Deep learning adjusts parameters to optimize the model through model training. Training algorithm is the key to model optimization and implementation. Therefore, the improvement of model training algorithm is of great...
详细信息
Deep learning adjusts parameters to optimize the model through model training. Training algorithm is the key to model optimization and implementation. Therefore, the improvement of model training algorithm is of great significance to deep learning. Based on the original gradientalgorithm, the paper proposes a new gradient descent algorithm AdaGrad Restricted by Windows (AdaRW) for Deep learning model training optimization. Aiming at the defects of AdaGrad algorithm, the new algorithm uses the subset of window to limit historical accumulation, so as to slow down the attenuation of learning rate, and improve the speed of model training. The paper constructs OceanTDA9, a Deep learning model of marine target detection for Synthetic Aperture Radar (SAR) data, and adopts the proposed AdaRW algorithm to train the model based on SAR data with a resolution of 10 m in the Bohai Sea. Experiment shows that the accuracy and loss of the algorithm are better than those of AdaGrad and Stochastic gradientdescent (SGD) algorithms, and the standard deviation is better than that of Adam algorithm.
In this article, we present a data-free method to solve differential equation using artificial neural networks (ANN). This method exploits the universal function approximation nature of a neural network to mimic a spe...
详细信息
In this article, we present a data-free method to solve differential equation using artificial neural networks (ANN). This method exploits the universal function approximation nature of a neural network to mimic a specified partial differential equation (PDE) and provide its solution. Specifically, use of simple feed-forward artificial neural network (FF-ANN) is shown to demonstrate the solution of 1-D second-order differential equations without use of any prior data. The article also demonstrates that the similarity in two PDEs is akin to similarity in the optimized weight matrices attained after the solution using FF-ANN. This property is then utilized for a transferred learning to enable faster convergence of a new PDE, based on the prior solution of a similar PDE. The concept is shown while considering a general form of second-order PDE, and considering specific cases of scalar in-homogeneous wave equation form and Poisson Equation form. Error convergence below 10(-6) is shown and the transferred learning process shows typical time acceleration by a factor of 1.5-3 for the considered equations.
Multivariate adaptive regression splines (MARS) has been adopted as a popular surrogate function to model the unknown relationships between input and output variables in complex systems. Searching optimal solutions on...
详细信息
Multivariate adaptive regression splines (MARS) has been adopted as a popular surrogate function to model the unknown relationships between input and output variables in complex systems. Searching optimal solutions on the complicated surrogate response surface of MARS serves as important tasks in various applications. In this paper, we present an efficient and effective approach to find a global optimal value of MARS models that incorporate two-way interaction terms which are products of truncated linear univariate functions (TITL-MARS). Specifically, with a MARS model consisting of linear and quadratic structures, we can reformulate the optimization problem on TITL-MARS into a mixed integer quadratic programming problem (TITL-MARS-OPT), which can be further solved in a more principled way. To illustrate the effectiveness of our proposed approach TITLMARS-OPT, we come up with a genetic algorithm and a gradient descent algorithm to solve the original version of TITL-MARS, and we compared the performance of the proposed approach with the benchmark algorithms. Numerical experiments are conducted on a spectrum of examples with different levels of complexity, including 6 existing models and a real world application in the wind farm optimization. Our proposed approach can find a global optimal successfully and efficiently while the comparison algorithms fail to find a global optimal solution. In the end, Python code for TITL-MARS and TITL-MARSOPT is made available on GitHub( https://***/JuXinglong/TITL-MARS-OPT). (c) 2022 Elsevier Inc. All rights reserved.
In this paper, we put forward numerical algorithms to simulate fractional HIV infected CD4(+)T cell model and fractional Schrodinger equation. First, we put forward numerical algorithm to solve fractional differential...
详细信息
In this paper, we put forward numerical algorithms to simulate fractional HIV infected CD4(+)T cell model and fractional Schrodinger equation. First, we put forward numerical algorithm to solve fractional differential systems with trigonometric neural network. Then trigonometric neural network method is used to simulate the fractional HIV infected CD4(+)T cell model. It is found that the infection equilibrium has the restriction condition for the number of virus particles released by each infected cell. Under this condition, the fractional system has a unique positive equilibrium. Second, we put forward numerical algorithm to solve fractional PDE with trigonometric neural network. Then this numerical method is used to simulate fractional Schrodinger equation, and some important results in fractional quantum mechanics are concluded by the corresponding numerical results.
A learning and perturbation based multi-stage pre-compensation method is proposed. The nonlinear compensation performance is enhanced by introducing amplitude and phase factors in each stage of pre-compensation, which...
详细信息
In this paper, a new method is proposed to find a feasible energy-efficient path between an initial point and goal point on uneven terrain and then to optimally traverse the path. The path is planned by integrating th...
详细信息
In this paper, a new method is proposed to find a feasible energy-efficient path between an initial point and goal point on uneven terrain and then to optimally traverse the path. The path is planned by integrating the geometric features of the uneven terrain and the biped robot dynamics. This integrated information of biped dynamics and associated cost (energy) for moving toward the goal point is used to define the value of a new speed function at each point on the discretized surface of the terrain. The value is stored as a matrix called the dynamic transport cost (DTC). The path is obtained by solving the Eikonal equation numerically by fast marching method (FMM) on an orthogonal grid, by using the information stored in the DTC matrix. One step of walk on uneven terrain is characterized by 10 footstep parameters (FSPs);these FSPs represent the position of swinging foot at the starting and ending time of the walk, orientation, and state (left or right) of support foot. A walking dataset was created for different walking conditions (FSPs), which the biped robot is likely to encounter when it has to walk on the uneven terrain. The corresponding energy optimal hip and foot trajectory parameters (HFTPs) are obtained by optimization using a genetic algorithm (GA). The created walk dataset is generalized by training a feedforward neural network (NN) using the scaled conjugate gradient (SCG) algorithm. The Foot placement planner gives a sequence of foot positions and orientations along the obtained path, which is followed by the biped robot by generating real-time optimal foot and hip trajectories using the learned NN. Simulation results on different types of uneven terrains validate the proposed method.
In this paper,the authors propose a novel smoothing descent type algorithm with extrapolation for solving a class of constrained nonsmooth and nonconvex problems,where the nonconvex term is possibly *** algorithm adop...
详细信息
In this paper,the authors propose a novel smoothing descent type algorithm with extrapolation for solving a class of constrained nonsmooth and nonconvex problems,where the nonconvex term is possibly *** algorithm adopts the proximal gradientalgorithm with extrapolation and a safe-guarding policy to minimize the smoothed objective function for better practical and theoretical ***,the algorithm uses a easily checking rule to update the smoothing parameter to ensure that any accumulation point of the generated sequence is an(afne-scaled)Clarke stationary point of the original nonsmooth and nonconvex *** experimental results indicate the effectiveness of the proposed algorithm.
Benefiting from area and power efficiency, memristors enable the development of neural network analog-to-digital converter (ADC) to break through the limitations of conventional ADCs. Although some memristive ADC (mAD...
详细信息
Benefiting from area and power efficiency, memristors enable the development of neural network analog-to-digital converter (ADC) to break through the limitations of conventional ADCs. Although some memristive ADC (mADC) architectures have been proposed recently, the current research is still at an early stage, which is mainly on the simulation and requires numerous target labels to train the synapse weights. In this paper, we propose a pipelined Hopfield neural network mADC architecture and experimentally demonstrate that such mADC has the capability of self-adaptive weight tuning. The proposed training algorithm is an unsupervised method originated from the random weight change (RWC) algorithm, which is modified to reduce the complexity of error feedback circuit to make it more hardware friendly. The synapse matrix could be adapted to the 1T1R crossbar array. For an 8-bit two-stage pipelined mADC, the proposed architecture in the simulation could achieve 7.69 fJ/conv FOM, 7.90 ENOB, 0.1 LSB INL, and 0.1 LSB DNL. And the experimental performance only achieves 1.56 pJ/conv FOM, 7.59 ENOB, 0.21 LSB INL, and 0.29 LSB DNL, which is mainly limited by the comparator's switching time.
In this paper, we focus on the path following control of a fixed-wing UAV and investigate how to avoid pop-up threats by onboard optimization. It contains two subproblems: how to bypass threat areas in time and how to...
详细信息
In this paper, we focus on the path following control of a fixed-wing UAV and investigate how to avoid pop-up threats by onboard optimization. It contains two subproblems: how to bypass threat areas in time and how to minimize deviations from the desired path. More importantly, the calculation must be completed in real-time. To address this, we design an accelerated model predictive controller with two stages. First, we incorporate multiple control objectives into the prediction model in a weighted manner. Then, by introducing the rectified linear unit penalties, it is transformed into an unconstrained problem, which is solved by gradient-based optimizers. Six state-of-the-art optimizers are compared in simulations. The results prove that the proposed control method can well deal with multiple static or moving threats, is robust to wind disturbance, and has high real-time performance.
Single-phase inverters are becoming increasingly important and popular because of the rise of distributed renewable energy. The mainstream single-phase inverter control methods include proportional resonance (PR) cont...
详细信息
ISBN:
(纸本)9798350339345
Single-phase inverters are becoming increasingly important and popular because of the rise of distributed renewable energy. The mainstream single-phase inverter control methods include proportional resonance (PR) control and proportional integral (PI) control, which need to adjust parameters according to manual experience. In this article, an echo state network (ESN) controller is designed to control L-filter single phase inverter. Firstly, the relevant factors affecting the modulation wave required by PWM are selected by analyzing the working principle of the inverter circuit. Secondly, the nonlinear relationship between them and the modulated wave is established by ESN and gradientdescent (GD) algorithm to achieve high-precision approximation of the modulated wave, which avoids the problem of artificial parameter adjustment. Finally, the simulation is carried out on MATLAB/Simulink platform. The experiments show that the ESN controller can not only accurately approximate the modulation signal required to realize the inverter control, but also compared with PR controller and PI controller, the ESN controller can realize stable grid-connection without adjustment in the initial state, and has good dynamic performance.
暂无评论