Weights Of Neural Network Research Articles

On Sequential Bayesian Inference for Continual Learning.

Sequential Bayesian inference can be used for continual learning to prevent catastrophic forgetting of past tasks and provide an informative prior when learning new tasks. We revisit sequential Bayesian inference and assess whether using the previous task's posterior as a prior for a new task can prevent catastrophic forgetting in Bayesian neural networks. Our first contribution is to perform sequential Bayesian inference using Hamiltonian Monte Carlo. We propagate the posterior as a prior for new tasks by approximating the posterior via fitting a density estimator on Hamiltonian Monte Carlo samples. We find that this approach fails to prevent catastrophic forgetting, demonstrating the difficulty in performing sequential Bayesian inference in neural networks. From there, we study simple analytical examples of sequential Bayesian inference and CL and highlight the issue of model misspecification, which can lead to sub-optimal continual learning performance despite exact inference. Furthermore, we discuss how task data imbalances can cause forgetting. From these limitations, we argue that we need probabilistic models of the continual learning generative process rather than relying on sequential Bayesian inference over Bayesian neural network weights. Our final contribution is to propose a simple baseline called Prototypical Bayesian Continual Learning, which is competitive with the best performing Bayesian continual learning methods on class incremental continual learning computer vision benchmarks.

Optimized neural network with self-adaptive MHBO algorithm for fault diagnosis

Bearing operation affects the usual operations of mechanical equipment. It is of much practical and theoretical concern to perform faults diagnosis in bearing. Accordingly, the extraction and election of faulty features aid in enhancing the accuracy of fault diagnosis. Nevertheless, the analysis endures from “(1) High dimension of the selected features, (2) Uncertainty of single sensor for data sampling.” Thereby, this work develops a fault diagnosis approach that contains two main phases “(1) Feature extraction and (2) Diagnosis.” At first, it extracts the “empirical wavelet transform, empirical mode decomposition, and wavelet transform” based features. Following this, the derived features are given to optimized neural network (NN) for diagnosis. Further, for progressing the diagnosis accuracy of the established scheme, the weights of NN are fine-tuned through self-adaptive marriage in honey bee optimization. At last, the supremacy of the presented approach is proved with respect to varied measures.

Intelligent routing algorithm for wireless sensor networks dynamically guided by distributed neural networks

Using reinforcement learning to adjust the power balance of sensor nodes dynamically is an essential approach for extending the lifetime of wireless sensor networks (WSNs), which makes sensor nodes adapt to changing environments by calculating optimal decision solutions in real time. However, as WSN grows, error propagation cannot keep up with topological changes in the networks, which prevents the sensor nodes’ routing decisions from being timely optimized. In this paper, we propose the DNN-HR algorithm, which optimizes the reinforcement learning process of WSN in three ways. First, this paper constructs a two-level learning structure. Discrete nodes perform reinforcement learning during the interaction with their neighboring nodes. Simultaneously, the neural networks that fit the V values guide the underlying nodes’ learning. Second, the algorithm divides the sensor networks into two-level subspaces according to their geographic locations, and cluster heads are elected in each subspace to form weight incremental propagation trees. Third, each sensor node sends the weight increments it learned to its first-level cluster head. At the first-level cluster head, the weight increments sent by multiple nodes are received and accumulated, and then the results are sent to the second-level cluster head. The base station acquires the weight increments of all the second-level cluster heads and then broadcasts the updated neural network weights to all the WSN nodes. Experiments determine the appropriate heuristic parameters. Compared to other algorithms, the experiments show that our proposed scheme significantly improves the packet delivery rate, node survival rate, and energy standard deviation of nodes.

Adaptive neural control with fast approximation for uncertain nonlinear systems: A novel composite learning approach

AbstractIn existing adaptive neural control approaches, only when the regressor satisfies the persistent excitation (PE) or interval excitation (IE) conditions, the constant optimal weights of neural network (NN) can be identified, which can be used to establish uncertainties in nonlinear systems. This paper proposes a novel composite learning approach based on adaptive neural control. The focus of this approach is to make the NN approximate uncertainties in nonlinear systems quickly and accurately without identifying the constant optimal weights of the NN. Hence, the regressor does not need to satisfy the PE or IE conditions. In this paper, regressor filtering scheme is adopted to generate prediction error, and then the prediction error and tracking error simultaneously drive the update of NN weights. Under the framework of Lyapulov theory, the proposed composite learning approach can ensure that approximation error of the uncertainty and tracking error of the system states converge to an arbitrarily small neighborhood of zero exponentially. The simulation results verify the effectiveness and advantages of the proposed approach in terms of fast approximation.

Prediction of Silicon Content in the Hot Metal of a Blast Furnace Based on FPA-BP Model

In the process of blast furnace smelting, the stability of the hearth thermal state is essential. According to the analysis of silicon content in hot metal and its change trend, the operation status of the blast furnace can be judged to ensure the stable and smooth operation of the blast furnace. Based on the error back-propagation neural network (BP), the flower pollination algorithm (FPA) is used to optimize the weight and threshold of the BP neural network, and the prediction model of silicon content is established. At the same time, the principal component analysis method is used to reduce the dimension of the input sequence to obtain relevant indicators. The relevant indicators are used as the input, and silicon content in the hot metal is used as the output, which is substituted into the model for training and utilizes the trained model to predict. The results show that the hit rate of the prediction model is 16% higher than the non-optimized BP prediction model. At the same time, the evaluation indicators and operation speed of the model are improved compared with the BP prediction model, which can be more accurately applied to predict the silicon content of the hot metal.

Compound learning based event-triggered adaptive attitude control for underwater gliders with actuator saturation and faults

This paper proposes an event-triggered compound learning adaptive attitude control scheme for underwater gliders (UGs) with unknown dynamics, actuator saturation, and actuator faults. To begin with, in the framework of adaptive dynamic surface control (DSC), a historical input–output data based neural network (NN) is employed to estimate and compensate for unknown dynamics, external disturbance and actuator faults as a whole. For the merit of the composite learning and integral concurrent learning techniques, the NN weights are updated in a compound learning fashion where instant tracking errors, prediction errors, and historical estimation errors are combined in the adaptive law, making the NN weights converge to their true value with a good transient performance. Furthermore, a hyperbolic tangent function is utilized to explicitly handle the input saturation constraint, and the singular problem is circumvented by a Nussbaum gain function. Finally, a novel event-triggering condition is designed to save communication resource and reduce energy consumption. Theoretical proofs are provided to ensure the semi-global uniformly ultimately bounded (SGUUB) stability of the closed-loop system. The effectiveness of the proposed scheme is verified by simulation studies under various conditions.

Identification and control of Maglev system using fractional and integer order PID controller

System identification techniques have proved to be the most effective methodologies for the modeling highly non-linear and system. For the purpose of real-time parameter estimation of a Maglev system, a Teaching Learning Based Optimization (TLBO) for updating the weights of Functional Link Artificial Neural Network (FLANN) model is proposed and implemented in this research. Moreover, we proposed a one & two-Degree of Freedom (one-DOF & two-DOF) Fractional Order PID (FOPID) controller, where the parameters are optimized by using the Teaching Learning Based Optimization (TLBO) and the recently proposed Black Widow Optimization (BWO) algorithm. To investigate the robustness of the proposed controller, a pulse signal disturbance is added at equal intervals of the output of the identified model of the Maglev system. It is found that the suggested two-DOF FOPID controller with TLBO performs better than its counterpart in terms of both in time domain specifications (i.e., maximum overshoot = 1.2648%, settling time = 1.3884 sec and rise time = 0.8685 sec) and in robustness analysis (i.e., system is sufficiently robust, because the infinity norms of the sensitivity and the complementary sensitivity functions of the system are less than two). The TLBO algorithm has performed better for both identification and optimization of controller parameter due to very less number of algorithmic parameter is as compared to other algorithm.

Development of a method of complex analysis and multidimensional forecasting of the state of intelligence objects

A method of complex analysis and multidimensional forecasting of the state of intelligence objects is proposed to increase the accuracy of their state assessment. The object of research is decision support systems. The subject of research is the process of decision-making in management problems using artificial intelligence methods. The hypothesis of research is to increase the efficiency of decision-making with a given assessment reliability. The proposed method is based on a combination of fuzzy cognitive and temporal models, an advanced cat swarm optimization algorithm and evolving artificial neural networks. The method has the following sequence of actions: ‒ input of initial data; ‒ processing of initial data taking into account uncertainty about the state of heterogeneous intelligence objects; ‒ construction of a fuzzy temporal ontological model of heterogeneous intelligence objects; ‒ conclusion on the state of heterogeneous intelligence objects; ‒ correction of the fuzzy temporal ontological model; ‒ building a fuzzy relational temporal cognitive model of heterogeneous intelligence objects and forecasting the state of the intelligence object; ‒ training knowledge bases on heterogeneous intelligence objects. The training procedure consists in learning the synaptic weights of the artificial neural network, the type and parameters of the membership function, as well as the architecture of individual elements and the architecture of the artificial neural network as a whole. The method makes it possible to increase the efficiency of data processing at the level of 18–25 % by using additional improved procedures. The proposed method should be used to solve the problems of evaluating complex and dynamic heterogeneous intelligence objects, characterized by a high degree of complexity.

Improvement of the solution search method based on the cuckoo algorithm

An improved method of finding solutions based on the cuckoo algorithm is proposed. The research object is the decision-making support systems. The research subject is the decision making process in management tasks using artificial intelligence methods. The hypothesis of the research is to increase the efficiency of decision making with a given assessment reliability. The proposed method is based on a combination of the cuckoo algorithm and evolving artificial neural networks. The method has the following differences: ‒ an additional processing of the source data takes place taking into account the uncertainty about the state of the control objects and the type of data noise about the state of the control object is additionally taken into account; ‒ the state model of the control object is adjusted taking into account the available computing resources of the system; ‒ added procedures to reduce the probability of detecting nests and reducing the length of the cuckoo’s step; ‒ knowledge bases about management objects are additionally taught. The training procedure consists in learning the synaptic weights of the artificial neural network, the type and parameters of the membership function and the architecture of individual elements and the architecture of the artificial neural network as a whole. The effectiveness of the proposed method was evaluated and it was established that the proposed modification provides a better value of the objective function compared to the results obtained by other authors and ensures the fulfillment of all restrictions. The specified example showed an increase in the efficiency of data processing at the level of 21–28 % due to the use of additional improved procedures. It is advisable to use the proposed method in decision making support systems of automated control systems.

Learning CHARME models with neural networks

In this paper, we consider a model called CHARME (Conditional Heteroscedastic Autoregressive Mixture of Experts), a class of generalized mixture of nonlinear (non)parametric AR-ARCH time series. The main objective of this paper is to learn the autoregressive and volatility functions of this model with neural networks (NN). This approach is justified thanks to the universal approximation capacity of neural networks. On the other hand, in order to build the learning theory, it is necessary first to prove the ergodicity of the CHARME model. We therefore show in a general nonparametric framework that under certain Lipschitz-type conditions on the autoregressive and volatility functions, this model is stationary, ergodic and $$\tau $$ -weakly dependent. These conditions are much weaker than those in the existing literature. Moreover, this result forms the theoretical basis for deriving an asymptotic theory of the underlying parametric estimation, which we present for this model in a general parametric framework. Altogether, this allows to develop a learning theory for the NN-based autoregressive and volatility functions of the CHARME model, where strong consistency and asymptotic normality of the considered estimator of the NN weights and biases are guaranteed under weak conditions. Numerical experiments are reported to support our theoretical findings.

Adaptive control for multi-agent systems with actuator fault via reinforcement learning and its application on multi-unmanned surface vehicle

In this paper, the adaptive cooperative control problem is studied for a class of second order multi-agent systems (MASs) with actuator fault via reinforcement learning (RL). The strategic utility function is approximated by the critic neural network (NN), and the uncertain dynamics in MASs are estimated by the action NN. The NN weight vectors are updated by employing the gradient descent strategy. Then, the distributed RL control strategy is developed to solve the consensus control design problem. In comparison with the existent RL control results, the actuator fault is taken into consideration in the controller design. The stability analysis is given based on the Lyapunov theory, and all signals in the closed-loop system are guaranteed to semi-globally uniformly ultimately bounded (SGUUB). Two simulation examples, including a multi-unmanned surface vehicle system, are presented to demonstrate the validation of this strategy.

Reconstruction of integrated hydraulic turbine characteristics curve based on classification weight neural network

To address the shortcomings in the drawing and processing of comprehensive characteristic curves of hydraulic turbines and the problems in the operation and management of small hydropower stations, this paper proposes a method for reconstructing the comprehensive characteristic curves based on classification weight neural network. Firstly, this study discretizes the comprehensive characteristic curve of the target turbine factory package to obtain multiple sets of discrete data points, and compares them with the actual operating data of small hydropower plants, and then pre-processes the operating data by the type and operating characteristics of the unit. Next, this paper uses XGBoost to classify the two types of data to obtain the leaf node optimal scores and uses the scores as the prediction weights of the deep neural network to grid the discrete points of the hydraulic turbine for prediction, and finally obtains the specific integrated characteristic graph of the unit. The method based on classification weight neural network can substantially improve the accuracy and reliability of composite characteristic curves while avoiding multivariate nonlinear mathematical problems, which provides a new method and idea for the reconstruction and expansion of the hydraulic turbine characteristic curve, and also provides a theoretical basis for the safe and economic operation of hydropower units.

Archive-based coronavirus herd immunity algorithm for optimizing weights in neural networks.

The success of the supervised learning process for feedforward neural networks, especially multilayer perceptron neural network (MLP), depends on the suitable configuration of its controlling parameters (i.e., weights and biases). Normally, the gradient descent method is used to find the optimal values of weights and biases. The gradient descent method suffers from the local optimal trap and slow convergence. Therefore, stochastic approximation methods such as metaheuristics are invited. Coronavirus herd immunity optimizer (CHIO) is a recent metaheuristic human-based algorithm stemmed from the herd immunity mechanism as a way to treat the spread of the coronavirus pandemic. In this paper, an external archive strategy is proposed and applied to direct the population closer to more promising search regions. The external archive is implemented during the algorithm evolution, and it saves the best solutions to be used later. This enhanced version of CHIO is called ACHIO. The algorithm is utilized in the training process of MLP to find its optimal controlling parameters thus empowering their classification accuracy. The proposed approach is evaluated using 15 classification datasets with classes ranging between 2 to 10. The performance of ACHIO is compared against six well-known swarm intelligence algorithms and the original CHIO in terms of classification accuracy. Interestingly, ACHIO is able to produce accurate results that excel other comparative methods in ten out of the fifteen classification datasets and very competitive results for others.

CNN-based Robust Sound Source Localization with SRP-PHAT for the Extreme Edge

Robust sound source localization for environments with noise and reverberation are increasingly exploiting deep neural networks fed with various acoustic features. Yet, state-of-the-art research mainly focuses on optimizing algorithmic accuracy, resulting in huge models preventing edge-device deployment. The edge, however, urges for real-time low-footprint acoustic reasoning for applications such as hearing aids and robot interactions. Hence, we set off from a robust CNN-based model using SRP-PHAT features, Cross3D [ 16 ], to pursue an efficient yet compact model architecture for the extreme edge. For both the SRP feature representation and neural network, we propose respectively our scalable LC-SRP-Edge and Cross3D-Edge algorithms which are optimized towards lower hardware overhead. LC-SRP-Edge halves the complexity and on-chip memory overhead for the sinc interpolation compared to the original LC-SRP [ 19 ]. Over multiple SRP resolution cases, Cross3D-Edge saves 10.32%~73.71% computational complexity and 59.77%~94.66% neural network weights against the Cross3D baseline. In terms of the accuracy-efficiency tradeoff, the most balanced version ( EM ) requires only 127.1 MFLOPS computation, 3.71 MByte/s bandwidth, and 0.821 MByte on-chip memory in total, while still retaining competitiveness in state-of-the-art accuracy comparisons. It achieves 8.59 ms/frame end-to-end latency on a Rasberry Pi 4B, which is 7.26× faster than the corresponding baseline.

Neural-network-based output feedback control for networked multirate systems: A bit rate allocation scheme

This paper deals with the neural-network (NN)-based output feedback control problem for a class of networked systems with unknown nonlinearities under the effects of bit rate constraints. Considering the physical conditions/requirements in practical applications, the sampling period of sensors is assumed to be different from the state updating period of the system. For the purpose of facilitating digital communications over networks, a group of encoders is utilized to convert the measurement signal into codewords with limited bit lengths. A so-called bit rate constraint is introduced to capture the bandwidth-limited nature of communication network. To handle the unknown nonlinearity of the multirate system, both the NN-based observer and NN-based controller are designed to generate the desired state estimates and control input signals, respectively. Then, a unified framework is established to analyze the boundedness of the estimation error and system state as well as the neural network weights. The effects of the bit rate constraint on the resultant control performance is also analyzed. Subsequently, sufficient conditions are derived to guarantee the existence of the required NN-based output feedback controller. A particle-swarm-optimization-based (PSO-based) algorithm is developed to co-design the desired controller parameter and the bit rate allocation strategy. Finally, an illustrative example is given to verify the effectiveness of the proposed control strategy.

Research on the health status evaluation method of rolling bearing based on EMD‐GA‐BP

AbstractTo more accurately evaluate the health state of rolling bearings, this paper proposes a health status evaluation method based on empirical pattern decomposition, genetic algorithm and BP neural network. Firstly, the vibration signal is decomposed by empirical mode decomposition (EMD) and the time domain features of each intrinsic mode function (IMF) component are extracted, and the signal‐to‐noise ratio (Snr) of the signal is improved effectively. Then, the initial threshold and weight of BP neural network are optimized by genetic algorithm, which effectively improves the Snr of the signal. Finally, the extracted features are input into the optimized BP neural network to realize the identification of different states of the bearing. The effectiveness of the method has been effectively verified in the bearing data of Case Western Reserve University bearing dataset and it has higher accuracy and robustness than other common evaluation methods.

Event‐triggered‐based integral reinforcement learning output feedback optimal control for partially unknown constrained‐input nonlinear systems

AbstractIn this paper, an adaptive output feedback event‐triggered optimal control algorithm is proposed for partially unknown constrained‐input continuous‐time nonlinear systems. First, a neural network observer is constructed to estimate unmeasurable state. Next, an event‐triggered condition is established, and only when the event‐triggered condition is violated will the event be triggered and the state be sampled. Then, an event‐triggered‐based synchronous integral reinforcement learning (ET‐SIRL) control algorithm with critic‐actor neural networks (NNs) architecture is proposed to solve the event‐triggered Hamilton–Jacobi–Bellman equation under the established event‐triggered condition. The critic and actor NNs are used to approximate cost function and optimal event‐triggered optimal control law, respectively. Meanwhile, the event‐triggered‐based closed‐loop system state and all the neural network weight estimation errors are uniformly ultimately bounded proved by Lyapunov stability theory, and there is no Zeno behavior. Finally, two numerical examples are presented to show the effectiveness of the proposed ET‐SIRL control algorithm.

Composite neural learning based appointed-time safe approach control under full-state constraints

This article investigates composite learning based appointed-time control strategy for spacecraft rendezvous and docking subject to model uncertainties, input saturation, and actuator faults. A novel prescribed performance function (PPF) is designed to constraint all system states and ensure the property of controller insensitive to initial conditions. Combining the artificial potential function (APF) method with transformed error states, a six-degree-of-freedom (6-DOF) coupled relative motion model with full-state and motion constraints is proposed. To obtain higher control accuracy and address the problem of complex calculations in network weight update, a low-complexity composite learning based neural network (NN) weight update law is investigated. The appointed-time controller is designed by utilizing backstepping method, and the dynamic surface technique is used to solve the “dimension explosion” problem. The proposed controller guarantees the stability of the system states in appointed time without the information of target. Three simulation scenarios are carried out to validate the effectiveness and robustness of the proposed controller.

Dynamic event-triggered control for MSVs via composite-learning-based adaptive neural approach

This article investigates the control issue of marine surface vehicles (MSVs) subject to uncertainties and input saturation. To ensure the smooth implementation of backstepping design framework, the non-smooth saturation model is replaced by a smooth function, which is analogous to the hyperbolic tangent function. To exact reconstruct the uncertainties including unknown internal nonlinear dynamic and external disturbance, an idea of separation reconstruction is proposed. That is, the adaptive neural network (NN) is utilized to online reconstruct the unknown internal nonlinear dynamic, and a NN-based disturbance observer is proposed to timely recover the external disturbance plus the reconstruction error of adaptive NN. To enhance the control accuracy, a serial-parallel estimation model (SPEM) is involved to capture the prediction error of MSVs’ velocity. Based on this, a novel adaptive law used to update the NN weight is designed by involving the velocity error, prediction error and estimation value of disturbance. Furthermore, a composite-learning-based adaptive neural control (CLBANC) law is designed. To decrease the physical wear of actuator, a dynamic event-triggered mechanism is established between the control law and actuator. Finally, a dynamic event-triggered (DET) CLBANC (DET-CLBANC) scheme is developed for MSVs. Theoretical analysis indicates that the proposed control scheme is endowed with the ability to ensure the boundedness of all signals in the whole closed-loop system of MSVs. Moreover, the effectiveness of DET-CLBANC scheme is verified by simulation and comparison.

Dual stage ensemble technique for intrusion detection in cloud computing

A capability of cloud-based IDS in identifying complicated and anonymous attacks is rising in the current era. However, unwanted delays hinder the detection rate. A malicious user might utilize vast quantities of computational power. The cloud provides to perform attacks both within and without the cloud. Furthermore, there are major challenges for intrusion detection due to the ease of the cloud and also the continual restructuring and movement of cloud resources. Intruder detection, feature extraction, and data processing are all included in the novel optimization-based Intrusion Detection System (IDS) paradigm that will be presented in this study. Data normalization is used to first pre-process the input data. Then, appropriate feature extraction is carried out, including the extraction of (a) raw features, (b) statistical features, then (c) higher-order statistical features using suggested kurtosis. The detection phase is then applied to the retrieved features. A two-stage ensemble method is suggested for finding intruders in clouds. Random forest (RF), Support Vector Machine (SVM), optimal Neural Network (NN), and RNN make up the suggested ensemble technique. The RF, SVM, and Optimized NN algorithms are directly fed the collected features. The output of these classifiers is then provided to the RNN classifier (i.e.), RF output to RNN1, SVM output to RNN2, and optimized NN output to RNN3. Then, the weighted average of RNN 1, 2, and 3 is considered as the final output. A Self Adaptive Salp Swarm Optimization optimizes the weights of NN for exact detection (SA-SSO). Finally, a test is conducted to confirm the developed model’s superiority.