dynamic Neural Network

description159 papers

group0 followers

lightbulbAbout this topic

Dynamic Neural Networks are a class of artificial neural networks that adapt their structure and parameters in response to changing input data or tasks. They enable real-time learning and flexibility, allowing the network to modify its architecture or weights dynamically to improve performance and efficiency in various applications.

lightbulbAbout this topic

Key research themes

1. How can dynamic computation graphs be efficiently implemented and optimized in neural network models?

This theme investigates the challenges and solutions for implementing neural networks whose computation graphs vary dynamically with each input, focusing particularly on techniques for efficient batching, modular design, and scalability. Such dynamic computation graphs (DCGs) are crucial for tasks involving structured inputs like parse trees, graphs, and sequences of varying length and structure, but present difficulties for standard static graph-based deep learning frameworks.

Deep Learning with Dynamic Computation Graphs

by Peter Norvig

2023, ArXiv

Key finding: Introduces dynamic batching, a novel technique that enables efficient batching of operations across different inputs with diverse graph structures, allowing networks with dynamic computation graphs to be trained and inferred... Read more

articleView Paper downloadDownload

DRAGNN: A Transition-based Framework for Dynamically Connected Neural Networks

by Ivan Bogatyy

2022

Key finding: Presents the Transition Based Recurrent Unit (TBRU) and associated DRAGNN framework to construct recurrent neural networks with dynamically built connections guided by intermediate activations, enabling explicit structural... Read more

articleView Paper downloadDownload

Modular Dynamic Neural Network: A Continual Learning Architecture

by Pedro Cardoso

2024, Applied Sciences

Key finding: Proposes a modular dynamic neural network architecture where independent sub-networks can be added incrementally to learn new classes without retraining the entire network, effectively handling catastrophic forgetting in... Read more

articleView Paper downloadDownload

keyboard_arrow_downShow more

2. What optimization strategies enable dynamic architecture neural networks to adapt and perform multi-task learning effectively?

This area explores methods to adapt neural network architectures dynamically during training to optimize performance on multi-output and multi-task problems, focusing on iterative layer addition and heuristic optimization. These approaches address the shortcomings of static architectures by facilitating model scalability, interpretability, and adaptability to heterogeneous outputs.

Heuristic optimisation of multi-task dynamic architecture neural network (DAN2)

by Sushen Zhang

2023, Neural Computing and Applications

Key finding: Develops an iterative heuristic method for optimizing a dynamic architecture neural network (DAN2) tailored for multi-task learning, enabling layers to be added sequentially starting from a minimal structure. The method... Read more

articleView Paper downloadDownload

Multilayer neural networks – as determined systems

by Ivan Kuno

2023, Computational Problems of Electrical Engineering

Key finding: Analyzes multilayer neural networks as deterministic dynamical systems, revealing bifurcation and chaotic dynamics influenced by the learning rate. Identifies an optimal learning speed that minimizes error by balancing... Read more

articleView Paper downloadDownload

Dynamic versus static neural network model for rainfall forecasting at Klang River Basin, Malaysia

by Mohd Taha

2023, Hydrology and Earth System Sciences

Key finding: Compares static (MLP, RBFNN) and dynamic (Input Delay Neural Network) models in rainfall forecasting, demonstrating that dynamic models better capture temporal dependencies and provide superior forecasting accuracy across... Read more

articleView Paper downloadDownload

keyboard_arrow_downShow more

3. How can external memory and recurrent architectures enhance learning of explicit and implicit knowledge in dynamic neural networks?

This line of research investigates the use of neural augmentations like external differentiable memory modules (e.g., Differentiable Neural Computers) and recurrent feedback mechanisms to improve the learning and retention of complex temporal dependencies, facilitating integration of explicit rule-based knowledge with implicit pattern recognition in sequence modeling.

Learning Explicit and Implicit Knowledge with Differentiable Neural Computer

by Mohamad Ivan Fanany and

2017

Key finding: Demonstrates that augmenting a Multi-Layer Perceptron controller with an external differentiable memory in a DNC architecture significantly improves the ability to learn both implicit representations (handwritten digit... Read more

articleView Paper downloadDownload

Mechanisms of human dynamic object recognition revealed by sequential deep neural networks

by H.Steven Scholte

2025, PLOS Computational Biology

Key finding: Finds that models integrating sequential lateral recurrence and adaptation mechanisms best predict human rapid serial visual presentation object recognition performance and neural response dynamics, suggesting that recurrent... Read more

articleView Paper downloadDownload

Active Learning in Recurrent Neural Networks Facilitated by a Hebb-like Learning Rule with Memory

by Frank Emmert-Streib

2017

Key finding: Shows that introducing memory into Hebb-like learning rules enables recurrent neural networks to actively select inputs during training, significantly reducing learning time for timing tasks compared to random learning,... Read more

articleView Paper downloadDownload

keyboard_arrow_downShow more

All papers in dynamic Neural Network

Evolving Bipedal Locomotion with a Dynamically-Rearranging Neural Network

by Peter Eggenberger Hotz

2025, Lecture Notes in Computer Science

Since highly complicated interaction dynamics exist, it is in general extremely dicult to design controllers for legged robots. So far various methods have been proposed with the concept of neural circuits, so{called Central Pattern... more

descriptionView Paper arrow_downwardDownload

An embodied account of serial order: How instabilities drive sequence generation

by Gregor Schoner

2025, Neural Networks

Learning and generating serially ordered sequences of actions is a core component of cognition both in organisms and in artificial cognitive systems. When these systems are embodied and situated in partially unknown environments, specific... more

descriptionView Paper arrow_downwardDownload

In silico methods in stability testing of hydrocortisone, powder for injections: Multiple regression analysis versus dynamic neural network

by Predrag Stankovic

2025, Chemical Industry

This article presents the possibility of using of multiple regression analysis (MRA) and dynamic neural network (DNN) for prediction of stability of Hydrocortisone 100 mg (in a form of hydrocortisone sodium succinate) freeze-dried powder... more

descriptionView Paper arrow_downwardDownload

Identification and control of continuous-time nonlinear systems via dynamic neural networks

by WAI LUN LO

2025, IEEE Transactions on Industrial Electronics

descriptionView Paper arrow_downwardDownload

Neural network models of eddy current multi-frequency system for nondestructive testing

by Tomasz Chady

2025, IEEE Transactions on Magnetics

descriptionView Paper arrow_downwardDownload

Character Rotation Absorption Using a Dynamic Neural Network Topology: Comparison With Invariant Features

by christophe choisy

2025, Proceedings of the 4th International Workshop on Pattern Recognition in Information Systems

This paper treats on rotation absorption in neural networks for multioriented character recognition. Classical approaches are based on several rotation invariant features. Here, we propose to use a dynamic neural network topology to... more

descriptionView Paper arrow_downwardDownload

Passivity Analysis of Dynamic Neural Networks with Different Time-scales

by ALEJANDRO CRUZ SANDOVAL

2025, Neural Processing Letters

Dynamic neural networks with different time-scales include the aspects of fast and slow phenomenons. Some applications require that the equilibrium points of the designed networks are stable. In this paper, the passivity-based approach is... more

descriptionView Paper arrow_downwardDownload

Using neural networks to control the process of plasma etching and deposition

by Gail Erten

2025, Proceedings of International Conference on Neural Networks (ICNN'96)

Neural architectures are proposed to model and control plasma etching and deposition processes in semiconductor wafer manufacturing. Static and dynamic neural networks are used to develop plant models and inverse models. A single-hidden... more

descriptionView Paper arrow_downwardDownload

Passivity Analysis for Neuro Identifier with Different Time-Scales

by ALEJANDRO CRUZ SANDOVAL

2024, Lecture Notes in Computer Science

descriptionView Paper arrow_downwardDownload

On-line training of recurrent neural networks with continuous topology adaptation

by Dragan Obradovic

2024, IEEE Transactions on Neural Networks

descriptionView Paper arrow_downwardDownload

THE ROLE OF SENSORY-MOTOR COORDINATION - Identifying Environmental Motion Dynamics with Dynamic Neural Networks

by Arul Selvan

2024, Proceedings of the Fifth International Conference on Informatics in Control, Automation and Robotics Service

We describe three recurrent neural architectures inspired by the proprioceptive system found in mammals; Exo-sensing, Ego-sensing, and Composite. Through the use of Particle Swarm Optimisation the robot controllers are adapted to perform... more

descriptionView Paper arrow_downwardDownload

Dynamic neural network control for non-linear systems: optimal neural network structure and stability analysis

by Thomas Stanford

2024, Chemical Engineering Journal

Design techniques for non-linear dynamic systems are closely related to their stability properties. Stability results can be used to design a reliable controller. This paper discusses the stability analysis of the dynamic neural network... more

descriptionView Paper arrow_downwardDownload

Dynamical neural network organization of the visual pursuit system

by Curt Deno

2024, IEEE Transactions on Biomedical Engineering

The central nervous system is a parallel dynamical system which connects sensory input with motor output for the performance of visual tracking. This paper applies elementary control system tools to extend dynamical neural network models... more

descriptionView Paper arrow_downwardDownload

Hydrologic impact of climate change in the Saguenay watershed: comparison of downscaling methods and hydrologic models

by Paulin Coulibaly

2024, Journal of Hydrology

Changes in global climate will have significant impact on local and regional hydrological regimes, which will in turn affect ecological, social and economical systems. However, climate-change impact studies on hydrologic regime have been relatively rare until recently, mainly because Global Circulation Models, which are widely used to simulate future climate scenarios, do not provide hourly or daily rainfall reliable enough for hydrological modeling. Nevertheless, more reliable rainfall series corresponding to future climate scenarios can be derived from GCM outputs using the so called 'downscaling techniques'. This study applies two types of statistical (a stochastic and a regression based) downscaling techniques to generate the possible future values of local meteorological variables such as precipitation and temperature in the Chute-du-Diable sub-basin of the Saguenay watershed in northern Québec, Canada. The downscaled data is used as input to two different hydrologic models to simulate the corresponding future flow regime in the catchment. In addition to assessing the relative potential of the downscaling methods, the paper also provides comparative study results of the possible impact of climate change on river flow and total reservoir inflow in the Chute-du-Diable basin. Although the two downscaling techniques do not provide identical results, the time series generated by both methods indicates a general increasing trend in the mean daily temperature values. While the regression based downscaling technique resulted in an increasing trend in the mean and variability of daily precipitation values, such a trend is not obvious in the case of precipitation time series downscaled with the stochastic weather generator. Moreover, the hydrologic impact analysis made with the downscaled precipitation and temperature time series as input to the two hydrological models suggest an overall increasing trend in mean annual river flow and reservoir inflow as well as earlier spring peak flows in the basin.

descriptionView Paper arrow_downwardDownload

Dynamic neural network-based Pulsed Plasma Thruster (PPT) fault detection and isolation for the attitude control system of a satellite

by arturo valdes

2024, 2008 IEEE International Joint Conference on Neural Networks (IEEE World Congress on Computational Intelligence)

NOTICE: The author has granted a nonexclusive license allowing Library and Archives Canada to reproduce, publish, archive, preserve, conserve, communicate to the public by telecommunication or on the Internet, loan, distribute and sell... more

Figure 6.2 shows the “High Level” FDI scheme that was proposed in Chapter 5 which is composed by three DNNs. As described in Chapter 5, Section 5.3, a set of three relative attitude variables and the sequence of pulses that are generated by the corresponding pair of thrusters

To avoid the occurrence of these failures, an FDI system must detect and isolate the faults to

in Figure 2.3, a qualitative model-based method called Fault-Tree technique is used as the engine unmanned space vehicle has been investigated by the approach that is proposed in [28]. As shown

A simulated ACS-model was used to generate the database for the automatic construction of as attributes with specific numerical value. Finally, the correct combination of this attributes the above fault-tree system. The attributes selected are: reaction wheel motor current, reaction analysis purposes. From the database example, preliminary information is extracted and labeled

The proposed FDI system in [27] includes two different structures: a dynamic neural perform the tasks of detection and isolation. As shown in Figure 2.5, neural networks are network for detection and a static neural network for isolation. The composition of a multilaye

Using a simulated ACS-model, healthy and faulty data set from a reaction wheel actuator

the z-axis (right-handed rule). Figure 3.1 shows the axes of the Earth Centered Inertial (ECI) Earth, parallel to the rotation axis of the Earth, and the y-axis is perpendicular to the x-axis and center of the Earth, but rotates with it. The x-axis points to the Greenwich meridian in the

satellite. This coordinate system is mainly used for attitude calculations and to describe relative positions and directions of the satellite. Figure 3.3 shows the system axes.

Figure 3.4. Orbit Reference Frame (Nadir pointing frame). 3.2.2 Attitude Representation

where the superscript ‘b’ indicates the satellite body fixed CS.

Similar to the dynamic equations, the kinematic equations can be written as quaternion quaternion and rates, Equations (3.13) and (3.14) can represent the kinematics of the spacecraft.

conditions given by Equation (3.22). leaderp where sys,i 18 the producible torque in the direction of the i-th axis. Finally, in order to avoid the activation of the actuators for small values of “““’Tyom,;, it is desirable to include a deadband

Figure 3.6. (a) Attitude Control Subsystem for a Leader Spacecraft and (b) Attitude Control Subsystem for a Follower Spacecraft. shows the complete ACS for the leader and the followers respectively. spacecraft are equipped with a system of Pulsed Plasma Thrusters (PPTs). Figure 3.6 (a) and (b) In Chapter 2, we have indicated that a controlled system is composed of the dynamic plant

(LES-8/9) [59], [60] and [94]. As shown in Figure 3.7, the main components are the capacitor, the Parallel-Plate Ablative PPT similar to the one utilized in the Lincoln Experimental Satellite 8/9

where x, is the position, x2 is the capacitor charge, x; is the velocity and x, is the current, v is the

we consider the so-called six-independent PPT configuration (refer to Figure 3.9). spacecraft where independent control actuation is achieved. Equations (3.36) and (3.37) show the

where T;,;; is the magnitude that the i-th thruster generates in the direction of the j-axis. The torque matrix and the total control torque that is applied to the satellite respectively:

Table 3.11. Nullspace Vectors for the Six-Independent PPT Configuration. Eth as shown in Equations (3.34) and (3.35) are specified in Table 3.TIII.

Figure 4.1. (c) Discharge Current with Measurement Noise (Healthy Scenario). equation (4.1). Figures 4.1 (a)-(c) show the behavior of capacitor voltage, thrust and discharge

Figure 4.2. (c) Discharge Current with Measurement Noise (Faulty Scenario 1). used. Figures 4.2 (a) — (b) show the resulting simulated responses.

Figure 4.3. (c) Discharge Current with Measurement Noise (Faulty Scenario 2). the width of the electrodes w = 0.0254 m is changed to w = 0.02195 m. For this faulty scenario,

faulty scenario, the resistance of the electrodes R,- = 0.00 Q is changed to R,, = 0.008 Q. Figures

detect if the pulse is healthy or faulty. The classification signature is called Health Status. The neural network considered in this thesis is a multilayer perceptron network whit neurons (corresponding to a single pulse of the thruster). The MAE is compared with the Threshold to having internal feedbacks. As presented in [105], [122]-[128], these special neurons allow the

details of the results obtained under this faulty scenario. are generated by the PPT, are correctly classified as healthy pulses. Table 4.VIII contains the

pulses generated. In the case of the PPT,, our FDI scheme classifies the first 28 pulses as healthy The actual pulses that are generated by the pair of thrusters are shown in Figure 4.15. The complete set of pulses generated by the PPT; is within the healthy interval (85 — 115) pN. On the scheme. According to this figure, the behavior of the PPT; is considered healthy for all the 64

are classified as faulty pulses. Table 4.[X shows the actual healthy and faulty pulses and the Table 4.1X. Actual and Detected Condition of the Generated Pulses (Faulty Scenario 2). other hand, only the first 8 pulses that are generated by the PPT; are healthy. The other 33 pulses

13 pulses are classified as faulty pulses. The 41 pulses generated by the PPT, are classified as

Figure 4.17. Actual Thrust Generated by PPT; and PPT, Under Faulty Scenario 3.

These results indicate that the proposed FDI scheme is capable of correctly classify the 96.09% of the healthy pulses and the 57.63% of the faulty pulses. The accuracy level (88.89%)

Figure 5.1. (a) Angular Rotation (in quaternion) of s/cy w.r.t. s/c; (Healthy Scenario).

Figure 5.1. (c) Sequence of the Pulses (+1/0/-1 states) of s/cy; w.r.t. s/c; (Healthy Scenario). racking errors, settling times and the number of generated pulses for s/c are shown in

Due to the loss of elasticity in the spring of the thruster PPT, the amount of force produced by Figure 5.2. (b) Angular Velocity (in degree/seconds) of s/cj; w.r.t. s/c; (Faulty Scenario 1). 5.2.2 Simulations for the Faulty Scenario 1

Table 5.VIII shows the performance parameters for s/cj under the Faulty scenario 1 conditions. Follower s/cy is not subjected to a faulty actuator, therefore its results are the same as

Figure 5.3. (b) Angular Velocity (in degree/seconds) of s/c w.r.t. s/c; (Faulty Scenario 2). (sedimentation) may produce force pulses 15% greater than the nominal values. Figures 5.3 (a}+

The performance parameters for s/cy under the Faulty scenario 2 conditions are shown in

reduction in the amount of thrust produced. A conductivity fault in the thruster PPT; of s/cy

filters and hyperbolic tangent sigmoid and linear activation functions for the neurons in the

After the training phase is completed, the parameters of the dynamic neural networks are

Similar to the procedure performed in the training phase, in the validation phase we simulate

The data set generated from the missions 1-6 include only the healthy data, therefore the

The number of pulses that are generated by PPT; and PPT) are 10 and 2890, respectively Figure 5.12 shows the real effect that this fault produces on the thruster PPT, and Table

5.XVI presents the results that are obtained by using our proposed FDI scheme. torques in +y and —y axes, respectively (see Chapter 3, Section 3.3.2).

thrusters PPT;/PPT,. The other two pairs of thrusters are identified as healthy.

threshold value, the pair of thrusters responsible for the rotational maneuvers around that axis is “High Level” FDI scheme is an efficient fault detection tool. Using the relative attitude variables,

Now, with the Logic Threshold Selection Block specified, the integrated FDI scheme is

residuals (MAE) are generated and then using the corresponding Threshold values the health

As shown in Figure 6.6, the first stage of our proposed “Integrated” FDI scheme estimates utside the healthy zone. On the other hand, PPT> did not generate pulses during the entire

variables to generate the corresponding residuals (SAE;). The results are shown in Table 6.V. the three components of the angular velocity and then compares the estimates with the actual

for PPT; and PPT) for all the simulated period. According to these status signals the entire set of

For the sake of brevity, figures such as Figures 6.6 to 6.8 are not shown for the following

The fault is declared in the pair of thrusters PPT;/ PPT;. Based on the number of pulses generated by the pair, the Logic Threshold Selection Block especifies the Threshold values for

The fault is declared in the pair of thrusters PPT;/ PPTs. Based on the number of pulse: generated by the pair, the Logic Threshold Selection Block especifies the Threshold values fc corresponding residuals (SAE,). The results are shown in Table 6.XI. 90th thrusters. Table 6.XII shows the resulting Thresholds. angular velocity and then compares the estimates with the actual variables to generate the

Figure 6.16. Misclassification of Pulses Generated by PPT; (Fault Scenario 3).

The fault is declared in the pair of thrusters PPT;/ PPT. Based on the number of pulses that

Fault Scenario 5 Results: Figures 6.19 shows the actual reduction in the amount of the torque generated by the thruster PPT; due to the fault injected after t = 400 sec. Meanwhile, some of the pulses that are generated by PPT; are outside the healthy zone, and the pulses that are generated by PPT, are healthy. The first stage of our proposed “Integrated” FDI scheme estimates the three

The fault is declared in the pair of thrusters PPT3/ PPT;. Based on the number of pulses that

The fault is declared in the pair of thrusters PPT;/ PPTs. Based on the number of pulses that

Fault Scenario 7 Results: Figures 6.25 shows the actual augmentation in the amount of the torque that is generated by the thruster PPT; due to the fault injected after t = 300 sec

Figure 6.25. Actual Sequence of Pulses Generated by PPT,/ PPT, for the Fault Scenario 7. The fault is declared in the pair of thrusters PPT;/ PPT;. Based on the number of pulses that are generated by the pair, the Logic Threshold Selection Block defines the Threshold values for

6.4. Confusion Matrix Results for the Integrated FDI Scheme The performance of our proposed “Integrated” FDI scheme is studied using the Confusion Matrix as described in Chapter 4, Section 4.4. Table 6.X XIX contains the results for the four Confusion Matrix parameters evaluated for each of the fault scenarios. The last row of this table shows the

And last, but definitely not least, I would like to thank my amazing wife Karla for her infinite (MIDEPLAN) for economically supporting me during the entire period of study and research.

important and research efforts in this field are needed.

L,+L,+Ly-. Equations (3.32) and (3.33) provide expressions for R, and L,,(t) respectively, and the

chapters of this thesis to simulate attitude maneuvers under normal and abnormal PPT’s

equation (4.1). 4.2.4 PPT Operation Under Faulty Scenario 3

equation (4.1). The simulations for the above three faulty scenarios generate faults with severity of 15%. which is 15% less than the nominal value and is just bellow the lower limit that is defined by

criterion. The mse for the entire data set is calculated and when mse < t.c. the training is stopped. until a termination criterion (t.c.) is satisfied. For t.c. we used the mean square error (mse

design an alternative fault diagnosis and health monitoring system.

Table 5.VI. Expected Settling Time and Tracking Error Values for s/cj and s/cp. simulations that are performed in this chapter have a final simulation time (tjna) of 3000 sec. he faulty cases described in Chapter 4 are used to simulate different faulty scenarios

Table 5.VIH. Since s/cy; and s/cz are identical spacecraft, the results of the above simulation are

the same as that of the healthy case (Table 5. VII). able 5. VIX. Since the follower s/cp is not subjected to faulty actuators therefore the results are

shown in Table 5.VX, one can conclude that the performance of the formation flying system

references that are expressed in the quaternions. The modeling capability of our DNN approach is illustrated in Figures 5.9 — 5.10. In these

Section 3.3.2, PPT; and PPT> generate torques in the +x and —x axes, respectively. is “considerably” higher than the amount of pulses generated by the other actuator (PPT)).

Finally, from this simulated faulty scenario, we can see that when thruster PPT; generates Based on the results shown in Table 5.XVII, we detect the presence of a fault in the pair of less torque than expected, the amount of pulses generated by this actuator is ‘considerably’ higher

represents a promising detection capability, however as far as the isolation purposes are

.IIT are used to represent the faulty pair of thrusters that is detected by the “High Level” FD]

Table 6.[I. Definition of Variables used by the “Integrated” FDI Scheme when PPT;./PPT, is the Faulty Pair of Thrusters.

Table 6.IV. Health Analysis Conditions Considered by Our Proposed “Integrated” FDI Approach. parameters of the faulty PPT as shown equations (6.1) — (6.4), namely details are shown in Table 6.IV. The injected faults are simulated by modifying the corresponding

Table 6.VIL. “Integrated” FDI Scheme Results for s/ey, under Fault Scenario 1. number 1. Table 6.VII shows the final health monitoring results that are obtained by using our

first stage of our proposed “Integrated” FDI scheme estimates the three components of the

Table 6.XIV. Health Analysis Results Obtained by our Proposed FDI approach for the Fault Scenario 4. corresponding residuals (SAE;). The results are shown in Table 6.XIV.

Table 6.XVII. Health Analysis Results Obtained by our Proposed FDI Approach for the Fault Scenario 5. generate the corresponding residuals (SAE;). The results are shown in Table 6.XVII. components of the angular velocity and then compares the estimates with the actual variables t¢

Table 6.XXX. Performance Results for Our Proposed “Low Level” and “Integrated” FDI Schemes. with the results that are obtained in Chapter 4.

misclassification rate as expressed as False Healthy and False Faulty parameters is quite

actuator or two or more actuators affected by faults concurrently) and the development of fault

descriptionView Paper arrow_downwardDownload

Minimizing Makespan on Identical Parallel Machines Using Neural Networks

by Souad TAHRAOUI

2024, Lecture Notes in Computer Science

A heuristic algorithm that uses iteratively LPT and MF approaches on different job and machine sets constructed by using the current solution is developed to solve a classical multiprocessor scheduling problem with the objective of... more

Table 1. Summary of the experimental frameworks errors are averaged fora group of 100 instances. The Tables also report the number of instances (out of 100) that have been solved to optimality (sub-columns labeled “o”). This number, for the families El, E2, and E3, was determined by a comparison to the lower bound value L,, whereas for the family E4 it was determined by a comparison to the optimum value. In Tables 2-5, the last rows report the overall average percentage errors on all instances (columns “gap”), and the total number of instances that have been solved to optimality (columns “o”). Finally, com- putational times are not reported given these were fractions of a second for all the algorithms.

Table 2. Computational results for E1 Copyright © 2015, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited

The results for El (Table 2) show that DJMS and LISTFIT algorithms always out- perform all other approaches in terms of the considered measures of accuracy (number of optimal solutions and average percentage er- rors). By comparing DJMS and LISTFIT, Table 2 shows that DJMS finds the optimal solution in 996 out of 1800 test problems, whereas the LISTFIT algorithm finds the optimum in 949. Moreover, the overall average percentage error of the new algorithm is practically identical to that of LISTFIT, but it is smaller than those of other algorithms. With respect to the average percentage errors, D/MS outperforms LISTFIT in 9 out of 18 experimental points, LJSTFIT outperforms DJMS in 3 points, while in the other 6 both methods perform identically. With respect to the number of optimal solu- tions, DJMS outperforms LISTFIT in 8 out of 18 experimental points, LJSTFIT outperforms DJMS in one point, while in the other 9 both methods perform identically. 24 International Journal of Operations Research and Information Systems, 6(1), 19-29, January-March 2015 considered measures of accuracy (number of

Table 4. Computational results for E3 International Journal of Operations Research and Information Systems, 6(1), 19-29, January-March 2015 25

descriptionView Paper arrow_downwardDownload

Minimizing Makespan on Identical Parallel Machines Using Neural Networks

by Houari Habiba

2024, Lecture Notes in Computer Science

descriptionView Paper arrow_downwardDownload

State space neural networks for short term rainfall-runoff forecasting

by Tsung-Yi Pan

2024, Journal of Hydrology

Rainfall-runoff processes are dynamic systems that are better described by a dynamic model. In this paper, a specific dynamic neural network, called state space neural network (SSNN), is modified to perform short term rainfall-runoff... more

descriptionView Paper arrow_downwardDownload

Robust Growing Hierarchical Self Organizing Map

by Héctor Allende

2024, Lecture Notes in Computer Science

The Growing Hierarchical Self Organizing Map (GHSOM) was introduced as a dynamical neural network model that adapts its architecture during its unsupervised training process to represents the hierarchical relation of the data. However,... more

descriptionView Paper arrow_downwardDownload

Robust Growing Hierarchical Self Organizing Map

by Héctor Allende Olivares

2024, Lecture Notes in Computer Science

descriptionView Paper arrow_downwardDownload

Robust Growing Hierarchical Self Organizing Map

by Sebastian Moreno

2023, Lecture Notes in Computer Science

descriptionView Paper arrow_downwardDownload

Towards Robustness in Neural Network Based Fault Diagnosis

by Krzysztof Patan

2023, International Journal of Applied Mathematics and Computer Science

Towards Robustness in Neural Network Based Fault DiagnosisChallenging design problems arise regularly in modern fault diagnosis systems. Unfortunately, classical analytical techniques often cannot provide acceptable solutions to such... more

descriptionView Paper arrow_downwardDownload

An intelligent character recognizer for Telugu scripts using multiresolution analysis and associative memory

by Arun Pujari

2023, Image and Vision Computing

The present work is an attempt to develop a commercially viable and a robust character recognizer for Telugu texts. We aim at designing a recognizer which exploits the inherent characteristics of the Telugu Script. Our proposed method... more

descriptionView Paper arrow_downwardDownload

Some stability properties of dynamic neural networks

by ALEJANDRO CRUZ SANDOVAL

2023, IEEE Transactions on Circuits and Systems I: Fundamental Theory and Applications

In this paper, the passivity-based approach is used to derive a tuning algorithm for a class of dynamic neural networks. Several stability properties, such as passivity, asymptotic stability, input-to-state stability and bounded... more

descriptionView Paper arrow_downwardDownload

A comment on "Comments on 'Necessary and sufficient condition for absolute stability of neural networks

by Vedat Tavsanoglu

2023, IEEE Transactions on Circuits and Systems I: Fundamental Theory and Applications

This letter points out that a new majorization of the derivative of the Lyapunov function used in the above letter, 1 gives a new, weaker condition on the interconnection matrix that is sufficient for the global asymptotic stability (GAS)... more

descriptionView Paper arrow_downwardDownload

Nonlinear adaptive trajectory tracking using dynamic neural networks

by Edgar Adan Jara Sanchez

2023, IEEE Transactions on Neural Networks

In this paper the adaptive nonlinear identification and trajectory tracking are discussed via dynamic neural networks. By means of a Lyapunov-like analysis we determine stability conditions for the identification error. Then we analyze... more

descriptionView Paper arrow_downwardDownload

Identification and control of unknown chaotic systems via dynamic neural networks

by Edgar Adan Jara Sanchez

2023, IEEE Transactions on Circuits and Systems I: Fundamental Theory and Applications

Identification and control problems for unknown chaotic dynamical systems are considered. Our aim is to regulate the unknown chaos to a fixed point or a stable periodic orbit. This is realized by following two contributions. First, a... more

Fig. 1. (a) Identification results. (b) Identification results. (c) Identification results. (0.2/1 + e °?*) — 0.05. P = diag(20, 20,20), r = 0.01. The identification results are shown in Fig. 1.

result, it is not easy to use them in any control loop. If apply the derived sliding mode identifier to the local optimal controller, we can avoid these large deviations.

(3) Chua’s Circuit: This is an interesting electronics system that displays rich and typical bifurcation and chaotic phenomena, such as double scroll and double hook [2]. To study the controlled circuit, we formulate its equation in the following form:

descriptionView Paper arrow_downwardDownload

Input Constraints Handling in an MPC/Feedback Linearization Scheme

by Richard Stobart

2023, International Journal of Applied Mathematics and Computer Science

Input Constraints Handling in an MPC/Feedback Linearization SchemeThe combination of model predictive control based on linear models (MPC) with feedback linearization (FL) has attracted interest for a number of years, giving rise to... more

descriptionView Paper arrow_downwardDownload

State estimation for a class of nonlinear differential games using differential neural networks

by Emmanuel Sanchez Garcia

2023, Proceedings of the 2011 American Control Conference

This paper deals with the problem of the state estimation for a certain class of nonlinear differential games, where the mathematical model of this class is completely unknown. Being thus, a Luenberger-like differential neural network... more

descriptionView Paper arrow_downwardDownload

Evolution and Development of Modular Control Architectures for 1D Locomotion in Six-legged Animats

by Jean-Arcady Meyer

2023, Connection Science

An evolutionary approach is used to design neural control architectures for six-legged animats. Using a geometry-oriented variation of the cellular encoding scheme and syntactic constraints that reduce the size of the genetic search... more

In this paper tion (DIVIDE) makes it possible for a mother cel daughter cell. A direction parameter (cv) and a dis ter (7) associated with that instruction specify the daughter cell to and one new afi current one is t we use a small set of general, low level, de- velopmental instructions (Table I). A cellular division instruc- to generate a tance parame- position of the be created in the coordinates of the local frame attached to the mother cell. Then, the local frame associated with the daughter cell is centered on that cell and is oriented the same way as the mother cell’s frame (Figure 1). Two instruc- tions (GROW and DRAW) respectively create one new efferent ferent connection. The cell to be linked to t by the instruction parameters (Figure 1). However, no conne tion is created i synaptic weight ify the values o instruction DIE f the target is outside the substrate’s limits. T f a cell’s time constant 7 and bias b. Lastly, t causes a cell to die. he closest to a target position that is specified of a new connection is given by the parameter w. Two additional instructions (SETTAU and SETBIAS) spec- in ne c- ne ne

To slow down convergence and to favor the apparition of eco- logical niches, we use a steady-state evolutionary algorithm that involves a population of N randomly generated well-formed programs distributed over a circle and whose mode of opera- tion is outlined in Figure 3.

We use an incremental approach that takes advantage of the zeometrical nature of the developmental model.

ig. 5. Left: Each leg has two degrees of freedom. Thanks to the antagonis- tic PS- (Power Strike) and RS- (Return Strike) muscles, the leg can rotate around an axis (Ay) orthogonal to the plane of the figure. The UP- and DOWN-muscles allow the position of the foot F' to be translated along the (Az) axis. The resting lengths of the tunable PS-, RS- and UP-muscles depend on the activity levels of the corresponding motoneurons u. The rest- ing length of the DOWN-muscle is supposed to be non-tunable. Right: The muscle model. A muscle is modeled as a spring of set stiffness & and of (possibly tunable) resting length / (after [30]).

x. 8. Left: The best subprogram 6 found in Experiment 2 (parameter values are not shown). This subprogram generates a tripod gait and is called LOCO! thereafter. Right: The corresponding artificial neural network after useless interneurons and connections have been pruned. Solid lines are excitatory connec tions, dotted lines are inhibitory connections. Fan-in connections arrive at the top and fan-out connections depart from the bottom of each neuron. The network contains 38 interneurons and 100 connections.

cursor cell of Module 2. Such subprograms can create up to four neurons and 16 connections. Figure 15 depicts the initial setup for the developmental process.

Fig. 16. Behavior of a good individual in the three phases of the evaluation. That individual responds correctly to both stimuli. The conditional protocol (described in the text) prevents the animats from just predicting the time of occurence of the stimuli.

descriptionView Paper arrow_downwardDownload

Modulation of neural connectivity during tongue movement and reading

by george james

2023, Human Brain Mapping

In a functional magnetic resonance imaging (fMRI) study, a novel connectivity analysis method termed within-condition interregional covariance analysis (WICA) was introduced for investigation into brain modulation during tongue movement... more

descriptionView Paper arrow_downwardDownload

Dynamical neural network organization of the visual pursuit system

by Curt Deno

2023, IEEE Transactions on Biomedical Engineering

descriptionView Paper arrow_downwardDownload

Nonlinear system identification using additive dynamic neural networks-two on-line approaches

by Roberto Griñó

2023, IEEE Transactions on Circuits and Systems I: Fundamental Theory and Applications

This paper proposes a class of additive dynamic connectionist (ADC) models for identification of unknown dynamic systems. These models work in continuous time and are linear in their parameters. Also, for this kind of model two on-line... more

descriptionView Paper arrow_downwardDownload

Robust Growing Hierarchical Self Organizing Map

by Dulce Omaña

2023, Lecture Notes in Computer Science

descriptionView Paper arrow_downwardDownload

Input-to-state stability (ISS) analysis for dynamic neural networks

by Edgar Sanchez

2023, IEEE Transactions on Circuits and Systems I: Fundamental Theory and Applications

In this paper a novel approach to assess the stability of dynamic neural networks is presented. Using a Lyapunov function, we determine conditions to guarantee input-to-state stability (ISS) which also ensures global asymptotic stability... more

descriptionView Paper arrow_downwardDownload

Dynamic neural networks for logic formula

by HAIMING (일반대학원 국제관계학과) LI,

2023

In this article, a kind of calculus method that is used to determine the truth-values of propositional logic formulae by means of the dynamic neural networks is proposed. The method can be executed mechanically and extended to fuzzy logic... more

descriptionView Paper arrow_downwardDownload

Nonlinear adaptive trajectory tracking using dynamic neural networks

by jose perdomo perez

2023, IEEE Transactions on Neural Networks

descriptionView Paper arrow_downwardDownload

GPS/INS integration utilizing dynamic neural networks for vehicular navigation

by Mohamed Bayoumi

2023, Information Fusion

Recently, methods based on Artificial Intelligence (AI) have been suggested to provide reliable positioning information for different land vehicle navigation applications integrating the Global Positioning System (GPS) with the Inertial... more

Fig. 1. Input-Delayed Neural Network architecture and single neuron calculations. 3. Methodology

Fig. 3. Test trajectories I showing the locations of GPS outages. The performance of the proposed IDNN-based GPS/INS integra- tion module was examined with two field tests (Tests I and Il) involving two different navigation systems. Test I utilizes CIMU navigation grade INS and a NovATel OEM-4 GPS receiver. Over the whole trajectory of Test I (shown in Fig. 3), no natural GPS out- ages were detected, and thus the position and velocity information obtained from this system in the differential mode will be used as a reference when evaluating the accuracy of the system. On the other hand, Test II utilizes a tactical grade INS system (IMU-Honey- well HG1700) and a NovAtel OEM4 GPS receiver. The trajectory (shown in Fig. 4) was totally different from that of Test I. A mini- mum of seven GPS satellites was available throughout Test II and

Fig. 2. Time line index for the model’ stages sequences.

Fig. 4. Test trajectories II showing the locations of GPS outages.

Fig. 5. The model performance during training, early stopping and prediction utilizing different value of X. The parameters associated with the IDNN modules change through the update procedure to achieve a predefined error target between the actual INS error (dPinsjcps OF 5Vinsjcps) and the IDNN output (dPins or OVjns). In parallel, the early stopping criterion is ap- plied utilizing input vectors from the INS and GPS position or velocity components on which the IDNN module was not trained. This input data (Piys or Vins) is presented to the trained IDNN mod- ule to examine its capability to accurately predict the correspond- ing output data (dPiys or 6Vins). AS Shown earlier in Fig. 2, a fraction X of a data window of size W (i.e. X « W) has to be chosen for the early stopping procedure. It is important to choose this part of the input data so that it is big enough to be representative of the The proposed IDNN - based GPS/INS module is examined and analyzed during both the update (training) and prediction modes of operations using the navigational grade INS data of Test I. The benefits of early stopping and the impact of the number of input delay elements are also explored. Performance validation on an- other trajectory while using other types of navigation systems is performed using the tactical grade INS data of Test II.

Fig. 6. The position error distribution along GPS outages as results from IDNN model Test I.

Fig. 7. The position error distribution along GPS outages as results from IDNN model Test II.

The highest position error utilizing different input delay sequences. Table 1

Summary of average position errors for 40 and 100s GPS outages and the EIR% using different models utilizing Test I. Table 2

Summary of position errors for 100 s GPS outages using different models utilizing Test II. Table 3 the output INS position or velocity error is modeled based on the present and past samples of the corresponding INS position or velocity. Consequently, the time dependence feature of the INS er- rors was appropriately modeled. In addition, the common over-fit- ting problems during the training procedure of the IDNN module were avoided by adopting both the early stopping criterion and the regularization procedure. The impact of different delay ele- ments at the input layer on the overall positioning accuracy was investigated. In comparison to conventional and recent AI - based GPS/INS integration techniques, the IDNN module showed superior performance during both short and long GPS outages. ASFP one the results are shown in Table 3 for 100s GPS outages. Obviously, IDNN module outperforms KF for all GPS outages and offers much better accuracy in most of the outages. On the other hand, ASFP provides less position error than IDNN for only four cases (highlighted in Table 3). However, one can notice that, for any of these four cases, there was only slight improvement (less than 2 m position error) over the IDNN module. Moreover, superior performance over the ASFP method can be noticed in Table 3 for most of the 8 GPS outages for both position components.

descriptionView Paper arrow_downwardDownload

THE ROLE OF SENSORY-MOTOR COORDINATION - Identifying Environmental Motion Dynamics with Dynamic Neural Networks

by wan amir fuad wajdi othman

2023, Proceedings of the Fifth International Conference on Informatics in Control, Automation and Robotics Service

descriptionView Paper arrow_downwardDownload

Modulation of neural connectivity during tongue movement and reading

by George James

2023, Human Brain Mapping

descriptionView Paper arrow_downwardDownload

The dynamics of integration and separation: ERP, MEG, and neural network studies of immediate repetition effects

by Brion Woroch

2023, Journal of Experimental Psychology: Human Perception and Performance

This article presents data and theory concerning the fundamental question of how the brain achieves a balance between integrating and separating perceptual information over time. This theory was tested in the domain of word reading by... more

descriptionView Paper arrow_downwardDownload

Application of the recurrent multilayer perceptron in modeling complex process dynamics

by Alexander G. Parlos

2023, IEEE Transactions on Neural Networks

A nonlinear dynamic model is developed for a process system, namely a heat exchanger, using the recurrent multilayer perceptron network as the underlying model structure. The recurrent multilayer perceptron is a dynamic neural network,... more

Fig. 1. Multilayer perceptron architecture. encountered in process systems. Fig. 1 depicts the RMLP network architecture used in this study.

Fig. 2. Transient responses for 10% to 15% of full power step steam flow rate increase with actuator and sensor noise.

Fig. 3. Transient responses for 10% to 15%. of full power step steam flow rate increase with actuator, process, and sensor noise.

Fig. 4. Normalized residuals for 10% to 15% of full power step steam flow rate increase with actuator, process, and sensor noise.

Fig. 6. Transient responses for 75% to 25% of full power ramp steam flow rate decrease (—3%/min) with actuator, process, and sensor noise; no on-line learning. Fig. 5. Transient responses for 30% to 20% of full power step steam flow rate decrease with actuator noise.

Fig. 7, Transient responses for 75% to 25% of full power ramp steam flow rate decrease (—3%/min) with actuator, process, and sensor noise; on-line learning.

Fig. 9. Transient responses for 10% to 15% of full power step steam flow rate increase with actuator and sensor noise; 12% boiling length uncertainty. Fig. 8. Transient responses for 10% to 15% of full power step steam flow rate increase with high level of actuator and sensor noise. model. Even though more complex in nature than previously reported developments, the enhanced learning capability of

EmpiricaL Process MODEL VALIDATION TEesT MATRIX TABLE I

descriptionView Paper arrow_downwardDownload

Projectional differential neural network observer with stable adaptation weights

by Alejandro Garcia

2023, 2008 47th IEEE Conference on Decision and Control

A class of dynamic neural network (DNN) observers involving a projection operator inside is considered. Such observers seem to be useful when an uncertain nonlinear system, affected by external perturbations, keeps its states in an a... more

descriptionView Paper arrow_downwardDownload

Classification of equinus in ambulatory children with cerebral palsy - discrimination between dynamic tightness and fixed contracture

by Herbert Witte

2023, Gait & Posture

In this study a generalised dynamic neural network (GDNN) was designed to process gait analysis parameters to evaluate equinus deformity in ambulatory children with cerebral palsy. The aim was to differentiate dynamic calf muscle... more

descriptionView Paper arrow_downwardDownload

Dynamic versus static neural network model for rainfall forecasting at Klang River Basin, Malaysia

by Mohd Taha

2023, Hydrology and Earth System Sciences

Rainfall is considered as one of the major components of the hydrological process; it takes significant part in evaluating drought and flooding events. Therefore, it is important to have an accurate model for rainfall forecasting.... more

where x is the output from the previous neuron, w is the weight of the output and k is the slope steepness of the sig- moid function. Extra neuron Xo is added in the input layer and hidden layer with the output value of 1. This input is called bias and its function is to stabilize computed output between 0 and 1. It does not have a link from the previous neuron. The network architecture of the MLP-NN is shown in Fig. 1. It contains three layers: input, hidden and output layers. Each layer consists of one or more neurons and there are two types of them. First, there are passive neurons that consider the input and output data. Another type is an active neuron that computes data input using Activation Transfer Function (ATF) and produces an output. The most common use of ATF in the hidden and output neuron is the sigmoid func- tion (Zhang et al., 1998; Fernando et al., 2000). The input into an active neuron is a summation of the previous neu- ron’s weighted outputs and the output is a result of applying the sigmoid function to the input. The process is shown in Fig. | and the equations for the input and output are

Fig. 2. Architecture of radial basis function neural network. Fig. 1. MLP-NN model architecture.

A. El-Shafie et al.: Dynamic versus static neural network model for rainfall forecasting

chosen as a sigmoid activation function of neuron i: The output of the IDNN, assuming that it has one output neu- ron j, a single hidden layer with m hidden neurons, and one input variable as shown in Fig. 4, is given by

Fig. 5. IDNN model architecture. Fig. 4. Input delay neural network architecture and single neuron calculations.

Fig. 6. Local authorities within Klang River Basin.

Fig. 7. Monthly actual rainfall records on Klang River on for period 1987-2008.

Fig. 8. Weekly actual rainfall records on Klang River on for period 1997-2008.

Fig. 9. Neural network model architecture utilized for rainfall fore- casting.

where 7, =number of peak rainfall greater than one-third of the mean peak rainfall observed; 7; =number of low rain- fall lower than one-third of the mean low rainfall observed. Coulibaly (2001) reported that both PFC and LFC provide better performance indicators for assessment of the forecast- ing model performance for the extreme rainfall events. As the model can provide low PFC or LFC, the model represents better fit.

Fig. 11. Training curve for monthly basis using the MLP-NN model.

Fig. 12. Neural network performance “RMSE and maximum error%” utilizing different architecture, (a) one hidden layer (b) and (c) two hidden layers.

Fig. 13. Performance of the MLP-NN model for monthly and weekly rainfall forecasting during training and testing stages.

Fig. 14. Performance of the RBFNN model for monthly and weekly rainfall forecasting during training and testing stages.

Fig. 15. Performance of the IDNN model for monthly and weekly rainfall forecasting during training and testing stages

Fig. 16. Forecasted and actual rainfall at Klang River (weekly ba- sis).

Fig. 17. Forecasted and actual rainfall at Klang River (monthly basis).

Table 1. Maximum RE% and correlation coefficient associated with the output of MLP-NN model on a weekly/monthly basis for years 2007 and 2008. model-based MLP-NN method, Table 1 illustrates the perfor- mance measures for both scenarios. It can be observed that scenario B outperformed scenario A and provides a better ac- curacy level. It might be due to that scenario A includes the rainfall record of the same month of the previous year, which might not correlate and relate to rainfall values of the same month in the current year, and thus, the model scenario A provides relatively poor forecasting accuracy.

Table 2. Maximum RE% and correlation coefficient associated with the output of ANN and ARMA models on a weekly/monthly basis for years 2007 and 2008. ment of the spread of the RBFNN model configuration is (as described in Sect. 2.1.2) the only parameter to be obtained. The spread (step size) is achieved by trial and error as well. The optimal values of the spread were found to be equal to 0.07 for the monthly and 0.03 for the weekly model. remarkable improvements in the correlation coefficient at both the weekly and monthly basis.

Table 3. RMSE and maximum RE for monthly and weekly rainfall forecasting model utilizing IDNN. for other river basins that might require introducing a higher order level. value of RMSE is of the IDNN model in the weekly rainfall forecasting model with only 7.3mm. It could be remarked that, generally, the performance for the weekly rainfall fore- casting is better than monthly rainfall forecasting. This is due to the inadequacy of historical data records on the monthly basis, which is 261 records, while for weekly, historical 621 rainfall records; thus, the model could capture most of tem- poral dimension of rainfall pattern and be able to provide lower forecasting error. In addition, it is obvious that the optimal results were received when using the IDNN method. Furthermore, with one-time step input delay, the IDNN is sufficient to achieve a significant level of accuracy. With re- spect to this observation, in Klang River Basin rainfall pat- tern, it might be the temporal dimension feature of rainfall that is in second order level. However, it could be inadequate

Table 5. Models performance for matching the rainfall consequences. could also be applicable to other studies in other river basins with different time step input delay, according to how far is the temporal dimension of rainfall pattern at this river basin. The results of the present study also show that the pro- posed IDNN achieves better accuracy than the MLP-NN and RBENN for the extreme rainfall pattern events; however, the IDNN is still less accurate if compared with the forecasting accuracy for non-extreme events. In addition, the proposed IDNN showed better accuracy for mimicking the consecutive rainfall pattern than the classical MLP-NN and /or RBFNN model.

Table 4. IDNN performance based on the peak and low flow error criteria for monthly and weekly forecasting

descriptionView Paper arrow_downwardDownload

The Mr. ArmHandOne Project A mazes roamer robot

by F. Pirri

2022

We present in this paper the crucial components of "Mr.ArmHandOne" project. We shall discuss topics concerning the robot architecture; the robot building process, including the mechanics and electronics solutions to the robot structure;... more

descriptionView Paper arrow_downwardDownload

Passivity Analysis for Neuro Identifier with Different Time-Scales

by Alejandro Sandoval

2022, Lecture Notes in Computer Science

descriptionView Paper arrow_downwardDownload

Passivity Analysis of Dynamic Neural Networks with Different Time-scales

by Alejandro Sandoval

2022, Neural Processing Letters

Dynamic neural networks with different timescales include the aspects of fast and slow phenomenons. Some applications require that the equilibrium points of the designed networks are stable. In this paper, the passivity-based approach is... more

descriptionView Paper arrow_downwardDownload

Some stability properties of dynamic neural networks

by Alejandro Sandoval

2022, IEEE Transactions on Circuits and Systems I: Fundamental Theory and Applications

descriptionView Paper arrow_downwardDownload

Trajectory generation and modulation using dynamic neural networks

by Pablo Zegers

2022, IEEE Transactions on Neural Networks

descriptionView Paper arrow_downwardDownload