Optimized self-adaptive PID speed control for

autonomous vehicles
Yassine Kebbati, Naïma Aït Oufroukh, Vincent Vigneron, Dalil Ichalal,
Dominique Gruyer

To cite this version:

Yassine Kebbati, Naïma Aït Oufroukh, Vincent Vigneron, Dalil Ichalal, Dominique Gruyer. Op-
timized self-adaptive PID speed control for autonomous vehicles. 26th International Conference
on Automation and Computing (ICAC 2021), Sep 2021, Portsmouth, United Kingdom. pp.1-6,
�10.23919/ICAC50006.2021.9594131�. �hal-03442081�

HAL Id: hal-03442081

https://hal.science/hal-03442081v1
Submitted on 3 Feb 2022

HAL is a multi-disciplinary open access L’archive ouverte pluridisciplinaire HAL, est

archive for the deposit and dissemination of sci- destinée au dépôt et à la diffusion de documents
entific research documents, whether they are pub- scientifiques de niveau recherche, publiés ou non,
lished or not. The documents may come from émanant des établissements d’enseignement et de
teaching and research institutions in France or recherche français ou étrangers, des laboratoires
abroad, or from public or private research centers. publics ou privés.
Proceedings of the 26th International Conference on Automation & Computing,
University of Portsmouth, Portsmouth, UK, 2-4 September 2021

Optimized self-adaptive PID speed control for autonomous vehicles

Yassine Kebbati1,* , Naima Ait-Oufroukh1 , Vincent Vigneron1,2 , Dalil Ichalal1 , and Dominique Gruyer3
1
IBISC-EA4526, Université Evry, Université Paris-Saclay, France
2
School of Applied Sciences (FCA),UNICAMP, Limeira, Brazil
3
IFSTTAR/ PICS-L, Université Gustave Eiffel, France
Email∗ : [email protected]

Abstract— The main control tasks in autonomous vehicles

are steering (lateral) and speed (longitudinal) control. PID
controllers are widely used in the industry because of their
simplicity and good performance, but they are difficult to
tune and need additional adaptation to control nonlinear
systems with varying parameters. In this paper, the longitudinal
control task is addressed by implementing adaptive PID control
using two different approaches: Genetic Algorithms (GA-PID)
and then Neural Networks (NN-PID) respectively. The vehicle
nonlinear longitudinal dynamics are modeled using Powertrain
blockset library. Finally, simulations are performed to assess
and compare the performance of the two controllers subject to Fig. 1: Forces and disturbances affecting vehicle dynamics.
external disturbances.
Index Terms— Autonomous Vehicle, Adaptive PID Control, on adaptive PID control using the inverse model theory,
Neural Networks, Genetic Algorithms, Optimization. [8] used the twiddle algorithm to tune a PID for trajectory
tracking and [9] used Genetic Algorithms (GA) to tune a
I. INTRODUCTION PID for controlling DC motors. [10] and [11] used NN-
The autonomous vehicle technology has been the main PID for frequency regulation and vehicle steering control
trend in the last decade. Although artificial intelligence and respectively. In [12], the authors used radial basis function
control researchers are quickly closing the gap to a real self- to adapt PID gains for the vehicle longitudinal control.
driving vehicle, several significant challenges still persist. A The research of [13] presented an adaptive PID with GA
fully autonomous vehicle requires fully automatic steering optimization for the lateral control task. However, most of
and speed control modules. Cruise control systems (CC) these methods are based on oversimplified models exclud-
control the speed of the vehicle and have been the subject of ing or simplifying external disturbances which significantly
extensive research: Osman et al. [1] and Simorgh et al. [2] affects the performance in real-life situations. In addition,
for instance showed that CC versions for highway driving most adaptation methods require the measurement of certain
worked only above certain speed limits as the braking part parameters that are not easily accessible. To address this
was not addressed [3]. Adaptive CC systems were introduced issue, two PID adaptation methods were applied to control
later on to tackle both acceleration and braking [4], these the vehicle velocity; the first method is based on a GA
systems are divided into high level and low level controllers. with a complete vehicle model taking into account external
The first computes the required acceleration or deceleration, disturbances, and the second one uses a feed forward neural
and the second actuates the throttle or the brake. These network (NN) that utilises only the error and control signals
systems allowed to automate vehicles even for urban driving. to adapt the PID gains. Simulations were carried to assess the
Many speed controllers are available in the literature. [5] performance of the designed controllers. Section II of this
worked on the sliding mode control for vehicle platooning paper deals with the modeling of the vehicle longitudinal
in uncertain topologies. [6] designed a preview controller dynamics. The design, optimization and adaptation of the
that integrates road slope, speed profile and vehicle dynamics controller are presented in section III. The performance of
to form an augmented optimal control. [7] used the model the controller is evaluated and discussed in section IV, and
predictive control (MPC) for low velocity tracking. Actually, section V concludes the paper and gives some perspectives
most control systems are based on the PID controller which for future work.
is easy to design and yet performs very well. However, the
II. V EHICLE M ODELING
non-linearities of the vehicle subsystems and the variety of
disturbances that face the vehicle such as road elevation, The longitudinal control of an autonomous vehicle regu-
wind speed and road conditions require the PID to be lates its speed to follow a speed profile. The latter is obtained
adaptive to these changing parameters. [2] and [3] worked from a speed planner or profiler based on changes of the
external environment like the shape of the road, traffic signs C. Wheel Dynamics
and weather conditions. The vehicle is mainly propelled by The wheel translates the gear torque into a traction force
the engine drive torque and stopped by the braking torque. that propels the vehicle. Thus, the angular speed of the wheel
determines the velocity of the vehicle as the following:
A. Vehicle Body Dynamics
Ft/b = R1w Tg/b

The longitudinal dynamics of the vehicle are obtained by (4)
ωw = R1w V
applying Newton’s second law as shown in Fig. 1, this results
in a simplified point mass model [14]: where Ft/b and Tg/b denote either the tractive force (Fig. 1)
 and the drive torque or the braking force and braking torque
F = M dv dt + Fa + Fg + Fr respectively. Rw and V are the respective wheel radius and
 t


Fg = M g sin θ vehicle longitudinal velocity.
(1)
 Fa = 12 ρCd Av 2


Fr = M gCr cos θ D. Tire Dynamics
The tires of a vehicle undergo different kinds of defor-
where the parameters are defined as: mations and behave in a highly nonlinear manner, the most
• M the total mass of the vehicle widely used longitudinal tire model is the magic formula or
• Ft the traction force (drive force) Pacejka model [16]:
• Fa the aerodynamic drag (air friction)
Fx = Fz D sin(C tan−1 [Bk − E[Bk − tan−1 (Bk)]]) (5)
• Fg the downgrade force (gravity pull)
• Fr the rolling resistance force with Fx and Fz being the tire longitudinal force and vertical
• ρ the air density load, k representing the wheel slip and B, C, D and E being
• Cr the rolling resistance coefficient empirical tire coefficients.
• Cd the drag coefficient
E. Brake Dynamics
• A the vehicle cross sectional area
dv
• dt the vehicle acceleration/deceleration.
The brakes of the vehicle can be either a drum or a disk
system. The disk braking system has been used in this paper
B. Engine Dynamics for both front and rear wheels. The braking torque delivered
by the mechanism is given as:
The engine and transmission dynamics make up the vehi-
cle’s propulsion system and are highly non-linear. For electric µP πBa2 Rm Npads
Tb = (6)
vehicles, the engine is an electric motor that supplies power 4
to the wheels through the transmission. The latter is linked where P, Rm and Ba are the applied brake pressure, the
to the electric drive through a gearbox which can be modeled mean radius of the brake pad and the brake actuator diameter
under the assumption of no losses by (2): respectively. µ and Npads are the disc-pad friction coefficient
 and the number of pads respectively.
Tg = kg Te F. Battery modeling
(2)
ωg = kg ωe
The electric motor is powered by the battery voltage ub
where ωg and ωe are the angular speeds of the gear and which can be modeled by a resistor Rb in series with an open
the engine, Tg and Te are the gear and electric drive circuit voltage Voc . Both of them depend on the battery state
torques and kg represents the ratio of the gearbox. The of charge (SOC):
electric torque is generated by the electric machine and its (
ub = Ns (Voc (SOC) + NIbp Rb (SOC))
model can be static or dynamic [15]. The dynamic model Rt (7)
depends on the type of machine whether it is a DC or SOC = C1b 0 NIbp dt
AC (synchronous\asynchronous) motor. It links the input where Ns , Np are the number of cells in series and in parallel
current\voltage to the output electric torque. The static model respectively, and the denominator Cb is the battery capacity.
which is used in this work and given in (3), is a data-driven The Powertrain blockset of MATLAB has been used to
model that is based on the electric drive efficiency map: build a high fidelity vehicle longitudinal model (see Fig. 2).
 The vehicle, motor and battery characteristic parameters were
 Te = Te-ref based on the Renault Zoe Vehicle (see table I). The rest of the
Te ωg (3) parameters are blockest default values. Moreover, both the
 Ib =
ub η k brake and accelerator pedal actuators are modeled as a first
with k = 1 if Te ωg ≥ 0, k = −1 otherwise. Te-ref repre- order system with time constants τa = 0.75s and τb = 1s
sents the reference electric torque obtained from the motor respectively [4]. The controller generates either a braking or
efficiency map, ub denotes the battery voltage, Ib and η are an acceleration command at a time. Based on the sign of
the battery current and motor efficiency respectively. Positive the control signal, the switching logic sends the command
currents discharge the battery while negative currents, as in to the braking system or to the engine such that a negative
regenerative braking, charge the battery. control signal means that the braking is activated, otherwise,
acceleration is activated [12].
 Fig. 2: Model Block Diagram

TABLE I: Simulation parameters.

M 1468kg Cd 0.29
B. GA-based PID Optimization
Cr 0.007 kg 3.4 Genetic algorithms are inspired from biological evolution
A 2.22m2 Rw 0.329m theory which is based on the process of natural selection
ρ 1.225kg/m3 Ns 96
µ 0.9 g 9.81m/s2 and genetics [9]. The algorithm uses population sets of
CG height 0.35m Npads 2 multiple potential encoded solutions called chromosomes.
Front axle to CG 1.455m η 95% Each chromosome is rated by a fitness function (10) that
Rear axle to CG 1.132m Cb 132Ah
Rm 0.1778m Nb 2 determines how good of a solution it is.
1
III. C ONTROLLER D ESIGN Fitnessvalue = (10)
Performanceindex
A. PID Controller GA algorithms use crossover and mutation operations (Fig.
The PID controller is heavily used in the industry due to 3) to enhance the existing solutions.
its simple design and satisfactory performance. The input to The crossover operation merges existing chromosomes to-
the controller is the error signal defined as the difference gether, while the mutation operation modifies the encoding of
between the set-point and the system output as in Fig. 6. existing chromosomes to yield new and better chromosomes.
The output of the controller depends on the PID gains and The fitness score of each chromosome is assigned by an
is governed by the following equation: objective function and only fit chromosomes are selected to
Z
de(t) evolve in the next operation. In the context of PID tuning,
u(t) = Kp e(t) + Ki e(τ )dτ + Kd (8) the objective function minimizes the error that is fed to
dt
the controller. The mean of squared error (MSE) has been
where Kp , Ki and Kd are the proportional, integral and used as the objective function for a consistent performance
derivative gains and e(t) is the error between set-point R comparison between GA-PID and NN-PID.
and output Y . The discrete form of the PID has been used The GA optimization tool in (Fig. 4) has been used.
in this work where the control law is obtained using the The process initiates a random population of a certain size
trapezoid method as follows: containing values (chromosomes) for the PID gains (Kp ,
Ki and Kd ) with which the vehicle longitudinal model is

 ∆u(k) = αe(t) + βe(k − 1) + γe(k − 2) simulated. The fitness value of each chromosome is assessed

 α = Kp + K Ki Ts through the MSE. Fit chromosomes with high fitness scores
Ts +
d
2
Ki Ts Kd (9) are selected as fit parents, while unfit chromosomes with low
 β = 2 − 2 Ts − Kp
 fitness scores are discarded. The GA then performs mutation
=K

γ d
Ts and crossover on the best fit parents and produces a new
with Ts being the sampling time. population. The number of best fit parents to be enhanced
is predefined as a percentage called elite count value. This
The performance of a PID controller is measured in whole process is repeated over many generations. Moreover,
terms of overshoot Os , rising time Rt and settling time the first selection of population individuals is random but
St . It depends on the choice of gains which is not straight bounded between a lower and an upper limit. The PID gains
forward, there are several tuning techniques such as root are optimized with regard to the varying parameters of the
locus, gain-phase margin and the widely used Ziegler Nichols model and new optimal PID gains are generated for each case
[17]. However, these methods are generally limited to linear
systems and have poor disturbance rejection. To account for
the effects of road angle and wind speed on the controller
performance, the designed model is simulated with GA to
optimize the gains of the PID controller. The choice of
this meta-heuristic optimization method is based on its good
performance reported in the literature [9]. Fig. 3: Genetic operations
 Fig. 6: Block diagram of the system

hj = f (Sj )
(12)
oj = g(Zj )
The gradient descent algorithm minimizes the MSE function
E(k) = 12 e(k)2 , which is back-propagated through the
Fig. 4: Genetic algorithm process. network using the chain rule [13]. The gradient of the error
term is given by:
of the varying parameters. A variety of cases were tested
δE(k) δE(k) δe(k) δy(k) δu(k) δoj
and the results were stored in a small data-set. The data- = (13)
h
δwij (k) δe(k) δy(k) δu(k) δoj δwijh
set is then used in a lookup table to select the optimized
gains that correspond to the varying parameters. This method where
allows continuous adaptation of the PID gains as the varying  δE(k) δe(k)
parameters change along a speed profile or drive cycle. 
 δe(k) δy(k) = e(k)
 δy(k)
≈ sign( YU (k)−Y (k−1)

(k)−U (k−1) )

δu(k)

C. Neural Network PID Adaptation 

δu(k)
= e(k) − e(k − 1)


The use of a NN inside the control loop makes it possible δo1
δu(k) (14)
to adapt the PID gains online. This method presents a better 
 δo2 = T2s (e(k) + e(k − 1))
δu(k)

= T1s (e(k) − 2e(k − 1) + e(k − 2))

and more suitable alternative to offline optimization with GA. 

 δo3
 δoj
The NN is trained online using the back-propagation of the 
 h = hi
δwij
error generated from previous PID gains. The architecture
of the NN consists of an input layer with four inputs, a δy(k)
The term δu(k) is not accessible in the network, there-
hidden layer with four neurons and an output layer with three
fore, it is approximated by sign( YU (k)−Y (k−1)
(k)−U (k−1) ), which shows
neurons (Fig. 5). The sigmoid activation function has been
whether the control law is improving. To avoid singularities,
used for the hidden layer, while contrary to existing works,
the expression sign((Y (k) − Y (k − 1)) × (U (k) − U (k − 1)) is
instead of the identity, the ReLU function has been used
used instead. For the hidden layer, the error term is defined
for the output layer to avoid predicting negative gains. The
as:
overall system is given in Fig. 6. δE(k) δE(k) δhj (k) δSj (k)
The inputs to the hidden layer (Sj ) and output layer (Zj ) i (k)
= i (k)
(15)
δwij δhj (k) δSj (k) δwij
i
are given by (11). wij are the weights between input and
h
hidden layer and wij are the weights between hidden and where: i = 1, .., ni + 1, and j = 1, .., nh .
output layer, nh and no represent the number of hidden units
 δE(k) δy(k) Pn0 δu(k) h
= −e(k) δu(k) i=1 δol (k) wjl (k)
 δhj (k)

and output units respectively. xi are the inputs to the network,

δhj (k)
the outputs of the hidden layer hj and the outputs of the δSj (k) = hj (k)(1 − hj (k)) (16)
 δSij (k) = xi (k)

network oj are given in (12) where f and g are the sigmoid

δwij (k)
and ReLU activation functions respectively.
( Pn+1 i The weights are updated according to the gradient descent
Sj = wij xi , for j = 1, .., nh rule as follows:
Pi=1nh +1 h (11)
Zj = i=1 wij hi , for j = 1, .., n0

h h δE(k)
 wij (k + 1) = wij (k) − η δw h
ij (k)
i i δE(k) (17)
 wij (k + 1) = wij (k) − η δwi (k)
ij

with η being the learning rate.

IV. S IMULATION R ESULTS AND D ISCUSSION
The designed PID controllers have been tested with the
vehicle longitudinal model given in section II. The GA
algorithm is used to optimize the gains of the PID (GA-PID)
for three scenarios:
• Sc1 A varying speed profile without disturbances (Vw =
Fig. 5: Neural network architecture. 0 and θ = 0).
 • Sc2 A varying speed profile subject to varying wind
speed and road slope (see Fig. 7).
• Sc3 Optimization for different values of Vref ,Vw and θ
separately.

Fig. 8: Speed tracking for Sc1 and Sc2

Fig. 7: Wind speed and road slope profiles

Both scenarios 1 and 2 produce a triplet of optimal gains

(Kp , Ki and Kd ) for the specific speed and disturbance
profiles. Scenario 2 shows the effect of the disturbances (vw
and θ) on the controller performance. Scenario 3 is adaptive
and more generalized and produces a data-set of optimized Fig. 9: Error variation for Sc1 and Sc2
gains corresponding to the changing values of Vref , Vw , and
θ. The GA-PID is compared to the NN-PID with optimized
initial weights and learning rate. Initialization of the weights
and choice of the learning rate are very important stages for
the performance of the NN-PID and its convergence speed.
The learning rate is selected by trial and error after the first
few simulations.
The GA optimization for scenarios 1 and 2 results in
the triplets {Kp1 = 999.75, Ki1 = 0.1, Kd1 = 0.3} and
{Kp2 = 383.79, Ki2 = 0.11, Kd2 = 128.84} respectively. Fig. 10: Acceleration/Brake variation for Sc1 and Sc2
The corresponding results are compared in Fig. 8-10 where
the speed tracking and error variation signals are in (m/s)
and the acceleration and braking signals are in percentage
(%). For scenario 3, the GA-PID gains are optimized for
Vref ∈ [0, 30] m/s, θ ∈ [−10◦ , 10◦ ] and Vw ∈ [−10, 15] m/s.
The performance of PID adaptation using the generated data-
set in a lookup table is compared to that of the NN-PID in
Fig. 11-13.
In scenario 1, the GA-PID produces a smoother and
faster response with no overshoots due to the absence of
disturbances. On the other hand, the inclusion of road slope
and wind speed profiles in scenario 2 increases the response
time which corresponds to real-life situations, this results Fig. 11: Speed tracking for Sc3 vs NN-PID.
in slower tracking with slight overshoots resulting in less
smooth responses.
The GA-PID in scenario 3 is smoother than the NN-PID
which is more aggressive since the adaptation takes place at
each iteration. On the other hand, the NN-PID is more robust
to disturbances and faster because it reacts immediately
whenever the error magnitude changes and adapts the PID
gains to quickly reject the disturbances. Overall, both GA-
PID and NN-PID insure adaptive control with very good
performance and disturbance rejection.
Fig. 12: Error variation for Sc3 vs NN-PID.
 faster than the NN-PID, while the latter was more adaptive
and more robust. Future work should evaluate other optimiza-
tion algorithms and address the smooth transition between
different PID gains as well as explore other controllers.
R EFERENCES
[1] K. Osman, M. F. Rahmat, and M. A. Ahmad, “Modelling and
controller design for a cruise control system,” Proc. 2009 5th Int.
Colloq. Signal Process. Its Appl. CSPA 2009, no. 1, pp. 254–258,
2009
[2] A. Simorgh, A. Marashian, and A. Razminia, “Adaptive PID Control
Fig. 13: Acceleration/Brake variation for Sc3 vs NN-PID. Design for Longitudinal Velocity Control of Autonomous Vehicles,”
Proc. - 2019 6th Int. Conf. Control. Instrum. Autom. ICCIA 2019, pp.
Table II compares the controller performance regarding the 1–6, 2019
M SE error and evaluates Rt , St and Os for a step response [3] J. E. A. Dias, G. A. S. Pereira, and R. M. Palhares, “Longitudinal
Model Identification and Velocity Control of an Autonomous Car,”
using different values for the hidden layer neurons (h) and IEEE Trans. Intell. Transp. Syst., vol. 16, no. 2, pp. 776–786, 2015
learning rate (lr). The GA-PID in scenario 2 has the best [4] H. Kim, D. Kim, I. Shu, and K. Yi, “Time-varying parameter adaptive
performance, while the NN-PID with (h=10, lr=0.01) is the vehicle speed control,” IEEE Trans. Veh. Technol., vol. 65, no. 2, pp.
581–588, 2016
second best controller. [5] F. Gao, X. Hu, S. E. Li, K. Li, and Q. Sun, “Distributed Adaptive
Sliding Mode Control of Vehicular Platoon with Uncertain Interaction
TABLE II: Controller performance for different h and lr. Topology,” IEEE Trans. Ind. Electron., vol. 65, no. 8, pp. 6352–6361,
Method MSE Rt (s) S t (s) O s (%) 2018
NN-PID (lr=0.01, h = 4) 7.596 0.348 0.445 4.6 [6] S. Xu, H. Peng, Z. Song, K. Chen, and Y. Tang, “Accurate and Smooth
NN-PID (lr=0.1 , h = 4) 7.605 0.37 0.445 4.59 Speed Control for an Autonomous Vehicle,” IEEE Intell. Veh. Symp.
Proc., vol. 2018-June, no. Iv, pp. 1976–1982, 2018
NN-PID (lr=0.8 , h = 4) 7.604 0.37 0.445 4.59
[7] F. Walz and S. Hohmann, “Model predictive longitudinal motion
NN-PID (lr=0.01 , h = 10) 7.555 0.348 0.445 4.6 control for low velocities on known road profiles,” Veh. Syst. Dyn.,
NN-PID (lr=0.8 , h = 15) 7.597 0.348 0.445 4.59 vol. 58, no. 8, pp. 1310–1328, 2020
[8] W. Farag and Z. Saleh, “Tuning of PID track followers for autonomous
NN-PID (lr=0.8 , h = 20) 7.557 0.348 0.445 4.6
driving,” 2018 Int. Conf. Innov. Intell. Informatics, Comput. Technol.
GA-PID(Sc1) 7.563 0.87 1.164 6.68 3ICT 2018, pp. 1–7, 2018
GA-PID(Sc2) 7.502 0.348 0.444 4.79 [9] A. A. M. Zahir, S. S. N. Alhady, W. A. F. W. Othman, and M. F.
Ahmad, “Genetic algorithm optimization of pid controller for brushed
GA-PID(Sc3) 7.534 0.449 0.616 1.52
DC motor,” Lect. Notes Mech. Eng., vol. 0, no. 9789811087875, pp.
427–437, 2018
[10] M. K. Debnath, R. Agrawal, S. R. Tripathy, and S. Choudhury,
V. CONCLUSIONS “Artificial neural network tuned PID controller for LFC investigation
including distributed generation,” Int. J. Numer. Model. Electron.
This paper dealt with the longitudinal control of au- Networks, Devices Fields, vol. 33, no. 5, pp. 1–17, 2020
tonomous vehicles where the vehicle dynamics were mod- [11] G. Han, W. Fu, W. Wang, and Z. Wu, “The lateral tracking control for
eled using the Powertrain blockset. The adaptive PID con- the intelligent vehicle based on adaptive PID neural network,” Sensors
(Switzerland), vol. 17, no. 6, pp. 1–15, 2017
troller was designed using two different approaches. The [12] L. Nie, J. Guan, C. Lu, H. Zheng, and Z. Yin, “Longitudinal speed
first approach (GA-PID) uses GA where the MSE has been control of autonomous vehicle based on a self-adaptive PID of radial
used as a cost function. The advantages of this approach are basis function neural network,” IET Intell. Transp. Syst., vol. 12, no.
6, pp. 485–494, 2018
(a) optimization with a complete vehicle model unlike other [13] B. Zhao, H. Wang, Q. Li, J. Li, and Y. Zhao. PID Trajectory Tracking
approaches which use simplified models in the form of trans- Control of Autonomous Ground Vehicle Based on Genetic Algorithm.
fer function or state space representation. (b) optimization Proceedings of the 31st Chinese Control and Decision Conference,
CCDC 2019, 3677–3682.
for a variety of cases and adaptation via a look-up table. The [14] Z. Sun and G. G. Zhu, Design and control of automotive propulsion
downside of this approach is the heavy calculations required. systems, 1st Editio. Boca Raton: CRC Press, 2014.
The second approach (NN-PID) was based on online learning [15] A. Desreveaux, M. Ruba, A. Bouscayrol, G. M. Sirbu and C. Martis,
”Comparisons of Models of Electric Drives for Electric Vehicles,”
with an MLP NN where the PID gains are adapted with back- 2019 IEEE Vehicle Power and Propulsion Conference (VPPC), Hanoi,
propagation. This approach has the following advantages: Vietnam, 2019, pp. 1-5
(a) automatic online PID adaptation and learning capability. [16] Pacejka, H. B. (2008). Vehicle System Dynamics : International
Journal of Vehicle Mechanics and Mobility. International Journal of
(b) adaptation only requires error and control signals. (c) fast Vehicle Mechanics and Mobility, (August 2012), 37–41.
instant reaction to error changes at every iteration. [17] A. Q. Ansari, Ibraheem, and S. Katiyar, “Application of ant colony
These characteristics make the NN-PID more robust and algorithm for calculation and analysis of performance indices for
adaptive control system,” Proc. Int. Conf. Innov. Appl. Comput. Intell.
better at rejecting disturbances. On the other hand, the GA- Power, Energy Control. with Their Impact Humanit. CIPECH 2014,
PID produces optimized PID gains in terms of response no. November, pp. 466–471, 2014
speed, precision and overshoot but does not generalize
well for different cases of disturbances and set-points. The
NN-PID has been found to be more efficient in terms of
adaptability and generalizes better. However, its performance
significantly depends on weights initialization and choice
of learning rate. In general, the GA-PID was smoother and