arXiv subjects

M. Sami Fadali

Publications and source records attributed to M. Sami Fadali.

4 recordsLinked to original sources

arXiv2018

Training Recurrent Neural Networks as a Constraint Satisfaction Problem

This paper presents a new approach for training artificial neural networks using techniques for solving the constraint satisfaction problem (CSP). The quotient gradient system (QGS) is a trajectory-based method for solving the CSP. This study converts the training set of a neural network into a CSP and uses the QGS to find its solutions. The QGS finds the global minimum of the optimization problem by tracking trajectories of a nonlinear dynamical system and does not stop at a local minimum of the optimization problem. Lyapunov theory is used to prove the asymptotic stability of the solutions with and without the presence of measurement errors. Numerical examples illustrate the effectiveness of the proposed methodology and compare it to a genetic algorithm and error backpropagation.

cs.LG↗

arXiv2018

Anomaly Classification in Distribution Networks Using a Quotient Gradient System

The classification of anomalies or sudden changes in power networks versus normal abrupt changes or switching actions is essential to take appropriate maintenance actions that guarantee the quality of power delivery. This issue has increased in importance and has become more complicated with the proliferation of volatile resources that introduce variability, uncertainty, and intermittency in circuit behavior that can be observed as variations in voltage and current phasors. This makes diagnostics applications more challenging. This paper proposes using quotient gradient system (QGS) to train two-stage partially recurrent neural network to improve anomaly classification rate in power distribution networks using high-fidelity data from micro-phasor measurement units (PMUs). QGS is a systematic approach to finding solutions of constraint satisfaction problems. We transform the PMUs data from the power network into a constraint satisfaction problem and use QGS to train a neural network by solving the resulting optimization problem. Simulation results show that the proposed supervised classification method can reliably distinguish between different anomalies in power distribution networks. Comparison with other neural network classifiers shows that QGS trained networks provide significantly better classification. Sensitivity analysis is performed concerning the number of PMUs, reporting rates, noise level and early versus late data stream fusion frameworks.

eess.SP↗

arXiv2018

Networked Model Predictive Control Using a Wavelet Neural Network

In this study, we use a wavelet neural network with a feedforward component and a model predictive controller for online nonlinear system identification over a communication network. The wavelet neural network (WNN) performs the online identification of the nonlinear system. The model predictive controller (MPC) uses the model to predict the future outputs of the system over an extended prediction horizon and calculates the optimal future inputs by minimizing a controller cost function. The Lyapunov theory is used to prove the stability of the MPC. We apply the methodology to the online identification and control of an unmanned autonomous vehicle. Simulation results show that the MPC with extended prediction horizon can effectively control the system in the presence of fixed or random network delay.

eess.SP↗

arXiv2018

Training Recurrent Neural Networks via Dynamical Trajectory-Based Optimization

This paper introduces a new method to train recurrent neural networks using dynamical trajectory-based optimization. The optimization method utilizes a projected gradient system (PGS) and a quotient gradient system (QGS) to determine the feasible regions of an optimization problem and search the feasible regions for local minima. By exploring the feasible regions, local minima are identified and the local minimum with the lowest cost is chosen as the global minimum of the optimization problem. Lyapunov theory is used to prove the stability of the local minima and their stability in the presence of measurement errors. Numerical examples show that the new approach provides better results than genetic algorithm and error backpropagation (EBP) trained networks.

eess.SP↗