Neural networks have emerged as a pivotal machine learning paradigm within contemporary artificial intelligence research, particularly for their adeptness at modeling complex input-output relationships. This paper addresses the training of shallow neural networks, focusing on optimizing the widely used Levenberg-Marquardt (LM) algorithm. While the LM algorithm facilitates faster convergence through adaptive learning rates and the use of Jacobian matrices, it also presents challenges in computational complexity and memory usage due to the growth of Jacobian matrices with network and dataset dimensions. To mitigate issues such as catastrophic forgetting and computational complexity and memory usage due to the growth of Jacobian matrices, this study proposes a Federated Multiple Dataset Levenberg-Marquardt (fmLM) algorithm. By employing distinct mask vectors for each dataset, the fmLM enables selective activation of network weights, allowing simultaneous training across multiple datasets while maintaining performance. Experimental results demonstrate that the fmLM algorithm achieves a promising error levels after training, effectively managing the balance between weight sharing and retention of prior learning. This approach not only enhances the efficiency of neural network training but also contributes a novel solution to catastrophic forgetting, underscoring the potential for improved generalization in neural networks.

Flight control actuators are the primary equipment of the Automatic Flight Control System (AFCS) that is used to provide short and long term stabilization. Flight control actuators are electrohydraulic actuators that are directly connected to the flight control surface. The hydraulic flow in these actuators is controlled using an Electro-Hydraulic Servo Valve (EHSV) with reference electrical command. Each EHSV has a null bias command to hold the valve in the null position. The null bias command depends on valve hysteresis, temperature, hydraulic pressure, and reference acceleration command. The null bias command and its variation reduce the tracking performance of the flight control actuators. In this article, we proposed a neural network bias compensator to compensate for the EHSV null bias command and improve the tracking performance of the controller. The nonlinear Hammerstein-Wiener model of the actuator was estimated from the test data. Then, a neural network bias compensator was designed in addition to the lead controller. The performance of the neural network bias compensator is analyzed through a series of simulations that demonstrate the desired qualities.

This paper presents a comprehensive implementation of a multilayer Neural Network (NN) controller for trajectory tracking in a 6 Degrees of Freedom (DoF) Lynx Helicopter Model, combining a passivity-based approach with real-time weight tuning and an inner-loop Linear Quadratic Regulator (LQR) for stability augmentation. The NN controller replaces traditional integral control, minimizing steady-state errors, enhancing reference tracking accuracy, and regulating off-axis states with a robust architecture that integrates filtered error signals, passivity-based terms, and system states. Using a forward pass for control generation and a backward pass for real-time weight adjustment via gradient descent, the controller ensures stability and convergence through adaptive gains and tuned learning rates. Time-domain simulations demonstrate precise tracking performance across diverse reference commands with minimal overshoot, smooth transients, and steady neural network weight convergence, validating the hybrid controller’s reliability under varying operating conditions.

This study addresses the challenges of 'catastrophic forgetting' and 'multi-task learning' encountered in the field of data classification and analysis, particularly with the use of Convolutional Neural Networks (CNNs). The aim of the study is to employ genetic algorithm (GA) to mitigate these issues. Methodologically, we have developed an optimization strategy that utilizes layer-based binary masks to tailor CNNs models for multiple dataset. GA serves as a heuristic search method to optimize a binary mask for each dataset. Experiments have been conducted on widely-used dataset such as MNIST, Fashion MNIST, and KMNIST. The obtained results are notably impressive, yielding classification accuracies of 76.25% for MNIST, 76% for Fashion MNIST, and 74.43% for KMNIST. These findings demonstrate that our proposed approach can generate high-performance models not only for a single task but also for multiple tasks.

Identification of dynamical systems from frequency response data (FRD) has extensively been studied and effective techniques have been developed. Given different FRD sets obtained from different systems and a fixed state space model structure, is it possible to find a constant parameter vector containing$(\mathbf {A},\mathbf {B},\mathbf {C},\mathbf {D})$quadruple's numerical content and a FRD-associated mask vector set that approximates the spectral information available in each FRD set? This article proposes a genetic algorithm based optimization approach to determine the real parameter vector$(\mathbf {A},\mathbf {B},\mathbf {C},\mathbf {D})$and the binary mask vector through a sequential optimization scheme. We study state space models for matching FRD from multiple systems. Results show that the proposed optimization approach solves the problem and compresses multiple dynamical models into a single masked one.

Satcom on the Move (SOTM) antennas are the primary devices for establishing satellite communication in both military and commercial applications. The main design parameters of the SOTM antennas are low cost, low weight, and high data rate. SOTM antennas are basically two or three degrees of freedom robotic manipulators with an antenna payload. In the classical approach, a position and stabilization controller is implemented in order to achieve a high data rate. Most applications use a tracking algorithm to find the maximum RF signal strength by planning a special trajectory for the end effector. In this article, SOTM antennas are modeled and controlled as if they are robotic manipulators. In addition, a neural network controller is implemented to control the robot manipulator and find the maximum RF signal. The neural network controller includes filtered computed torque control (CTM), robustifying signal, and 2 layers neural network structure. The filtered CTM and robustifying signal ensure the closed-loop characteristic, while the neural network structure eliminates nonlinearities and generates the required torque to find the maximum RF signal. The results obtained through a series of simulations demonstrate the desired qualities.

Research on signals collected from the human heart has been a core subject area as the heart displays a rich set of dynamical information that needs careful analysis for medical diagnosis and treatment. The acquisition of the electrical activity signals is a convenient way to analyze, control, evaluate and understand the heart. Electrocardiography (ECG) measurements are used to categorize the heartbeat behaviors to achieve classification. ECG heartbeat signal classification methods range from classical signal processing to convolutional neural networks. Heterogeneous Harmonization of Heartbeat Signals for Arrhythmia Detection (H3-SAD) method based CNN is proposed in this study. H3-SAD method differs from other methods in the literature with its robust and tempered classification ability against heterogeneous spectrums of ECG Signal by targeting being a part of high mobility lifestyle. Literature studies have reasonable estimation rates for MIT-BIH Dataset but not for the heterogeneous acquisition of data in real-life applications. The key point that tempers our classification algorithm is applied dynamic augmentation details towards different signal sources and input values that adduct data to real-life, and heterogeneous augmentation-based CNN architecture.

The navigation of mobile robots using heuristic algorithms is one of the important issues in computer and control sciences. Path planning and obstacle avoidance are current topics of navigational challenges for mobile robots. The major drawbacks of conventional methods are the inability to plan motion in a dynamic and unknown environment, failure in crowded and complex environments, and inability to predict the velocity vector of obstacles and non-optimality of the synthesised path. This paper presents a novel path planning approach using a grasshopper algorithm for navigation of a mobile robot in dynamic and unknown environments. To accomplish this goal, two different approaches are presented. First, a sensory system is used to detect the obstacles and then a new method is developed to predict and avoid static and dynamic obstacles while the velocity of obstacles is unknown. The robot uses the obtained information and finds a collision-free, optimal and safe path. The controller proposed in this paper is tested in crowded and complex environments. Simulation results show that the approach is successful in all test environments. Also, the proposed controller is compared with several heuristic methods. The comparison work stipulates that the introduced controller here is promising in terms of running time, optimality, stability and failure rate.

In this study, a controller is designed by using neural networks. Performance of this controller is compared with an another controller which is designed by using classical control methods. They are both tested with the same flight scenario to analyse their capabilities. This comparison provides comprehension about the dynamical capabilities of the neural networks as a controller, so that in further studies, robustness against model uncertainties and sudden unplanned airframe changes may be achieved with a neurocontroller that is subjected to online training.

In recent decades, one of the challenging problems is path planning for autonomous vehicle in dynamic environments with along static or moving obstacles. The main aim of these researches is to reduce congestion, accidents and improve safety. We propose an optimal path planning using model predictive controller (MPC) which automatically decides about the mode of maneuvers such as lane keeping and lane changing. For ensuring safety, we have additionally used two different potential field functions for road boundary and obstacles where the road potential field keeps the vehicle for going out of the road boundary and the obstacle potential field keep the vehicle away from obstacles. We have tested the proposed path planning on the different scenarios. The obtained results represent that the proposed method is effective and makes reasonable decision for different maneuvers by observing road regulations while it ensures the safety of autonomous vehicle.

Movement sequence of a real air combat flight contains valuable information that can be used to infer artificial air combat learning. There are different ways to control unmanned aerial vehicles for a given flight path. But identifying the best move at the time being relative to an enemy air craft requires learning flight experience from real air combat fighters. This paper shows how to set up learning and control environment with adaptive neuro fuzzy inference system for maneuver decisions using real F-16 flight information. Real flight information is also utilized to justify the test results.

Adaptive Neural Fractional Order Proportional Integral Derivative (FOPID) controller is designed to control a missile using Proportional Navigation Guidance PNG system. The proposed FOPID controller is intended to improve the performance of PNG system in terms of miss distance accuracy and stability. A new tuning procedure has been proposed by applying hybrid neural genetic algorithm in order to align the missile attitude with the Line of Sight (LOS) angle. Genetic algorithm is used first in which it has fast convergence speed at the initial tuning stages, but near the global optimum values the tuning process becomes very slow, therefore neural technique is used to train a neural network which has faster tuning speed near the optimal values, and the tuning accuracy becomes much better. The tuning method has been compared with Ziegler-Nichols which is applied on PID controller and the results showed the superiority of the proposed tuning method. The need for fractional order type of feedback control system is justified by the nonlinear nature of the proportional navigation system and the dynamics of the missiles, to which many alternatives were applied in the literature using less accurate controllers while the proposed control system proved to have more accuracy hitting the target with less value of miss distance and more stability in the angle of attack and the motion during flight as well as the performance of the controller in terms of second and infinity norms.

This paper presents an FPGA based synthesizable offline UAV local path planner implementation using Evolutionary Algorithms for 3D unknown environments. A Genetic Algorithm is selected as the path planning algorithm and all units of it are executed on a single FPGA board. In this study, Nexys 4 Artix-7 FPGA board is selected as the target device and Xilinx Vivado 2015.4 software is used for synthesis and analysis of HDL design. Local path planner is designed in a way that it has two flight modes: free elevation flight mode and fixed elevation flight mode. Designed FPGA based local path planner which has 74 MHz operating frequency and 62% logic slice utilization is tested into two different unknown environments generated by a LIDAR sensor. Results show that both GAs are efficient path planning algorithms for UAV applications and FPGAs are very suitable platforms for flight planning periphery.

This paper considers a novel model based switching control scheme. The philosophy of the approach is to design a conventional linear or nonlinear feedback control scheme for a nominal plant model and to force the true system states to that of the nominal model by introducing a switching term. In the demonstrated example, a twin rotor system is considered. Feedback controller for the nominal system is designed using the backstepping method and the results show that the proposed technique is successful.

The applications of Unmanned Aerial Vehicles (UAVs) require robust control schemes that can alleviate disturbances such as model mismatch, wind disturbances, measurement noise, and the effects of changing electrical variables, e.g., the loss in the battery voltage. Proportional Integral and Derivative (PID) type controller with noninteger order derivative and integration is proposed as a remedy. This paper demonstrates that a neural network can be trained to provide the coefficients of a Finite Impulse Response (FIR) type approximator, that approximates to the response of a given analog PIλDμcontroller having time varying action coefficients and differintegration orders. The results obtained show that the neural network aided FIR type controller is very successful in driving the vehicle to prescribed trajectories accurately. The response of the proposed scheme is highly similar to the response of the target PIλDμcontroller and the computational burden of the proposed scheme is very low.

This paper describes an emerging tool for industry: fractional order systems. Conventional understanding of the notion of derivative and integral uses integer orders and our sense is mature in their physical interpretations. Derivatives or integrals of fractional orders are generalizations of the concept containing the classical cases and solutions based on fractional order operators utilize the full flexibility offered by the mathematical definitions. The interest of the industry to fractional order systems lie in the fact that complicated modules can be simplified significantly and practical applications can be diverse. This paper describes linear and nonlinear cases with necessary stability and performance considerations for the benefit of a practicing engineer exploiting informatics in industry.

This paper focuses on the swing up and stabilization control of a rotary inverted pendulum system with linear quadratic regulator (LQR), sliding mode control (SMC) and fuzzy logic control (FLC). The inverted pendulum, a popular control application exists in several forms and due to its widespread use for prototyping control schemes we present experimental results obtained on a rotary version of the pendulum. The paper develops the dynamical model and introduces the implementation of the considered schemes comparatively.

Representation of knowledge within a neural model is an active field of research involved with the development of alternative structures, training algorithms, learning modes and applications. Radial Basis Function Neural Networks (RBFNNs) constitute an important part of the neural networks research as the operating principle is to discover and exploit similarities between an input vector and a feature vector. In this paper, we consider nine architectures comparatively in terms of learning performances. Levenberg-Marquardt (LM) technique is coded for every individual configuration and it is seen that the model with a linear part augmentation performs better in terms of the final least mean squared error level in almost all experiments. Furthermore, according to the results, this model hardly gets trapped to the local minima. Overall, this paper presents clear and concise figures of comparison among 9 architectures and this constitutes its major contribution.

This paper presents a novel parameter adjustment scheme to improve the robustness of fuzzy sliding-mode control achieved by the use of an adaptive neuro-fuzzy inference system (ANFIS) architecture. The proposed scheme utilizes fractional-order integration in the parameter tuning stage. The controller parameters are tuned such that the system under control is driven toward the sliding regime in the traditional sense. After a comparison with the classical integer-order counterpart, it is seen that the control system with the proposed adaptation scheme displays better tracking performance, and a very high degree of robustness and insensitivity to disturbances are observed. The claims are justified through some simulations utilizing the dynamic model of a 2-DOF direct-drive robot arm. Overall, the contribution of this paper is to demonstrate that the response of the system under control is significantly better for the fractional-order integration exploited in the parameter adaptation stage than that for the classical integer-order integration.

Describes a multi-layered hierarchical motion planning strategy for a class of self-reconfigurable modular robotic systems, I-Cubes. The approach is based on the synthesis of motion on the basis of metacubes, which have a particular structure possessing 8 Cubes and 16 Links. The developed strategy organizes the metacube motions and the corresponding cube-level motions. At the lowest level, link motions are generated. The resulting system is demonstrated to be capable of performing a pre-specified task of moving from one position/shape to another. The paper describes the latest results of our planning strategy through some experimentally justified examples.

A general backpropagation algorithm is proposed for feedforward neural network learning with time varying inputs. The Lyapunov function approach is used to rigorously analyze the convergence of weights, with the use of the algorithm, toward minima of the error function. Sufficient conditions to guarantee the convergence of weights for time varying inputs are derived. It is shown that most commonly used backpropagation learning algorithms are special cases of the developed general algorithm.

This paper discusses the stabilizability of artificial neural networks trained by utilizing the gradient information. The method proposed constructs a dynamic model of the conventional update mechanism and derives the stabilizing values of the learning rate. This is achieved by integrating the error backpropagation (EBP) technique with variable structure systems (VSS) methodology, which is well known with its robustness to environmental disturbances. In the simulations, control of a three degrees of freedom anthropoid robot is chosen for the evaluation of the performance. For this purpose, a feedforward neural network structure is utilized as the controller.

In this paper, a novel method for the establishment of a sliding mode in a nonlinear system is presented. The method discussed aims to minimize a cost measure, which is a function of the switching surface. For this purpose, an adaptive fuzzy controller is selected and the parameters of the defuzzifier are adjusted such that the cost is minimized and the system is enforced to behave in sliding mode. The paper considers a 2-DOF direct drive robotic manipulator as the test bed and a standard fuzzy system is used as the controller. The results obtained clearly stipulate that a sliding motion can be achieved by appropriately tuning the parameters of the controller.

This paper discusses the stabilizability of artificial neural networks trained by utilizing the gradient information. The method proposed constructs a dynamic model of the conventional update mechanism and derives the stabilizing values of the learning rate. The stability in this context corresponds to the convergence in adjustable parameters of the neural network structure. It is shown that the selection of the learning rate as imposed by the proposed algorithm results in stable training in the sense of Lyapunov. Furthermore, the algorithm devised filters out the high frequency dynamics of the gradient descent method. The excitation of this dynamics typically occurs in the presence of noise and abruptly changing the parameters of the mapping being learned. This adversely influences the learning performance that can be attained during a training cycle. A natural consequence following this excitation is divergence in parameter space. The method analyzed in this paper integrates the gradient descent technique with variable structure systems methodology, which is well known for its robustness to environmental disturbances. In the simulations, control of a three degrees of freedom anthropoid robot is chosen for the evaluation of the performance. For this purpose, a feedforward neural network structure is utilized as the controller. Highly nonlinear dynamics of the plant, existence of a considerable amount of observation noise, and the adverse effects of gravitational forces constitute the difficulties to be alleviated by the neurocontroller trained with the proposed method. In order to come up with a fair comparison, the results obtained with the pure gradient descent technique with the same initial conditions are also presented and discussed. © 2000 John Wiley & Sons, Inc.

Develops a training methodology for computationally intelligent systems utilizing gradient information in parameter updating. The devised scheme uses the first-order dynamic model of the training procedure and applies the variable structure systems approach to control the training dynamics. This results in an optimal selection of the learning rate, which is continually updated as prescribed by the adopted strategy. The parameter update rule is then mixed with the conventional error backpropagation method in a weighted average. The paper presents an analysis of the imposed dynamics, which is the response of the training dynamics driven solely by the inputs designed by a variable structure control approach. The analysis continues with the global stability proof of the mixed training methodology and the restrictions on the design parameters. The simulation studies presented are focused on the advantages of the proposed scheme with regards to the compensation of the adverse effects of the environmental disturbances and its capability to alleviate the inherently nonlinear behavior of the system under investigation. The performance of the scheme is compared with that of a conventional backpropagation. It is observed that the method presented is robust under noisy observations and time varying parameters due to the integration of gradient descent technique with variable structure systems methodology. In the application example studied, control of a two degrees of freedom direct-drive robotic manipulator is considered. A standard fuzzy system is chosen as the controller in which the adaptation is carried out only on the defuzzifier parameters.

In this study, a novel method for identification and control of nonlinear systems is developed. The method proposed realizes the dynamics of a system by employing the Runge-Kutta method at the upper level. The intermediate level of the strategy constructs the architecture utilizing an adaptive neuro fuzzy inference system. The overall system is able to imitate the behavior of a complex dynamic system with a few rules or to control the system with high accuracy. The proposed method has been applied to a two degrees of freedom direct drive SCARA robot.