A complex EKF-RTRL neural network

Coelho, Pedro Henrique Gouvêa

doi:10.1109/ijcnn.2001.939003

Cited by 13 publications

(8 citation statements)

References 10 publications

Supporting

Mentioning

Contrasting

Unclassified

Order By: Relevance

“…Where R and I mean real and imaginary parts respectively, W a is a scalar, W b is a 1-by-13 matrix, C is a scalar and The neural network is trained by a Kalman filter for complex valued signals as described by the first author in [11]. The training procedure was enhanced in this paper by the use of heuristic fine-tuning techniques for the Kalman filtering.…”

Section: A the Equalizer Structurementioning

confidence: 99%

See 1 more Smart Citation

Preliminary studies on parameter aided EKF-CRTRL equalizer training for fast fading channels

Coelho

Neto

2011

The 2011 International Joint Conference on Neural Networks

View full text Add to dashboard Cite

This paper shows an enhanced training for the EKF-RTRL (Extended Kalman Filter -Real Time Recurrent Learning) single neuron Equalizer using heuristic mechanisms on the training algorithms enabling them to make the training process initial conditions set-up more automatic. The method uses a parameter which evolves accordingly in the training period. The equalizer is used for fast fading selective frequency channels using the WSS_US (Wide Sense Stationary -Uncorrelated Scattering) model. The EKF-RTRL is a symbol by symbol neural equalizer. The performance results here presented depicts several scenarios regarding the channel variation speed. The performance considered in this paper is the symbol error rate (SER). I. INTRODUCTIONHIS paper presents a single state neuron equalizer for fast fading channels. Traditional equalizers perform well over fixed channels, but show poor results for time varying channels. Indeed, many real life channels, like mobile channels particularly in fast fading scenarios are timevarying. This motivates the use of dynamic neural networks as equalizer structures rather than using conventional static structures. Neural networks are known to produce quite good nonlinear mapping of the inverse response of the channel [1] which is of paramount importance in the equalization problem. Recurrent neural networks are within the class of neural networks the most adequate for handling equalization in presence of time-varying channels [1]. The study of fast convergence and high performance channel impulse response tracking equalizers has been carried out by communication systems researchers, in order to use them in digital communication applications in which the effects of fast and frequency selective fading strongly impacts the system performance [2,3]. Such is the case of TDMA digital mobile communication systems. The present work reports preliminary results involving such issues and aims to initiate such studies by using initial conditions parameter based strategies using neural equalizers trained by extended Kalman filters. One of the pioneers work on recurrent neural networks applied to equalizers was done by Kechriotis et al. [4] proposing a fully Recurrent Neural Network Equalizer typically with a small number of neurons which outperforms the linear transversal equalizer

show abstract

Section: A the Equalizer Structurementioning

confidence: 99%

“…Details of the Kalman trained recurrent neural network algorithm can be found in [11]. Table I summarizes the recurrent neural network Kalman training for the proposed equalizer including the tuning procedure just described.…”

Section: A the Equalizer Structurementioning

confidence: 99%

Preliminary studies on parameter aided EKF-CRTRL equalizer training for fast fading channels

Coelho

Neto

2011

The 2011 International Joint Conference on Neural Networks

View full text Add to dashboard Cite

show abstract

“…Known disadvantages of gradient-based methods are slow convergence rates and long training symbols necessary for suitable performance of neural networks based devices.. In order to overcome such problems Kalman filter trained neural networks has been considered in the literature [2], [3], [4], [5], [6].…”

Section: Introductionmentioning

confidence: 99%

Complex RTRL Neural Networks Fast Kalman Training

Coelho

Neto

2007

Seventh International Conference on Intelligent Systems Design and Applications (ISDA 2007)

View full text Add to dashboard Cite

This paper presents an Extended FastKalman filter training procedure for complex RTRL neural networks. Fast Kalman training methods use the framework for Extended Kalman filtering techniques which proved to be efficient but quite computationally demanding particularly when a large number of states is involved. In standard Kalman filtering algorithms the number of multiplications is proportional to the square of the number of states while in Fast Kalman algorithms that number is proportional to the number of states. Fast Kalman/Extended Kalman filter complex RTRL training algorithms inherit the convergence capabilities of the standard Extented Kalman filter techniques and are adequate for complex RTRL neural networks training involving a large number of states. Simulations were carried out which indicate the success of the method. Seventh International Conference on Intelligent Systems Design and Applications 0-7695-2976-3/07 $25.00

show abstract

“…The essence of the recursive EKF procedure is that an approximate covariance matrix is generated, which encapsulates second-order information about the training problem considered and the elements of the matrix evolve during the training process. Since Singhal and Wu introduced the EKF training algorithm in [20] in the context of static forward neural networks (FNNs), the EKF has constituted the basis of computationally efficient neural network based training techniques that facilitate the application of FNNs and RNNs in diverse problems such as pattern classification [15], [16], control [21], [22], channel equalization [23], [24], [25], [26], etc.…”

Section: Training Algorithms For Fcrnnsmentioning

confidence: 99%

Recurrent Neural Network Based Narrowband Channel Prediction

Liu

Yang

Hanzo

2006 IEEE 63rd Vehicular Technology Conference

View full text Add to dashboard Cite

Abstract-In this contribution, the application of fully connected recurrent neural networks (FCRNNs) is investigated in the context of narrowband channel prediction. Three different algorithms, namely the real time recurrent learning (RTRL), the global extended Kalman filter (GEKF) and the decoupled extended Kalman filter (DEKF) are used for training the recurrent neural network (RNN) based channel predictor. Our simulation results show that the GEKF and DEKF training schemes have the potential of converging faster than the RTRL training scheme as well as attaining a better MSE performance.

show abstract

A complex EKF-RTRL neural network

Cited by 13 publications

References 10 publications

Preliminary studies on parameter aided EKF-CRTRL equalizer training for fast fading channels

Preliminary studies on parameter aided EKF-CRTRL equalizer training for fast fading channels

Complex RTRL Neural Networks Fast Kalman Training

Recurrent Neural Network Based Narrowband Channel Prediction

Contact Info

Product

Resources

About