning. On the

67
O.A. Somefun et al. Annual Reviews in Control 52 (2021) 65–74

contrary, significantly, the design problem (tuning) of PID controllers 5. Plant-model based methods
has attracted unceasing theoretical and practical research interest (As-
tröm & Hägglund, 2018; Vilanova, & Visioli, 2012; Wang, Yan, Li, Notably, a greater percentage of PID control design methods (fixed
& Sun, 2018). To achieve best performance in terms of stability and and adaptive) available are highly dependent on the knowledge of de-
accuracy, the single best control design solution to tuning a closed- rived mathematical model approximations from first principles or fitted
loop system controlled by a PID (which can be termed a closed PID-loop) mathematical model approximations of the actual underlying physical
remains an open question in many applications (Guo, 2020; Saab, 2017; dynamical system from experimental data (Astrom, & Hagglund, 2006;
Smithm, 2018a). One reason for this is the absence of perfect mod- Aström, & Hägglund, 2017; Ellis, 2012; O’Dwyer, 2009).
els for describing real-world systems which are inherently uncertain In modern terms, considering trade-offs between performance and
systems (Ang, Chong, & Li, 2005; Dastjerdi et al., 2018; Jantzen, & robustness (Garpinger, Hägglund, & Åström, 2014; Grimholt & Sko-
Jakobsen, 2016; Sung, Lee, & In-Beu. Lee, 2009). gestad, 2018; Segovia, Hägglund, & Åström, 2014), the identified pa-
The systematic selection of the optimum set of PID parameters
rameters of these plant models and structures are explicitly used as
have therefore been categorized as an NP-hard problem in terms of
design constants to set the PID parameters (Levine, 2011; Li & Wang,
complexity (Koszaka, Rudek, & Pozniak-Koszalka, 2006). This implies
2016; Skogestad, 2006; Sung et al., 2009; Visioli, 2006). Corresponding
there is no unique solution. It also means that tuning for guaranteed
adaptive methods achieve parameter estimations using derived plant
accurate and stable control performance can be burdensome even for
model parameters. For LTI (Linear Time-Invariant) MIMO systems, such
very common servomechanism applications (Grimholt, & Skogestad,
tuning methods proposed to guarantee robust closed-loop stability can
2018; Killingsworth, & Krstic, 2006; Roux-Oliveira, Costa, Pino, & Paz,
2019). be found in Cao, and Chen (2014), Gundes, and Ozguler (2007) and
Indeed, the literature as evident in Vilanova and Visioli (2012) on Saab, and Toukhtarian (2015). In this category, some methods like
PID design (systematic tuning) for single-input single output (SISO) and in Reynoso-Meza, Garcia-Nieto, Sanchis, and Blasco (2013) employ
multiple-input multiple-output (MIMO) systems is rich and extensively multi-objective optimization for tuning. These methods generally in-
wide. These methods use one or more of classical control theories volve minimization of some performance index. The use of Lyapunov
or optimal, robust and adaptive control theories, which are subsets stability and linear matrix inequalities (LMI) analysis also form the
under nonlinear control theory (Silva et al., 2005; Slotine, & Li, 1991). basis of some other tuning methods (Mendoza, Zavala-Río, Santibáñez,
Earlier surveys of fixed and adaptive PID tuning methods have been & Reyes, 2015).
presented in Aström, Hägglund, Hang, and Ho (1993), Cominos, and Other modern methods in the category mix of plant-model knowl-
Munro (2002) and Koivo, and Tanttu (1991). Similarly, in Aström and edge and optimization objectives include: the general model predictive
Hägglund (2018) a more recent overview on the place of PID control control (MPC) (Klaučo, & Kvasnica, 2019), H-infinity (Diaz-Rodriguez
design as an active research in adaptive control (Black, Haghi, & Ariyur, et al., 2019), quantitative feedback theory (QFT) (Comasòlivas, Esco-
2014) and the need for automatic tuning algorithms instead of simple bet, & Quevedo, 2012; Mercader, Åström, Baños, & Hägglund, 2017)
rules was presented. and linear quadratic Gaussian (LQG) control methods.
A major pre-requisite for systematic control design is that some As most engineered dynamical systems are cyber–physical systems
form of model is always needed for guaranteed control (Campestrini with actually complex, unknown dynamics and large time-varying time
et al., 2017; Gevers, 1996; Van Den Hof & Schrama, 1995). This is constants (Wang, Davison, & Davison, 2013). From a linear dynami-
known as system identification (Aström, & Wittenmark, 1971; Gevers, cal systems perspective, a plant-model based approach by system (or
2005; Ljung, & Vicino, 2005), where a computational model is ex- process) identification, such as the first-order plus time delay model
plicitly identified either by first principles of physical laws or either is useful and more so, reasonable in some sense (Bai, & Roth, 2019;
explicitly or implicitly from the experimental input–output data of a Forrai, 2013; Lynch, Marchuk, & Elwin, 2016; Samad, 2017; Samad,
dynamic system, also known as training data in the machine-learning & Annaswamy, 2013). Notwithstanding, these models are imperfect
literature (Ljung, Chen, & Mu, 2020; Pillonetto, 2018). This aim of
representations of the physical systems, which although may operate
identifying models from data apart for control purposes may be for
in a linear sense, are actually both non-linear and uncertain (Guo,
simulations. The practice of system identification has therefore shaped
2020). Therefore, these plant-model based approaches may not fully
the way controllers are designed in order to achieve certain closed-loop
exploit the best quality of control accuracy using the PID. According
performance metrics.
to a ASEA Brown Boveri (ABB) survey in Koelsch (2014), after some
In this context, the methods of tuning PID controllers can there-
time these tuning methods, which are commercially available become
fore be categorized (See Fig. 3.) into three general perspectives based
unreliable. Also, it has consistently been reported in the case of process
on the concept of a model for describing or representing either the
closed-loop system or the open-loop system to be controlled. Namely: industries, that such identification, although useful for testing, can be
Plant-model based methods, which require identification of a plant- a time consuming and expensive process (Jantzen & Jakobsen, 2016;
model structure and knowledge of its parameters for tuning. These Roux-Oliveira et al., 2019).
methods may or may not be data-driven; Plant-model free methods, A summary list of the general shortcomings associated with plant-
which are purely data-driven approaches with optimization for tuning; model based methods include:
and Hybrid methods, which combine a kind of priori plant-model
1. imperfect models compared to the actual uncertain, complex
knowledge with data-driven approaches for tuning.
plant dynamics
Also, it should be noted that tuning can be done offline or online.
2. significant retuning may be frequently needed
When tuning is done online, we mean that the controller is tuned while
3. time-consuming identification process
it is providing the control function in a closed-loop. When tuning is
done offline, we mean that the controller is tuned while it has not 4. usually fixed tuning rules, but most modern methods make
yet started providing the control function in a closed-loop. Further, it heavy use of numerical optimization techniques
should be noted that controllers are tuned on the fundamental basis
that power supply to the physical dynamic system (including the chip 6. Plant-model free methods
hosting the controller, its actuators and sensors) is averagely always
constant and reliable. In practise, the degradation of power supply such Therefore, as the challenges of plant-modelling being time-
as in battery-powered systems, means set-points cannot be reached, consuming, costly and complex or good plant models being math-
disturbances cannot be rejected and hence failure in control. ematically unavailable surrounds process identification for control.
In detail, we continue the discussion of these categorical methods This has motivated interest in design methods that are not plant-
and their shortcomings in the following sections. model based. Deviating from model-based conventions are model-free

68
O.A. Somefun et al. Annual Reviews in Control 52 (2021) 65–74

methods, which can essentially be categorized as purely data-driven 7. Hybrid methods

control design methods.
In this case, what is often meant, is the absence of a mathematical Hybrid methods cover methods that combine both the use of a form
model of the plant for tuning the controller parameters and that under of plant-model knowledge (not necessarily a parametric plant-model
certain assumptions, there is no explicit identification of the parameters structure) with data-driven plant model-free approaches for tuning.
of a certain model structure from data. In this category, there are methods, that depending on the way
These methods for tuning PID gains are usually based on a reference control techniques (optimal control methods, adaptive control meth-
model from closed-loop input–output data experiments and minimiza- ods, pole-placement methods, computational-intelligence methods) are
tion of some objective function (or performance index). A notable used, could be classified as using both or one of plant-model based and
feature of plant model-free methods, is that they realistically assume plant-model free characteristics for tuning.
that the true system is inherently nonlinear and uncertain.
Furthermore, there are other cases, where, despite being signifi-
Among them, one popular technique is the Relay Feedback Test
cantly plant-model free, some form of implied model or characteristic of
(RFT) (Boiko, 2013), specifically the Relay Auto-tuning method (As-
the system from observed data is used to directly tune the PID controller
tröm & Hägglund, 2018). Relay Auto-tuning, uses modified Ziegler–
parameters, whether adaptively or in a fixed form.
Nichols (Z–N) tuning rules to automate the frequency response identi-
In the first case, a lot of recent works on PID tuning design employ
fication of a dynamical system (gain and phase margin information).
optimal control techniques, and tend towards mixing different robust
Although, most recent variants involve multi-objective optimization,
design objectives, such as specification of gain and phase margins, by
the RFT method still involves some form of process identification and
sometimes the use of plant-model knowledge. Primarily, this approach employing constrained optimization and multi-objective optimization.
relies on describing nonlinear function analysis, and so frequency limit Particularly, considering the attenuation of measurement noise (filter-
cycles are common. The accuracy of the identified critical frequency ing) and time-delay, different tuning approaches and studies such as
point of oscillation is also still a subject of active research (Hornsey, in Larsson, and Hägglund (2011), Mercader et al. (2017), Sekara, and
2012; Zeng et al., 2019). Matausek (2009) and Veronesi, and Visioli (2018) have been carried
Particularly notable in the model-free category, are gradient-based out.
design methods which explicitly minimize a cost function (performance- Again, a notable method in this category, is the dominant eigen-
index) for real-time optimization, such as the unfalsified method, value assignment, crudely known as pole-placement methods. Such ap-
iterative feedback tuning (IFT) (Ho, Hong, Hansson, Hjalmarsson, & proaches, of which a greater percent are plant-model based, are treated
Deng, 2003), iterative learning control (ILC) (Moore, & Xu, 2000), in works such as Datta et al. (2013), Diaz-Rodriguez et al. (2019), Han,
and extremum-seeking (ES) method summarized in Killingsworth and and Bhattacharyya (2018), Keel and Bhattacharyya (2016), Srivastava,
Krstic (2006). In Paz, Oliveira, Pino, and Fontana (2020) and Roux- and Pandit (2016), Wang, Han, Liu, and He (2020) and Zítek, Fišer, and
Oliveira et al. (2019), through ES, the gains of the PID control law were Vyhlídal (2013). An attracting point of this method, is that performance
adaptively tuned for functional neuromuscular electrical stimulation, specifications such as stability and dynamic performance directly corre-
specifically to control and coordinate the positioning of the arm–elbow late to the assignment and distribution of dominant closed-loop eigen-
joints of stroke patients. The assumption in this case was that the values (poles). The added use of multi-objective optimization is also
neuromusculoskeletal system is input-to-state stable (ISS). This implies very common in some of these approaches.
that for any bounded control input, both the state and output variables In addition, more recently, there is a trend, observed in works
are also bounded. such as Almabrok, Psarakis, and Dounis (2018), Ekinci, and Hekımoğlu
On the other hand, for the IFT, usually multiple offline closed loop (2019), Ekinci, Hekımoğlu, and Izci (2021), Hekimoğlu (2019), Izci,
experiments have to be carried out (Lequin, 1997) to estimate the ob- and Ekinci (2021) and Mandava, and Vundavilli (2019) which pro-
jective function’s gradient. The ILC approach leverages repetition, but pose metaheuristic (evolutionary) optimization algorithms as new PID
to generally achieve perfect tracking performance, it requires physical tuning methods. These methods consider multi-objective optimization,
systems to satisfy an identical initialization condition, which is a very using evolutionary optimization and specification of performance con-
restrictive condition (Guan, Zhu, Wang, & Liu, 2014; Preitl et al., 2007).
straints such as stability, bandwidth, and robustness. Although, these
Another is Virtual Reference Feedback Tuning (VRFT) (Cao & Chen,
methods are often posed as model-free methods, in their design for
2014; Formentin, Savaresi, & Re, 2012) or fictitious reference iterative
control purposes, they still often require the need of a plant model of
tuning (FRIT) (Ashida, Wakitani, & Yamamoto, 2017; Kano et al.,
the system to be controlled for simulation purposes. At the basic level,
2010; Tasaka, Kano, Ogawa, Masuda, & Yamamoto, 2009; Wakitani,
they can be viewed as stochastic search methods. Although promising,
Yamamoto, & Gopaluni, 2019). An alternative model-free learning (op-
they have not found wide use in the process industry, since any serious
timization) framework to ES and IFT for LTI MIMO systems (assumed
uncertainty in the parameters of the system can lead to instability, and,
unknown) was proposed in Wang et al. (2013), which also involves
in some cases, the converged solutions are less than optimal (Dastjerdi
iteratively carrying out closed-loop experiments. There is still however
no guarantee that the tuned gains will always approach the optimum. et al., 2018; Li et al., 2019). Also, sometimes, before they are executed,
Likewise, a summary list of the general problems associated with they depend on some initial settings from another tuning method as
plant-model free based tuning methods include: good starting-points for optimization. Some recent attempts at practical
implementations of such techniques can be found in Ekinci, Izci, and
1. possess restrictive design conditions for systems in practise. Hekımoğlu (2021) and Chen, Bowels, Zhang, and Fuhlbrigge (2019).
2. more computationally expensive, and sometimes unreliable In the second case, one notable method, was proposed in a study
error-minimization optimizations on data. by Jantzen and Jakobsen (2016). Originating from fuzzy control study,
3. time-consuming data-driven experiments to ensure data quality the central argument in this approach is that the time-consuming
4. use of, sometimes, complicated objective-function formulations system identification of plant-models for PID tuning can be reduced
5. since no explicit model is used, some trial and error values, or to considering the physical meaningful performance specification of
simulations may be needed to arrive at satisfactory settings. desired settling-time based on operator knowledge and system oper-
6. nominal control is only as good as data, data may not cap- ating behaviour. Also, more recently, Zhong, Huang, and Guo (2020)
ture all actual operating conditions, so actual control may be proposed a tuning method based on a parametric bandwidth formula
sub-optimal. of the extended state observer (ESO), which uses pole-placement and
7. may require a try-test-adjust approach in setting the values of model-reference following paradigms to connect the active disturbance
certain hyperparameters used for tuning. rejection control (ADRC) concept to design the PID law.

69
O.A. Somefun et al. Annual Reviews in Control 52 (2021) 65–74

The use of explicit reference models, is common in tuning methods of PID gains to tune as the number of controlled output variables
which use ideas of model following adaptive control. Examples are increase. For instance, in the control of a system with 6 motors, this
direct (or indirect) model reference adaptive methods and existing would mean simultaneous tuning of at least 18 parameters or 24
self-tuning methods (Aggarwal, & O’Reilly, 2006; Bobál, Böhm, Fessl, parameters in the case of added filter constants. It is believed that this
& Machácek, 2005). Most of these methods usually suffer from poor complexity can be reduced to some minimum level. As such, adaptive
transient performance or (and) may involve some expensive online but robust PID control design will continue to attract research interest.
parameter optimizations and estimations. Specifically, alternative tuning frameworks for streamlined control
Furthermore as hybrid methods, tuning approaches that exploit the in the closed-loop data-driven category, without the overhead of time-
artificial intelligence theories of fuzzy-logic, artificial neural-networks, consuming plant-model identification and multi-objective numerical
and reinforcement learning for PID design exist (Jafari, & Dhaouadi, optimization, would help to improve the PID’s widespread adoption
2011; Kofinas, & Dounis, 2019; Malekabadi, Haghparast, & Nasiri, and contribute to better product quality (Chen et al., 2019; Jantzen
2018; Marino, & Neri, 2019; Mendel, 2017; Pirasteh-Moghadam et al., & Jakobsen, 2016), hence reduction in the costs and complexities
2020; Savran, & Kahraman, 2014; Shipman, & Coetzee, 2019; Srivas- encountered with PID tuning.
tava et al., 2018; Taeib, & Chaari, 2015; Wang, Cheng, & Sun, 2007).
These methods have generally been applied to the control of systems 9. Future: Dynamical system description for Control by the Infor-
with the assumption that they are inherently nonlinear. Fuzzy-logic mation of Settling-time
PID control techniques fuzzify the system error states, however, it is
afflicted by time-consuming rule parameter optimization problems (Bai
Accordingly, there needs to be a renewed interest in the search
& Roth, 2019). On the other hand, deep learning techniques involving
for automatic adaptive tuning algorithms that can help streamline the
neural networks and reinforcement learning methods have been noted
costs of PID tuning and reduce the complexities involved. Several rela-
to often give poor results in real-time control practise (Retch, 2020).
tively recent works such as in Bucz and Kozáková (2018), Jantzen and
Again, a summary list of the general problems or characteristic
Jakobsen (2016), Nguyen, and Nguyen (2018), Viteckova and Vitecek
issues associated with hybrid tuning methods include:
(2015) and Wang et al. (2018) have all emphasized the important
1. dominant use of different types of heuristics or optimization consideration of settling-time for both streamlining and improving PID
techniques tuning.
2. inherits some problems from its two basis methods, so tuning In an attempt to resolve this PID tuning dilemma, a novel proposal
solution may be less than optimal and cause instability may be to consider the Closed-PID loop model (CPLM) Following
3. poor transient performance, if real-time Control (CPLMFC) Theory.
4. expensive real-time parameter optimizations, may give poor CPLM. First, the PID law by its structure is a controller with its own
results in practise. internal and implicit closed-loop reference model that can be called the
closed-PID loop model (CPLM), which is the implicit ideal closed-loop
Therefore, having considered these categories, we see a trend in
response that it tries to reference.
the use of numerical optimization techniques in each of these general
Low-frequency Bandwidth. Second, the CPLM can be viewed as a
categories. In the next section, we move on to discuss the effect of
specific case of dominant eigenvalue assignment of complex-conjugate
this from the perspective of control software use-case in real-time, as
poles through the specification of its low-frequency bandwidth.
opposed to simulation time.
Closed-loop Settling-time. Third, the closed-loop settling-time in
connection to the low-frequency bandwidth can be used a sufficient
8. Tuning software and numerical optimization
systems dynamics description for control. This way, settling-time identi-
fication then becomes a robust test for the maximum capability (limits)
In Zhang, and Guo (2019) and Zhao, and Guo (2020), the authors
of a feedback system (Guo, 2020) as observed in the time-domain.
propose new universal theorems for the design of the PID law focused
Interestingly, the frequency (bandwidth) parameter and the finite set-
on dynamical systems as nonlinear and uncertain. The authors highlight
that precise understanding and rationale for the remarkable practical tling behaviour (FSB) (Rao, & Bernstein, 2001) of a dynamical system
effectiveness and capability of PID control is still limited. This view is are connected. Using the low-frequency bandwidth parameter, implicit
affirmed by statistics from industrial process automation, which show dominant eigenvalues of the CPLM can be assigned through the integral
that: although the continued evolution of hardware and software for and derivative(s) time-constants to ensure the actual closed PID-loop
discrete-time control has advanced the use of PIDs, the weakness of system stably and accurately follows the CPLM. The proportional gain
poor tuning is still prevalent (Allan, 2018). Interestingly, a reactive could then be constrained to be an adaptive parameteric estimation
effect of this is the requirement of expensive tuning software. This may of the control input(including disturbances)-to-output gain, ensuring
explain why commercial tuning software is big business in the control global stability and accuracy of the overall feedback system.
industry today (Koelsch, 2014). Therefore, the implication of the CPLMFC Theory is that the PID law
Most auto-tuning software are heavily based on either one-shot or can in real-time adaptively tune itself using the CPLM in connection
real-time plant-model fitting and the use of numerical optimization to an identified settling-time of the closed-loop system. An advantage
techniques. A downside to this approach is that since actual physical of such a tuning framework, is that it requires no knowledge of the
systems deviate from this estimated models. They tend to embody some parameters of a plant (dynamical system) model structure or its order,
form of nonlinearity, noise and unpredictable perturbations. There- and no explicit minimization of an objective-function for optimization.
fore, inaccurate estimations through such tuning software may lead to This theory was recently introduced in a master’s dissertation (Some-
closed-loop performance degradation (Allan, 2018). The effects of such fun, 2021) and applied to the real-time adaptive (online tuning) control
poorly tuned closed PID-loops have been noted to include: increase in of a differential-drive mobile robot. Note that, a preliminary form of
process variability, increased energy consumption, operating costs and the CPLMFC theory was applied in Somefun, Akingbade, and Dahunsi
reduction or unimproved product quality. (2020) for the robust speed-control of dc-motors.
Therefore, despite the reliance on numerical optimization present in However, this brings up the issue of how to reliably identify such a
tuning methods and their implementation as software, many physical closed-loop settling-time in an automatic manner and online manner,
dynamical systems still fall back to being tuned manually or heuristi- directly from data, that is by real-time signal processing, instead of
cally in practise (Smith, 2018b). The complexity of this task is evident manually or offline. Again, with the adaptive nature (online tuning),
in the search-space of numerical optimization functions and the number of such a framework, comes the need for safety guarantees.

70
O.A. Somefun et al. Annual Reviews in Control 52 (2021) 65–74

Therefore, in the CPLMFC framework, the need for a robust online Consequently, the minimum dead-time, in a control loop is the loop’s
settling-time identification process becomes very important. The devel- sampling-time, and the maximum dead-time is the sum of all delay-time
opment of such robust online algorithms and methods for automatic intervals in the loop, from the output to the input. Since, controllers
identification of the closed-loop input–output settling-time as a suffi- sample observed measurements, act, then wait until next sampling-
cient system dynamics property, then becomes a necessary problem to time, and also, communication protocols and scheduling operations of
be solved. the embedded operating system contribute additional layers of dead-
Limits of Feedback and the Information of Time. Again, the time and uncertainty to the overall system. Therefore, the controlled
information of time, has important effects on a closed-loop system. system or process has sampling-times of the controller’s processor and
Since feedback can generally be viewed as a causal function of the communication protocols built naturally into its structure as dead-time
observed output from a dynamical system (Guo, 2002). Its maximum intervals.
performance is constrained to the limitations of the closed-loop’s actu- Fast and slow are relative terms defined by the inherent dynamics
ator, sensor and the time taken to execute the control-input. Therefore, of the system (including its actuator, sensor, and processor). This is
the test for the maximum capability of a dynamical system, is to why selecting a sampling period (not too fast and not too slow) that
simply probe what it cannot do. In the time-domain, the limits of such satisfies the scheduling and control requirements is such an important
performance is the system’s settling-time. In other words, a system and difficult task.
(closed or open) cannot react or stabilize itself faster than its dynamics Sampling slower than the system allows, will have a negative im-
allow. An attempt to violate this, especially in practise, often leads pact on performance, like loss of data fidelity, introduction of extra
to excessive oscillations and overshoots in the time-response of such dead-time. Sampling faster than the system allows, will have a nega-
systems.
tive impact on performance, like unnecessary computational overload
Settling-Time. To further understand the above statements, we
and overhead, higher demand on high-end sensing, actuating and
recall that practical, canonical physical dynamical systems are inher-
computing resources, with no performance improvement.
ently energy dissipating or storage systems (Tewari, 2002). On the
The control engineer would need to choose a sampling time such
application of an arbitrary step input, the output of such systems do
that data can be sampled, processed, and transferred by the software
not instantaneously take a steady-state value, but takes some amount
within the limits of the real time system’s or hardware’s working
of finite time to settle to an average value. This input–output delay-time
frequency. Anything too more or less will degrade the stability margins
interval to settle to an average value is a characteristic property of all
(and response) of the system.
controllable and stable dynamic systems. This finite input–output delay
Therefore, the achievable accuracy and stability of a PID controlled
time interval, can be regarded as the settling-time, being the aggregated
time delay dynamics that captures the input–output transition of such physical dynamical system is strongly related to time. Actuating faster
systems. or slower than the maximum time (or bandwidth) dynamics generating
Ultimately, the settling-time embodies the observable performance a signal leads to false reconstruction or observations and hence poor
indices of a control system’s reliability, which is in the time-domain control. This is also a dilemma (Aström, & Wittenmark, 2013; Bai &
(that is: how rapidly a fluctuating measured process output takes to Roth, 2019; Franklin, Powell, & Emami-Naeini, 2019) especially for
settle, and with what amount of maximum overshoot the measured tuning the PID controller in real-time. One recommended best practice,
output takes to settle to become exponentially close (accurate) in value is that the control loop’s sampling-time should be at least 10 times
to the desired process output or goal) (Atherton, 2015). faster than the sampling-time of the measured process. There is still
Dead-Time. Also, the interval of time known as settling-time is a no universal rule, only heuristics like (9).
superset of the dead-time (transport-lag) known by control researchers The relation of the information of time to tuning a feedback-
and engineers to cause difficulty in PID tuning, and so, degraded, controlled system is therefore, arguably, the most important.
non-smooth, limited regulation of feedback systems.
The first known timing effect, as seen during open-loop operation, is 10. Concluding remarks
the inherent physical constraint that does not allow the observed output
to change from its previous value until some finite period of time after This paper has provided a succinct review of the rich and persistent
the change in the control input. Nothing can be done about this. literature on tuning PID controllers, also known as PID control design.
The second known timing effect, as seen during closed-loop opera- The review started with an introductory discussion on feedback theory
tion, is that it degrades the closed-loop transient, simply because with in relation to the PID control law and its representations. Then the PID
respect to time, the closed-loop is trying to control the observed output tuning problem was discussed next. This led to a general classification
faster than its realistic limits. Targeting dead-time control, there exists of available PID design methods into the following three perspectives:
significant quantity of research since the early periods of the twentieth plant-model based, plant-model free, and hybrid methods. The review
century (Normey-Rico, 2007). also highlighted certain generic problems with the methods discussed,
Finally, the third timing effect, is in the case of the sampling- as encountered in practice when tuning the PID law for controlling
time (Clair, 1993; Kerrigan, 2015; Lequin, 1997) to be discussed next. real-world physical dynamical systems. Next, the connection between
On Sampling-Time. To conclude this section, it seems useful to tuning software and numerical optimization methods was discussed.
have a short note on sampling-time 𝜏𝑠 for control. Finally, closed-loop settling-time identification for control was pro-
Basically, there are at least two sampling-times used in realistic posed as a sufficient dynamical-systems descriptor to replace traditional
control design such as illustrated by (9). Hence, as opposed to most plant-model parametric identification for control design.
simulations, a real-time feedback system is viewed as a multi-rate
Also, we remark that compared to using numerical optimization
discrete-time system. First is the open-loop sampling-time, the rate at
techniques on the PID law, there is a need for the control community to
which actuator (process input) and the sensor (process output) vari-
start viewing the PID as a natural optimization algorithm. In any case,
ables are sampled and measured. Next is the feedback (or control-loop
it remains to see if, in the foreseeable future, there will be a universally
or closed-loop) sampling-time that specifies how often the controller
acceptable adaptive and robust solution to the dilemma faced with the
samples the measured process variable and then computes and trans-
automatic tuning of a PID controller for long-term operation.
mits a new control input signal to the actuator. There is also the
Because of the popularity of the PID in practise, future successful
computational dead-time, which is recommended to be a very small
attempts at resolving its tuning dilemma can lead to significant money-
negligible time.
savings investment and technological revolution in the engineering
𝜏𝑠(𝑦) = 𝑙 𝜏𝑠(𝑢) , where: 𝑙 ≥ 100 (recommended) (9) field of control design, learning theory and optimization.

71
O.A. Somefun et al. Annual Reviews in Control 52 (2021) 65–74

Declaration of competing interest Bucz, Š., & Kozáková, A. (2018). PID control for industrial processes, Advanced meth-
ods of PID controller tuning for specified performance. http://dx.doi.org/10.5772/
intechopen.76069.
The authors declare that they have no known competing finan- Campestrini, L., Eckhard, D., Bazanella, A. S., & Gevers, M. (2017). Data-driven model
cial interests or personal relationships that could have appeared to reference control design by prediction error identification. Journal of the Franklin
influence the work reported in this paper. Institute, 354, 2628–2647.
Cao, Y., & Chen, X. B. (2014). An output-tracking-based discrete PID-sliding mode
control for MIMO systems. IEEE/ASME Transactions on Mechatronics, 19, 1183–1194.
Acknowledgement http://dx.doi.org/10.1109/TMECH.2013.2275743.
Chang, P. H., & Jung, J. H. (2009). A systematic method for gain selection of
This research received no specific grant from any funding agency in robust PID control for nonlinear plants of second-order controller canonical form.
IEEE Transactions on Control Systems Technology, 17, 473–483. http://dx.doi.org/10.
the public, commercial, or not-for-profit sectors.
1109/TCST.2008.2000989.
Chen, H., Bowels, S., Zhang, B., & Fuhlbrigge, T. (2019). Controller parameter
References optimization for complex industrial system with uncertainties. Measurement and
Control, 52, 888–895. http://dx.doi.org/10.1177/0020294019830108.
Clair, D. W. S. (1993). Controller tuning and control loop performance: ‘‘PID without the
Abdelaty, B. G., Ahmed, A. H., & Ouda, A. N. (2018). Fixed set point weighting
math’’:A primer. Straight-Line Control Company.
2DOF PID controller for control processes. Engineering Mathematics, 2(21), http:
Comasòlivas, R., Escobet, T., & Quevedo, J. (2012). Automatic design of robust PID
//dx.doi.org/10.11648/j.engmath.20180201.13.
controllers based on QFT specifications. IFAC Proceedings Volumes, 45, 715–720.
Abramovitch, D. Y., Hoen, S., & Workman, R. (2008). Semi-automatic tuning of
http://dx.doi.org/10.3182/20120328-3-IT-3014.00121.
PID gains for atomic force microscopes. In 2008 American control conference (pp.
Cominos, P., & Munro, N. (2002). PID controllers: Recent tuning methods and design
2684–2689). http://dx.doi.org/10.1109/ACC.2008.4586898.
to specification. IEE Proceedings D (Control Theory and Applications), 149, 46–53.
Aggarwal, V., & O’Reilly, U. (2006). A self-tuning analog proportional-integral-
http://dx.doi.org/10.1049/ip-cta:20020103.
derivative (PID) controller. In First NASA/ESA conference on adaptive hardware and
Dastjerdi, A. A., Saikumar, N., & HosseinNia, S. H. (2018). Tuning guidelines for
systems (pp. 12–19). http://dx.doi.org/10.1109/AHS.2006.12.
fractional order PID controllers: Rules of thumb. Mechatronics, 56, 26–36. http:
Allan, K. (2018). Pros and cons of autotuning—the big story. https://www.controleng.
//dx.doi.org/10.1016/j.mechatronics.2018.10.004.
com/articles/pros-and-cons-of-autotuning-the-big-story/. Dastjerdi, A. A., Vinagre, B. M., Chen, Y., & HosseinNia, S. H. (2019). Linear fractional
Almabrok, A., Psarakis, M., & Dounis, A. (2018). Fast tuning of the PID controller in order controllers a survey in the frequency domain. Annual Reviews in Control, 47,
an HVAC system using the big bang–big crunch algorithm and FPGA technology. 51–70. http://dx.doi.org/10.1016/j.arcontrol.2019.03.008.
Algorithms, 11(146), http://dx.doi.org/10.3390/a11100146. Datta, A., Ho, M.-T., & Bhattacharyya, S. P. (2013). Structure and synthesis of PID
Ang, K. H., Chong, G., & Li, Y. (2005). PID control system analysis, design, and controllers. Springer Science & Business Media.
technology. IEEE Transactions on Control Systems Technology, 13, 559–576. http: Diaz-Rodriguez, I. D., Han, S., & Bhattacharyya, S. P. (2019). Analytical design of
//dx.doi.org/10.1109/TCST.2005.847331. PID controllers. Springer International Publishing, http://dx.doi.org/10.1007/978-
Araki, M., & Taguchi, H. (2003). Two-degree-of-freedom PID controllers. International 3-030-18228-1.
Journal of Control, Automation and Systems, 1, 401–411. Díaz-Rodríguez, I. D., Oliveira, V. A., & Bhattacharyya, S. P. (2015). Modern design
Ashida, Y., Wakitani, S., & Yamamoto, T. (2017). Design of an implicit self-tuning PID of classical controllers: Digital PID controllers. In 2015 IEEE 24th international
controller based on the generalized output. IFAC-PapersOnLine, 50, 13946–13951. symposium on industrial electronics (pp. 1010–1015). http://dx.doi.org/10.1109/ISIE.
http://dx.doi.org/10.1016/j.ifacol.2017.08.2216. 2015.7281610.
Aström, K. J., & Hägglund, T. (1995). PID controllers - Theory, design, and tuning (2nd Ekinci, S., & Hekımoğlu, B. (2019). Improved kidney-inspired algorithm approach
ed.). ISA. for tuning of PID controller in AVR system. IEEE Access, 7, 39935–39947. http:
Astrom, K. J., & Hagglund, T. (2006). Advanced PID control. ISA Instrumentation, //dx.doi.org/10.1109/ACCESS.2019.2906980.
Systems and Automation Society. Ekinci, S., Hekımoğlu, B., & Izci, D. (2021). Opposition based henry gas solubility
Aström, K. J., & Hägglund, T. (2017). Design methods: PID control. In The control optimization as a novel algorithm for PID control of DC motor. Engineering Science
handbook: Control system fundamentals (2nd ed.). (pp. 9–70–9–90). http://dx.doi. and Technology, an International Journal, http://dx.doi.org/10.1016/j.jestch.2020.
org/10.1201/b10383. 08.011.
Aström, K. J., & Hägglund, T. (2018). Advanced PID control. Ekinci, S., Izci, D., & Hekımoğlu, B. (2021). Optimal FOPID speed control of DC
Aström, K. J., Hägglund, T., Hang, C. C., & Ho, W. K. (1993). Automatic tuning and motor via opposition-based hybrid manta ray foraging optimization and simulated
adaptation for PID controllers - a survey. Control Engineering Practice, 1, 699–714. annealing algorithm. Arabian Journal for Science and Engineering, 46, 1395–1409.
http://dx.doi.org/10.1016/0967-0661(93)91394-C. http://dx.doi.org/10.1007/s13369-020-05050.
Aström, K. J., & Wittenmark, B. (1971). Problems of identification and control. Journal Ellis, G. (Ed.), (2012). Control system design guide. Boston: Butterworth-Heinemann,
of Mathematical Analysis and Applications, 34, 90–113. http://dx.doi.org/10.1016/B978-0-12-385920-4.01001-8.
Aström, K. J., & Wittenmark, B. (2013). Computer-controlled systems: Theory and design. Formentin, S., Savaresi, S. M., & Re, L. D. (2012). Non-iterative direct data-driven
Courier Corporation. controller tuning for multivariable systems: Theory and application. IET Control
Atherton, D. P. (2015). Setting the parameters of proportional–integral–derivative Theory & Applications, 6, 1250–1257. http://dx.doi.org/10.1049/iet-cta.2011.0204.
Forrai, A. (2013). Embedded control system design: A model based approach. Heidelberg
controllers. Measurement and Control, 48, 273–277. http://dx.doi.org/10.1177/
; New York: Springer.
0020294015600476.
Franklin, G. F., Powell, J. D., & Emami-Naeini, A. (2019). Feedback control of dynamic
Bai, Y., & Roth, Z. S. (2019). Classical and modern controls with microcontrollers:
systems (8th ed.). Pearson.
Design, implementation and applications. In Advances in industrial control, Springer
Garpinger, O., Hägglund, T., & Åström, K. J. (2014). Performance and robustness trade-
International Publishing.
offs in PID control. Journal of Process Control, 24, 568–577. http://dx.doi.org/10.
Bennett, S. (2001). The past of PID controllers. Annual Reviews in Control, 25, 43–53.
1016/j.jprocont.2014.02.020.
http://dx.doi.org/10.1016/S1367-5788(01)00005-0.
Gevers, M. (1996). Identification for control. Annual Reviews in Control, 20, 95–106.
Bhaya, A., & Kaszkurewicz, E. (2007). A control-theoretic approach to the design of zero Gevers, M. (2005). Identification for control: From the early achievements to the revival
finding numerical methods. IEEE Transactions on Automatic Control, 52, 1014–1026. of experiment design. European Journal of Control, 11, 335–352.
http://dx.doi.org/10.1109/TAC.2007.899109. Gibson, T. E., Annaswamy, A. M., & Lavretsky, E. (2013). On adaptive control with
Bingi, K., Ibrahim, R., Karsiti, M. N., Hassan, S. M., & Harindran, V. R. (2018). closed-loop reference models: Transients, oscillations, and peaking. IEEE Access, 1,
A comparative study of 2DOF PID and 2DOF fractional order PID controllers 703–717. http://dx.doi.org/10.1109/ACCESS.2013.2284005.
on a class of unstable systems. Archives of Control Sciences, 28, 635–682. http: Gibson, T. E., Qu, Z., Annaswamy, A. M., & Lavretsky, E. (2015). Adaptive output
//dx.doi.org/10.24425/acs.2018.125487. feedback based on closed-loop reference models. IEEE Transactions on Automatic
Black, W. S., Haghi, P., & Ariyur, K. B. (2014). Adaptive systems: History, tech- Control, 60, 2728–2733. http://dx.doi.org/10.1109/TAC.2015.2405295.
niques, problems, and perspectives. Systems, 2, 606–660. http://dx.doi.org/10. Grimholt, C., & Skogestad, S. (2018). Optimal PI and PID control of first-order plus
3390/systems2040606. delay processes and evaluation of the original and improved SIMC rules. Journal
Bobál, V., Böhm, J., Fessl, J., & Machácek, J. (2005). Advanced textbooks in control of Process Control, 70, 36–46.
and signal processing, Digital self-tuning controllers: Algorithms, implementation and Guan, W., Zhu, Q., Wang, X.-D., & Liu, X.-H. (2014). Iterative learning control
applications. London: Springer-Verlag. design and application for linear continuous systems with variable initial states
Boiko, I. (2013). Advances in industrial control, Non-parametric tuning of PID controllers: based on 2-D system theory. Mathematical Problems in Engineering, 2014, 1–5.
A modified relay-feedback-test approach. London: Springer-Verlag. http://dx.doi.org/10.1155/2014/970841.
Borase, R. P., Maghade, D. K., Sondkar, S. Y., & Pawar, S. N. (2020). A review of PID Gundes, A. N., & Ozguler, A. B. (2007). PID stabilization of MIMO plants. IEEE
control, tuning methods and applications. International Journal of Dynamics and Transactions on Automatic Control, 52, 1502–1508. http://dx.doi.org/10.1109/TAC.
Control, http://dx.doi.org/10.1007/s40435-020-00665-4. 2007.902763.

72
O.A. Somefun et al. Annual Reviews in Control 52 (2021) 65–74

Guo, L. (2002). Exploring the maximum capability of adaptive feedback. International Lessard, L., Recht, B., & Packard, A. (2016). Analysis and design of optimization
Journal of Adaptive Control and Signal Processing, 16, 341–354. http://dx.doi.org/ algorithms via integral quadratic constraints. SIAM Journal on Optimization, 26,
10.1002/acs.713. 57–95. http://dx.doi.org/10.1137/15M1009597.
Guo, L. (2020). Feedback and uncertainty: Some basic problems and results. Annual Levine, W. S. (2011). The control systems handbook: Control system advanced methods.
Reviews in Control, 49, 27–36. http://dx.doi.org/10.1016/j.arcontrol.2020.04.001. In Electrical engineering handbooks, Engineering handbook (2nd ed.). CRC Press.
Han, S., & Bhattacharyya, S. P. (2018). PID controller synthesis using a𝜎-Hurwitz Li, S., & Wang, J. (2016). Research on engineering tuning methods of PID controller
stability criterion. IEEE Control Systems Letters, 2, 525–530. http://dx.doi.org/10. parameters and its application. In D.-S. Huang, K. Han, & A. Hussain (Eds.),
1109/LCSYS.2018.2842784. Intelligent computing methodologies lecture notes in computer science, (pp. 563–570).
Hekimoğlu, B. (2019). Optimal tuning of fractional order PID controller for DC motor Springer International Publishing.
speed control via chaotic atom search optimization algorithm. IEEE Access, 7, Li, L., Zheng, N., & Wang, F. (2019). On the crossroad of artificial intelligence: A
38100–38114. http://dx.doi.org/10.1109/ACCESS.2019.2905961. revisit to Alan Turing and Norbert Wiener. IEEE Transactions on Cybernetics, 49,
Hershberger, W. A. (1990). Control theory and learning theory. American Behavioral 3618–3626. http://dx.doi.org/10.1109/TCYB.2018.2884315.
Scientist, 34, 55–66. http://dx.doi.org/10.1177/0002764290034001006. Ljung, L., Chen, T., & Mu, B. (2020). A shift in paradigm for system identification.
Ho, W. K., Hong, Y., Hansson, A., Hjalmarsson, H., & Deng, J. W. (2003). Relay auto- International Journal of Control, 93, 173–180. http://dx.doi.org/10.1080/00207179.
tuning of PID controllers using iterative feedback tuning. Automatica, 39, 149–157. 2019.1578407.
http://dx.doi.org/10.1016/S0005-1098(02)00201-7. Ljung, L., & Vicino, A. (2005). Guest editorial: Special issue on system identification.
Hornsey, S. (2012). A review of relay auto-tuning methods for the tuning of PID-type IEEE Transactions on Automatic Control, 50, 1473. http://dx.doi.org/10.1109/TAC.
controllers. Reinvention: an International Journal of Undergraduate Research, 5. 2005.856638.
Hu, B., & Lessard, L. (2017). Control interpretations for first-order optimization Lynch, K. M., Marchuk, N., & Elwin, M. L. (2016). PID feedback control. In Embedded
methods. arXiv:1703.01670 [cs, math], arXiv:1703.01670. computing and mechatronics with the PIC32 (pp. 375–385). Elsevier, http://dx.doi.
Huba, M. (2015). Filter choice for an effective measurement noise attenuation in PI and org/10.1016/B978-0-12-420165-1.00023-8.
PID controllers. In 2015 IEEE international conference on mechatronics (pp. 46–51). Malekabadi, M., Haghparast, M., & Nasiri, F. (2018). Air condition’s PID controller fine-
http://dx.doi.org/10.1109/ICMECH.2015.7083946. tuning using artificial neural networks and genetic algorithms. Computers, 7(32),
Huba, M., & Vrančič, D. (2018). Introduction to the discrete time pidmn control for http://dx.doi.org/10.3390/computers7020032.
the IPDT plant. IFAC-PapersOnLine, 51, 119–124. http://dx.doi.org/10.1016/j.ifacol. Mandava, R. K., & Vundavilli, P. R. (2019). An optimal PID controller for a biped
2018.07.140. robot walking on flat terrain using MCIWO algorithms. Evolutionary Intelligence, 12,
Izci, D., & Ekinci, S. (2021). Comparative performance analysis of slime mould algo- 33–48. http://dx.doi.org/10.1007/s12065-018-0184-y.
rithm for efficient design of proportional–integral–derivative controller. Electrica, Marino, A., & Neri, F. (2019). PID tuning with neural networks. In Intelligent information
21, 151–159. http://dx.doi.org/10.5152/electrica.2021.20077. and database systems (pp. 476–487). Cham: Springer, http://dx.doi.org/10.1007/
Jafari, R., & Dhaouadi, R. (2011). Advances in reinforcement learning, Adaptive PID control 978-3-030-14799-0_41.
of a nonlinear servomechanism using recurrent neural networks. http://dx.doi.org/10. Mendel, J. M. (2017). Uncertain rule-based fuzzy systems: Introduction and new directions
5772/13020. (2nd ed.). Springer International Publishing.
Mendoza, M., Zavala-Río, A., Santibáñez, V., & Reyes, F. (2015). A generalised PID-type
Jantzen, J., & Jakobsen, C. (2016). Turning PID controller tuning into a simple
control scheme with simple tuning for the global regulation of robot manipulators
consideration of settling time. In 2016 European control conference (pp. 370–375).
with constrained inputs. International Journal of Control, 88, 1995–2012. http:
http://dx.doi.org/10.1109/ECC.2016.7810313.
//dx.doi.org/10.1080/00207179.2015.1027272.
Kano, M., Tasaka, K., Ogawa, M., Ootakara, S., Takinami, A., & Takahashi, S. (2010).
Mercader, P., Åström, K. J., Baños, A., & Hägglund, T. (2017). Robust PID design based
Practical direct PID/I-PD controller tuning and its application to chemical processes.
on QFT and convex–concave optimization. IEEE Transactions on Control Systems
In 2010 IEEE international conference on control applications (pp. 2426–2431).
Technology, 25, 441–452. http://dx.doi.org/10.1109/TCST.2016.2562581.
http://dx.doi.org/10.1109/CCA.2010.5611129.
Minorsky, N. (1922). Directional stability of automatically steered bodies. Journal of
Katebi, R. (2007). Modelling, simulation and control of large power plants. IFAC
the American Society for Naval Engineers, 34, 280–309. http://dx.doi.org/10.1111/
Proceedings Volumes, 40, 3–14. http://dx.doi.org/10.3182/20070709-3-RO-4910.
j.1559-3584.1922.tb04958.x.
00002.
Moore, K. L., & Xu, J.-X. (2000). Editorial: Special issue on iterative learning
Keel, L., & Bhattacharyya, S. (1997). Robust, fragile, or optimal? IEEE Transactions on
control. International Journal of Control, 73, 819–823. http://dx.doi.org/10.1080/
Automatic Control, 42, 1098–1105. http://dx.doi.org/10.1109/9.618239.
002071700405798.
Keel, L., & Bhattacharyya, S. (2016). Robustness and fragility of high order controllers:
Nam, Y.-J. (2016). Comparison study of time delay control (TDC) and uncertainty and
A tutorial. In 2016 IEEE conference on control applications (pp. 191–202). http:
disturbance estimation (UDE) based control. In 2016 16th international conference
//dx.doi.org/10.1109/CCA.2016.7587837.
on control, automation and systems (pp. 749–756). http://dx.doi.org/10.1109/ICCAS.
Kerrigan, E. C. (2015). Feedback and time are essential for the optimal control of
2016.7832415.
computing systems. IFAC-PapersOnLine, 48, 380–387. http://dx.doi.org/10.1016/j.
Nguyen, N. H., & Nguyen, P. D. (2018). Overshoot and settling time assignment with
ifacol.2015.11.309.
PID for first-order and second-order systems. IET Control Theory & Applications, 12,
Killingsworth, N. J., & Krstic, M. (2006). PID tuning using extremum seeking: Online,
2407–2416. http://dx.doi.org/10.1049/iet-cta.2018.5076.
model-free performance optimization. IEEE Control Systems Magazine, 26, 70–79.
Nicolai, M. (1922). Automatic steering device.
http://dx.doi.org/10.1109/MCS.2006.1580155. Normey-Rico, J. E. (2007). Control of dead-time processes. In Advanced textbooks in
Klaučo, M., & Kvasnica, M. (2019). MPC-based reference governors: theory and case control and signal processing, London: Springer-Verlag.
studies. In Advances in industrial control. Springer International Publishing, http: O’Dwyer, A. (2009). Handbook of PI and PID controller tuning rules. Imperial College
//dx.doi.org/10.1007/978-3-030-17405-7. Press.
Koelsch, J. (2014). Tuning tools maintain harmony in PID loops. https://www. Paz, P., Oliveira, T. R., Pino, A. V., & Fontana, A. P. (2020). Model-free neuromuscular
automationworld.com/products/software/article/13311005/. electrical stimulation by stochastic extremum seeking. IEEE Transactions on Control
Kofinas, P., & Dounis, A. I. (2019). Online tuning of a PID controller with a fuzzy Systems Technology, 28, 238–253. http://dx.doi.org/10.1109/TCST.2019.2892924.
reinforcement learning mas for flow rate control of a desalination unit. Electronics, Peretz, Y. (2018). A randomized algorithm for optimal PID controllers. Algorithms,
8(231), http://dx.doi.org/10.3390/electronics8020231. 11(81), http://dx.doi.org/10.3390/a11060081.
Koivo, H. N., & Tanttu, J. T. (1991). Tuning of PID conrollers: Survey of SISO and MIMO Petráš, I. (2012). Tuning and implementation methods for fractional-order controllers.
techniques. IFAC Proceedings Volumes, 24, 75–80. http://dx.doi.org/10.1016/S1474- Fractional Calculus & Applied Analysis, 15, 282–303. http://dx.doi.org/10.2478/
6670(17)51299-9. s13540-012-0021-4.
Koszaka, L., Rudek, R., & Pozniak-Koszalka, I. (2006). An idea of using reinforcement Johnson, M. A., & Moradi, M. H. (Eds.), (2005). PID control: New identification and design
learning in adaptive control systems. In International conference on networking, inter- methods. London: Springer-Verlag, http://dx.doi.org/10.1007/1-84628-148-2.
national conference on systems and international conference on mobile communications Pillonetto, G. (2018). System identification using kernel-based regularization: New
and learning technologies (p. 190). http://dx.doi.org/10.1109/ICNICONSMCL.2006. insights on stability and consistency issues. Automatica, 93, 321–332. http://dx.
52. doi.org/10.1016/j.automatica.2018.03.065.
Koszewnik, A., Pawłuszewicz, E., & Ostaszewski, M. (2021). Experimental studies of Pirasteh-Moghadam, M., Saryazdi, M. G., Loghman, E., E. , A. K., & Bakhtiari-Nejad, F.
the fractional PID and TID controllers for industrial process. International Journal of (2020). Development of neural fractional order PID controller with emulator. ISA
Control and Automation Systems, 19, 1847–1862. http://dx.doi.org/10.1007/s12555- Transactions, 106, 293–302. http://dx.doi.org/10.1016/j.isatra.2020.06.014.
020-0123-4. Podlubny, I. (1999). Fractional-order systems and controllers. In Fractional differential
Larsson, P.-O., & Hägglund, T. (2011). Control signal constraints and filter order equations: Vol. 198, Mathematics in science and engineering. Elsevier, http://dx.doi.
selection for PI and PID controllers. In Proceedings of the 2011 American control org/10.1016/S0076-5392(99)80028-9.
conference (pp. 4994–4999). http://dx.doi.org/10.1109/ACC.2011.5991112. Poljak, B. T., & Tsypkin, Y. (1973). Pseudogradient adaptation and training algorithms.
LeCun, Y., Bengio, Y., & Hinton, G. (2015). Deep learning. Nature, 521, 436–444. Automation and Remote Control, 34.
Lequin, O. (1997). Optimal closed-loop PID tuning in the process industry with Preitl, S., Precup, R.-E., Preitl, Z., Vaivoda, S., Kilyeni, S., & Tar, J. K. (2007). Iter-
the ‘‘iterative feedback tuning’’ scheme. In 1997 European control conference (pp. ative feedback and learning control. Servo systems applications. IFAC Proceedings
3931–3936). http://dx.doi.org/10.23919/ECC.1997.7082734. Volumes, 40, 16–27. http://dx.doi.org/10.3182/20070709-3-RO-4910.00004.

73
O.A. Somefun et al. Annual Reviews in Control 52 (2021) 65–74

Rao, V., & Bernstein, D. (2001). Naive control of the double integrator. IEEE Control Tepljakov, A., Alagoz, B. B., Yeroglu, C., Gonzalez, E. A., Hosseinnia, S. H., &
Systems Magazine, 21, 86–97. http://dx.doi.org/10.1109/37.954521. Petlenkov, E. (2021). Towards industrialization of FOPID controllers: A survey on
Recht, B. (2019). A tour of reinforcement learning: The view from continuous control. milestones of fractional-order control and pathways for future developments. IEEE
Annual Review of Control, Robotics, and Autonomous Systems, 2, 253–279. http: Access, 9, 21016–21042. http://dx.doi.org/10.1109/ACCESS.2021.3055117.
//dx.doi.org/10.1146/annurev-control-053018-023825. Tewari, A. (2002). Modern control design: With MATLAB and SIMULINK. Wiley.
Retch, B. (2020). Reflections on the learning-to-control renaissance. In Artificial intelli- Tsypkin, Y. Z. (1971). Adaptation and learning in automatic systems. In Mathematics
gence and control: Plenary talk from the 2020 IFAC virtual world congress. Germany: in science and engineering. New York: Academic Press.
IFAC. Vagia, M. (2012). PID controller design approaches-Theory, tuning and application to
Reynoso-Meza, G., Garcia-Nieto, S., Sanchis, J., & Blasco, F. X. (2013). Controller tuning frontier areas.
by means of multi-objective optimization algorithms: A global tuning framework. Van Den Hof, P. M., & Schrama, R. J. (1995). Identification and control—closed-loop
IEEE Transactions on Control Systems Technology, 21, 445–458. http://dx.doi.org/10. issues. Automatica, 31, 1751–1770.
1109/TCST.2012.2185698. Veronesi, M., & Visioli, A. (2018). Optimized retuning of PID controllers for TITO pro-
Roux-Oliveira, T., Costa, L. R., Pino, A. V., & Paz, P. (2019). Extremum seeking- cessses. IFAC-PapersOnLine, 51, 268–273. http://dx.doi.org/10.1016/j.ifacol.2018.
based adaptive PID control applied to neuromuscular electrical stimulation. 06.076.
Anais da Academia Brasileira de Ciências, 91, http://dx.doi.org/10.1590/0001- Vilanova, R., & Visioli, A. (2012). PID control in the third millennium: Lessons learned
3765201820180544. and new approaches. Springer Science & Business Media.
Roy, S., & Kar, I. N. (2020). Time-delayed control (TDC): Design issues and solutions. In Vinagre, B. M., Monje, C. A., Calderón, A. J., & Suárez, J. I. (2007). Fractional PID
S. Roy, & I. N. Kar (Eds.), Adaptive-robust control with limited knowledge on systems controllers for industry application. A brief introduction. Journal of Vibration and
dynamics: An artificial input delay approach and beyond studies in systems, decision Control, 13, 1419–1429.
and control (pp. 23–39). Singapore: Springer, http://dx.doi.org/10.1007/978-981- Visioli, A. (2006). Practical PID control. Springer Science & Business Media.
15-0640-6_2. Viteckova, M., & Vitecek, A. (2015). 2DOF controller tuning. International Journal of
Saab, S. S. (2017). Development of multivariable PID controller gains in presence of Engineering Research in Africa, 18, 57–64. http://dx.doi.org/10.4028/www.scientific.
measurement noise. International Journal of Control, 90, 2692–2710. http://dx.doi. net/JERA.18.57.
org/10.1080/00207179.2016.1263760. Wakitani, S., Yamamoto, T., & Gopaluni, B. (2019). Design and application of a
Saab, S. S., & Toukhtarian, R. (2015). A MIMO sampling-rate-dependent controller. IEEE database-driven PID controller with data-driven updating algorithm. Industrial
Transactions on Industrial Electronics, 62, 3662–3671. http://dx.doi.org/10.1109/ and Engineering Chemistry Research, http://dx.doi.org/10.1021/acs.iecr.9b00704 (p.
TIE.2014.2375252. acs.iecr.9b00704).
Samad, T. (2017). A survey on industry impact and challenges thereof [technical Wang, W. (2012). The new design strategy on PID controllers. In M. Vagia (Ed.), PID
activities]. IEEE Control Systems Magazine, 37, 17–18. http://dx.doi.org/10.1109/ controller design approaches - Theory, tuning and application to frontier areas (pp.
MCS.2016.2621438. 229–251). InTech, http://dx.doi.org/10.5772/34670.
Samad, T., & Annaswamy, A. (2013). New edition of CSS’s ‘‘the impact of control Wang, Liuping (2020). PID Control System Design and Automatic Tuning Using
technology’’ report [publication activities]. IEEE Control Systems Magazine, 33, 21. Matlab/Simulink (Wiley - IEEE). WILEY.
http://dx.doi.org/10.1109/MCS.2012.2234933. Wang, X.-s., Cheng, Y.-h., & Sun, W. (2007). A proposal of adaptive PID controller based
Savran, A., & Kahraman, G. (2014). A fuzzy model based adaptive PID controller on reinforcement learning. Journal of China University of Mining and Technology, 17,
design for nonlinear and uncertain processes. ISA Transactions, 53, 280–288. http: 40–44. http://dx.doi.org/10.1016/S1006-1266(07)60009-1.
//dx.doi.org/10.1016/j.isatra.2013.09.020. Wang, W. S. W., Davison, D. E., & Davison, E. J. (2013). Controller design for mul-
Schmidhuber, J. (2015). Deep learning in neural networks: An overview. Neural tivariable linear time-invariant unknown systems. IEEE Transactions on Automatic
Networks, 61, 85–117. http://dx.doi.org/10.1016/j.neunet.2014.09.003. Control, 58, 2292–2306. http://dx.doi.org/10.1109/TAC.2013.2258812.
Segovia, V. R., Hägglund, T., & Åström, K. J. (2014). Measurement noise filtering for Wang, H., Han, Q.-L., Liu, J., & He, D. (2020). Discrete-time filter proportional–
PID controllers. Journal of Process Control, 24, 299–313. http://dx.doi.org/10.1016/ integral– derivative controller design for linear time-invariant systems. Automatica,
j.jprocont.2014.01.017. 116, Article 108918. http://dx.doi.org/10.1016/j.automatica.2020.108918.
Sekara, T. B., & Matausek, M. R. (2009). Optimization of PID controller based
Wang, H., Luo, Y., An, W., Sun, Q., Xu, J., & Zhang, L. (2020). PID controller-based
on maximization of the proportional gain under constraints on robustness and
stochastic optimization acceleration for deep neural networks. IEEE Transactions
sensitivity to measurement noise. IEEE Transactions on Automatic Control, 54,
on Neural Networks and Learning Systems, 1–13. http://dx.doi.org/10.1109/TNNLS.
184–189. http://dx.doi.org/10.1109/TAC.2008.2008359.
2019.2963066.
Shipman, W. J., & Coetzee, L. C. (2019). Reinforcement learning and deep neural
Wang, X., Yan, X., Li, D., & Sun, L. (2018). An approach for setting parameters for two-
networks for PI controller tuning. IFAC-PapersOnLine, 52, 111–116. http://dx.doi.
degree-of-freedom PID controllers. Algorithms, 11(48), http://dx.doi.org/10.3390/
org/10.1016/j.ifacol.2019.09.173.
a11040048.
Silva, G. J., Datta, A., & Bhattacharyya, S. P. (2005). PID controllers for time-delay
Werbos, P. J. (1975). Beyond regression: New tools for prediction and analysis in the
systems. In Control engineering. Birkhäuser Basel.
behavioral sciences (Ph.D. thesis).
Skogestad, S. (2006). Tuning for smooth PID control with acceptable disturbance
Wiener, N. (1950). The human use of human beings: Cybernetics and society. Boston:
rejection. Industrial and Engineering Chemistry Research, 45, 7817–7822. http://dx.
Houghton Mifflin.
doi.org/10.1021/ie0602815.
Yang, G.-Z. (2020). Coronavirus pandemic: A call to action for the robotics community.
Slotine, J.-J. E., & Li, W. (1991). Applied nonlinear control volume 199. NJ: Prentice hall
Yu, W. (2018). Pid admittance control in task space. In W. Yu (Ed.), PID control
Englewood Cliffs.
with intelligent compensation for exoskeleton robots (pp. 139–158). Academic Press,
Smith, J. (2018b). Pros and cons of autotuning control: Part 2.
http://dx.doi.org/10.1016/B978-0-12-813380-4.00008-6, Chapter 8.
Smithm, J. (2018a). Pros and cons of autotuning control: Part 1. https://www.
Zeng, D., Zheng, Y., Luo, W., Hu, Y., Cui, Q., & Li, Q. (2019). Research on improved
controleng.com/articles/pros-and-cons-of-autotuning-control-part-1/.
auto-tuning of a PID controller based on phase angle margin. Energies, 12(1704),
Somefun, O. (2021). Design of two-degree-of-freedom (2DOF) PID tuning algorithms for
http://dx.doi.org/10.3390/en12091704.
the differential drive mobile robot using a singular specification of settling-time (Ph.D.
Zhang, J., & Guo, L. (2019). Theory and design of PID controller for nonlinear uncertain
thesis), Federal University of Technology Akure, http://dx.doi.org/10.13140/RG.2.
systems. IEEE Control Systems Letters, 3, 643–648. http://dx.doi.org/10.1109/LCSYS.
2.18698.49605/1.
Somefun, O., Akingbade, K., & Dahunsi, F. (2020). Speed control of DC motors: 2019.2915306.
Optimal closed PID-loop model predictive control. Universal Journal of Control and Zhao, C., & Guo, L. (2020). Towards a theoretical foundation of PID control for
Automation, 8, 9–21. http://dx.doi.org/10.13189/ujca.2020.080102. uncertain nonlinear systems. arXiv:2010.06864 [math], arXiv:2010.06864.
Srivastava, S., P. G. , A., Gupta, M., Prasannakumar, N., & Rudola, A. (2018). A Zhong, S., Huang, Y., & Guo, L. (2020). A parameter formula connecting PID and ADRC.
comparative study of PID and neuro-fuzzy based control schemes for a 6-DoF Science China. Information Sciences, 63, Article 192203. http://dx.doi.org/10.1007/
robotic arm. Journal of Intelligent & Fuzzy Systems, 35, 5317–5327. http://dx.doi. s11432-019-2712-7.
org/10.3233/JIFS-169814. Zhong, Q.-C., Kuperman, A., & Stobart, R. K. (2011). Design of UDE-based con-
Srivastava, S., & Pandit, V. S. (2016). A PI/PID controller for time delay systems with trollers from their two-degree-of-freedom nature. International Journal of Robust and
desired closed loop time response and guaranteed gain and phase margins. Journal Nonlinear Control.
of Process Control, 37, 70–77. http://dx.doi.org/10.1016/j.jprocont.2015.11.001. Ziegler, J. G., & Nichols, N. B. (1942). Optimum settings for automatic controllers.
Sung, S. W., Lee, J., & In-Beu. Lee, L. (2009). Process identification and PID control. Transactions of the ASME, 64.
John Wiley & Sons, http://dx.doi.org/10.1002/9780470824122.ch5. Zítek, P., Fišer, J., & Vyhlídal, T. (2013). Dimensional analysis approach to dominant
Taeib, A., & Chaari, A. (2015). Tuning optimal PID controller. International Journal of three-pole placement in delayed PID control loops. Journal of Process Control, 23,
Modelling, Identification and Control, 23, 140–147. http://dx.doi.org/10.1504/IJMIC. 1063–1074. http://dx.doi.org/10.1016/j.jprocont.2013.06.001.
2015.068872. Zweiri, Y. H., Seneviratne, L. D., & Althoefer, K. (2005). Stability analysis of a three-
Tasaka, K., Kano, M., Ogawa, M., Masuda, S., & Yamamoto, T. (2009). Direct PID term backpropagation algorithm. Neural Networks, 18, 1341–1347. http://dx.doi.
tuning from closed-loop data and its application to unstable processes. Transactions org/10.1016/j.neunet.2005.04.007.
of the Institute of Systems, Control and Information Engineers, 22, 137–144. http:
//dx.doi.org/10.5687/iscie.22.137.

74