NEURAL NETWORKS FOR TRANSMISSION OVER NONLINEAR MIMO CHANNELS

Size: px
Start display at page:

Download "NEURAL NETWORKS FOR TRANSMISSION OVER NONLINEAR MIMO CHANNELS"

Transcription

1 NEURAL NETWORKS FOR TRANSISSION OVER NONLINEAR IO CHANNELS by AL UKHTAR AL-HINAI A thesis submitted to the Department of Electrical and Computer Engineering in conformity with the requirements for the degree of aster of Science (Engineering) Queen s University Kingston, Ontario, Canada August, 2007 Copyright AL uhtar AL-Hinai, 2007

2 To y Father who taught me the virtue of modesty To my other who taught me the meaning of ambition

3 Abstract ultiple-input ultiple-output (IO) systems have gained an enormous amount of attention as one of the most promising research areas in wireless communications. However, while IO systems have been extensively explored over the past decade, few schemes acnowledge the nonlinearity caused by the use of high power amplifiers (HPAs) in the communication chain. When HPAs operate near their saturation points, nonlinear distortions are introduced in the transmitted signals, and the resulting IO channel will be nonlinear. The nonlinear distortion is further exacerbated by the fading caused by the propagation channel. The goal of this thesis is: ) to use neural networs (NNs) to model and identify nonlinear IO channels; and 2) to employ the proposed NN model in designing efficient detection techniques for these types of IO channels. In the first part of the thesis, we follow a previous wor on modeling and identification of nonlinear IO channels, where it has been shown that a proposed bloc-oriented NN scheme allows not only good identification of the overall IO input-output transfer function but also good characterization of each component of the system. The proposed scheme employs an ordinary gradient descent based algorithm to update the NN weights during the learning process and it assumes only real-valued inputs. In this thesis, natural gradient (NG) descent is used for training the NN. oreover, we derive an improved variation of the previously proposed NN scheme to avoid the input type restriction and allow for complex modulated inputs as well. We also investigate the scheme tracing capabilities of time-varying nonlinear IO channels. Simulation results show that NG i

4 descent learning significantly outperforms the ordinary gradient descent in terms of convergence speed, mean squared error (SE) performance, and nonlinearity approximation. oreover, the NG descent based NN provides better tracing capabilities than the previously proposed NN. The second part of the thesis focuses on signal detection. We propose a receiver that employs the neural networ channel estimator (NNCE) proposed in part one, and uses the Zero-Forcing Vertical Bell Laboratories Layered Space-Time (ZF V-BLAST) detection algorithm to retrieve the transmitted signals. Computer simulations show that in slow time-varying environments the performance of our receiver is close to the ideal V- BLAST receiver in which the channel is perfectly nown. We also present a NN based linearization technique for HPAs, which taes advantage of the channel information provided by the NNCE. Such linearization technique can be used for adaptive data predistortion at the transmitter side or adaptive nonlinear equalization at the receiver side. Simulation results show that, when higher modulation schemes (>6-QA) are used, the nonlinear distortion caused by the use of HPAs is greatly minimized by our proposed NN predistorter and the performance of the communication system is significantly improved. ii

5 Acnowledgements I would lie to than my supervisor, Dr. ohamed Ibnahla for his help, guidance, and encouragement. I would also lie to than Dr. Aboelmagd Noureldin, Dr. Saeed Gazor, and Dr. Il-in Kim for their valuable comments during my defense. I extend my gratitude to all the members of the Satellite and obile Communication Lab for their friendship and help. I would especially lie to than Ali Alamdar for taing the time to proof-read my thesis. A special than you goes to the inistry of Higher Education in the Sultanate of Oman for their financial support. I am infinitely grateful for my friend Tricia Armson for her long time encouragement and support. Without the smile of the moon I would be lost in the void. y deepest gratitude goes to my parents. To my om who endured the patience of time waiting for this moment that I may return home. To my Dad who always believed in the importance of education. Than you for believing in me and giving me the motivation to continue even through the trying times. Dad you made it possible to have the education to buy all the shoes I need. iii

6 Contents Abstract...i Acnowledgements...iii Contents...iv List of Figures...vi List of Tables...xi List of Tables...xi Summery of Abbreviations and Symbols...xii Abbreviations... xii Symbols... xiii Chapter... Introduction.... otivation Bacground and Literature Review Neural Networ V-BLAST Detection Algorithm HPA Linearization Thesis Contribution Thesis Outline... 6 Chapter Nonlinear IO Channel odel and Neural Networ Scheme Nonlinear IO Channel odel Neural Networ Scheme Learning Algorithms Nonlinear Adaptive Algorithms Linear Adaptive Algorithms... 5 Chapter Neural Networ odeling and Identification of Nonlinear IO Channels Study Case odeling and Identification of Nonlinear IO Channels L IO Systems ( > 2, L > 2) Tracing of Time-Varying Nonlinear IO Channels iv

7 Chapter Improved Neural Networ odeling and Identification Scheme Channel odel Identification Structure Learning Algorithm Simulation results Chapter Applications Channel odel A Neural Networ Based V-BLAST Receiver The NNCE V-BLAST Detection Algorithm LS-NN Equalizer NN predistorter Simulation Chapter Conclusion and Future Wor Conclusion Suggestions for Future Wor... 0 Reference...02 v

8 List of Figures Fig. 2.: Nonlinear IO channel... 8 Fig. 2.2: NN identification structure... 0 Fig. 2.3: Adaptive system diagram... 2 Fig. 3.: SE vs. µ: comparison between BP and NG based algorithms ( = L = 2, N = 5, SNR = 60 db, Zero-mean white Gaussian input)... 9 Fig. 3.2: Smoothed SE curve (LS-BP, = L = 2, µ = 0. 09, N = 5, SNR = 60 db, Zeromean white Gaussian input)... 2 Fig. 3.3: Evolution of the normalized weights (LS-BP, = L = 2, µ = 0. 09, N = 5, SNR = 60 db, Zero-mean white Gaussian input)... 2 Fig. 3.4: HPA nonlinearity g (x) and normalized NN (x) (LS-BP, = L = 2, µ = 0. 09, N = 5, SNR = 60 db, Zero-mean white Gaussian input) Fig. 3.5: HPA nonlinearity g 2 (x) and normalized NN 2 (x) (LS-BP, = L = 2, µ = 0. 09, N = 5, SNR = 60 db, Zero-mean white Gaussian input) Fig. 3.6: Smoothed SE curve (RLS-BP, = L = 2, µ = 0. 09, N = 5, SNR = 60 db, Zeromean white Gaussian input) Fig 3.7: Evolution of the normalized weights (RLS-BP, = L = 2, µ = 0. 09, N = 5, SNR = 60 db, Zero-mean white Gaussian input) Fig. 3.8: HPA nonlinearity g (x) and normalized NN (x) (RLS-BP, = L = 2, µ = 0. 09, N = 5, SNR = 60 db, Zero-mean white Gaussian input) Fig. 3.9: HPA nonlinearity g 2 (x) and normalized NN 2 (x) (RLS-BP, = L = 2, µ = 0. 09, N = 5, SNR = 60 db, Zero-mean white Gaussian input) Fig. 3.0: Smoothed SE curve (LS-NGBP, = L = 2, µ = , N = 5, SNR = 60 db, Zero-mean white Gaussian input) Fig. 3.: Evolution of the normalized weights (LS-NGBP, = L = 2, µ = , N = 5, SNR = 60 db, Zero-mean white Gaussian input) Fig. 3.2: HPA nonlinearity g (x) and normalized NN (x) (LS-NGBP, = L = 2, µ = , N = 5, SNR = 60 db, Zero-mean white Gaussian input) Fig. 3.3: HPA nonlinearity g 2 (x) and normalized NN 2 (x) (LS-NGBP, = L = 2, µ = , N = 5, SNR = 60 db, Zero-mean white Gaussian input) Fig. 3.4: Smoothed SE curve (RLS-NGBP, = L = 2, µ = , N = 5, SNR = 60 db, Zero-mean white Gaussian input) Fig. 3.5: Evolution of the normalized weights (RLS-NGBP, = L = 2, µ = , N = 5, SNR = 60 db, Zero-mean white Gaussian input) Fig. 3.6: HPA nonlinearity g (x) and normalized NN (x) (RLS-NGBP, = L = 2, µ = , N = 5, SNR = 60 db, Zero-mean white Gaussian input) Fig. 3.7: HPA nonlinearity g 2 (x) and normalized NN 2 (x) (RLS-NGBP, = L = 2, µ = , N = 5, SNR = 60 db, Zero-mean white Gaussian input) Fig. 3.8: Smoothed SE curves: comparison between BP and NG based algorithms ( = L = 2, N = 5, SNR = 60 db, Zero-mean white Gaussian input) Fig. 3.9: Smoothed SE curves for N = 3, 5 and 7 (LS-BP, = L = 2, µ = 0. 09, SNR = 60 db, Zero-mean white Gaussian input)... 3 vi

9 Fig. 3.20: Smoothed SE curves for N = 3, 5 and 7 (LS-NGBP, = L = 2, µ = , SNR = 60 db, Zero-mean white Gaussian input)... 3 Fig. 3.2: SE vs. µ : comparison between BP and NG based algorithms ( = L = 3, N = 5, SNR = 60 db, Zero-mean white Gaussian input) Fig. 3.22: Smoothed SE curve (LS-BP, = L = 3, µ = 0. 07, N = 5, SNR = 60 db, Zeromean white Gaussian input) Fig. 3.23: Evolution of the normalized weights (LS-BP, = L = 3, µ = 0. 07, N = 5, SNR = 60 db, Zero-mean white Gaussian input) Fig. 3.24: HPA nonlinearity g (x) and normalized NN (x) (LS-BP, = L = 3, µ = 0. 07, N = 5, SNR = 60 db, Zero-mean white Gaussian input) Fig. 3.25: HPA nonlinearity g 2 (x) and normalized NN 2 (x) (LS-BP, = L = 3, µ = 0. 07, N = 5, SNR = 60 db, Zero-mean white Gaussian input) Fig. 3.26: HPA nonlinearity g 3 (x) and normalized NN 3 (x) (LS-BP, = L = 3, µ = 0. 07, N = 5, SNR = 60 db, Zero-mean white Gaussian input) Fig. 3.27: Smoothed SE curve (RLS-BP, = L = 3, µ = 0. 09, N = 5, SNR = 60 db, Zeromean white Gaussian input) Fig. 3.28: Evolution of the normalized weights (RLS-BP, = L = 3, µ = 0. 09, N = 5, SNR = 60 db, Zero-mean white Gaussian input) Fig. 3.29: HPA nonlinearity g (x) and normalized NN (x) (RLS-BP, = L = 3, µ = 0. 09, N = 5, SNR = 60 db, Zero-mean white Gaussian input) Fig. 3.30: HPA nonlinearity g 2 (x) and normalized NN 2 (x) (RLS-BP, = L = 3, µ = 0. 09, N = 5, SNR = 60 db, Zero-mean white Gaussian input) Fig. 3.3: HPA nonlinearity g 3 (x) and normalized NN 3 (x) (RLS-BP, = L = 3, µ = 0. 09, N = 5, SNR = 60 db, Zero-mean white Gaussian input) Fig. 3.32: Smoothed SE curve (LS-NGBP, = L = 3, µ = 0. 0, N = 5, SNR = 60 db, Zeromean white Gaussian input) Fig. 3.33: Evolution of the normalized weights (LS-NGBP, = L = 3, µ = 0. 0, N = 5, SNR = 60 db, Zero-mean white Gaussian input) Fig. 3.34: HPA nonlinearity g (x) and normalized NN (x) (LS-NGBP, = L = 3, µ = 0. 0, N = 5, SNR = 60 db, Zero-mean white Gaussian input) Fig. 3.35: HPA nonlinearity g 2 (x) and normalized NN 2 (x) (LS-NGBP, = L = 3, µ = 0. 0, N = 5, SNR = 60 db, Zero-mean white Gaussian input) Fig. 3.36: HPA nonlinearity g 3 (x) and normalized NN 3 (x) (LS-NGBP, = L = 3, µ = 0. 0, N = 5, SNR = 60 db, Zero-mean white Gaussian input)... 4 Fig. 3.37: Smoothed SE curve (RLS-NGBP, = L = 3, µ = , N = 5, SNR = 60 db, Zero-mean white Gaussian input)... 4 Fig. 3.38: Evolution of the normalized weights (RLS-NGBP, = L = 3, µ = , N = 5, SNR = 60 db, Zero-mean white Gaussian input) Fig. 3.39: HPA nonlinearity g (x) and normalized NN (x) (RLS-NGBP, = L = 3, µ = , N = 5, SNR = 60 db, Zero-mean white Gaussian input) Fig. 3.40: HPA nonlinearity g 2 (x) and normalized NN 2 (x) (RLS-NGBP, = L = 3, µ = , N = 5, SNR = 60dB, Zero-mean white Gaussian input) Fig. 3.4: HPA nonlinearity g 3 (x) and normalized NN 3 (x) (RLS-NGBP, = L = 3, µ = , N = 5, SNR = 60 db, Zero-mean white Gaussian input) vii

10 Fig. 3.42: Smoothed SE curves: comparison between BP and NG based algorithms ( = L = 3, N = 5, SNR = 60 db, Zero-mean white Gaussian input) Fig. 3.43: Time varying case - SE vs. µ: comparison between BP and NG based algorithms ( = L = 2, N = 5, SNR = 60 db, Zero-mean white Gaussian input) Fig. 3.44: Time varying case - Smoothed SE curves: comparison between LS-BP and RLS- BP (f d = 0.000, = L = 2, N = 5, SNR = 60 db, Zero-mean white Gaussian input) Fig. 3.45: Time varying case: comparison between LS-BP and RLS-BP (f d = 0.000, = L = 2, N = 5, SNR = 60 db, Zero-mean white Gaussian input) Fig. 3.46: Time varying case - HPA nonlinearity g (x) and normalized NN (x): comparison between LS-BP and RLS-BP (f d = 0.000, = L = 2, N = 5, SNR = 60 db, Zero-mean white Gaussian input) Fig. 3.47: Time varying case - HPA nonlinearity g 2 (x) and normalized NN 2 (x): comparison between LS-BP and RLS-BP (f d = 0.000, = L = 2, N = 5, SNR = 60 db, Zero-mean white Gaussian input) Fig. 3.48: Time varying case - Smoothed SE curves: comparison between LS-NGBP and RLS-NGBP (f d = 0.000, = L = 2, N = 5, SNR = 60 db, Zero-mean white Gaussian input) Fig. 3.49: Time varying case: comparison between LS-NGBP and RLS-NGBP (f d = 0.000, = L = 2, N = 5, SNR = 60 db, Zero-mean white Gaussian input) Fig. 3.50: Time varying case - HPA nonlinearity g (x) and normalized NN (x): comparison between LS-NGBP and RLS-NGBP (f d = 0.000, = L = 2, N = 5, SNR = 60 db, Zero-mean white Gaussian input) Fig. 3.5: Time varying case - HPA nonlinearity g 2 (x) and normalized NN 2 (x): comparison between LS-NGBP and RLS-NGBP (f d = 0.000, = L = 2, N = 5, SNR = 60 db, Zero-mean white Gaussian input)... 5 Fig. 3.52: Time varying - case Smoothed SE curves: comparison between BP and NG based algorithms (f d = 0.000, = L = 2, N = 5, SNR = 60 db, Zero-mean white Gaussian input) Fig. 3.53: Time varying - case Smoothed SE curves: RLS-NGBP performance comparison for different forgetting facotrs (f d = 0.000, = L = 2, N = 5, SNR = 60 db, Zero-mean white Gaussian input) Fig. 3.54: RBF identification structure Fig. 3.55: SE vs. µ: RBF identification structure ( = L = 2, SNR = 60 db, Zero-mean white Gaussian input) Fig. 3.56: Smoothed SE curves: comparison between NN scheme and RBF structure (f d = 0.000, = L = 2, N = 5, SNR = 60 db, Zero-mean white Gaussian input) Fig. 3.57: RBF identification of the time-varying channel: h (n) and normalized w (n) (f d = 0.000, = L = 2, N = 5, SNR = 60 db, Zero-mean white Gaussian input) Fig. 3.58: RBF HPA nonlinearity approximation: g (x) and normalized NN (x): (f d = 0.000, = L = 2, N = 5, SNR = 60 db, Zero-mean white Gaussian input) Fig. 3.59: RBF HPA nonlinearity approximation: g 2 (x) and normalized NN 2 (x): (f d = 0.000, = L = 2, N = 5, SNR = 60 db, Zero-mean white Gaussian input) Fig. 4.: Improved NN identification structure Fig. 4.2: Improved NN model - SE vs. µ : comparison between LS-BP and LS-NGBP algorithms ( = L = 2, N = 5, SNR = 60 db, 6-QA) Fig. 4.3: Improved NN model - Smoothed SE curves: comparison between LS-BP and LS- NGBP algorithms ( = L = 2, N = 5, SNR = 60 db, 6-QA) Fig. 4.4: Improved NN model - Evolution of the normalized weights (LS-BP, = L = 2, µ = 0. 05, N = 5, SNR = 60 db, 6-QA) viii

11 Fig. 4.5: Improved NN model - Evolution of the normalized weights (LS-NGBP, = L = 2, µ = 0.003, N = 5, SNR = 60 db, 6-QA) Fig. 4.6: Improved NN model - HPA A/A characteristic of g (x): true curve and normalized NN models, (o) and (*) represent the three 6-QA amplitudes and their corresponding outputs for LS-NGBP and LS-BP ( = L = 2, N = 5, SNR = 60 db ) Fig. 4.7: Improved NN model - HPA A/A characteristic of g 2 (x): true curve and normalized NN models, (o) and (*) represent the three 6-QA amplitudes and their corresponding outputs for LS-NGBP and LS-BP ( = L = 2, N = 5, SNR = 60 db ) Fig. 4.8: Improved NN model - HPA A/P characteristic of g (x): true curve and normalized NN models, (o) and (*) represent the three 6-QA amplitudes and their corresponding outputs for LS-NGBP and LS-BP ( = L = 2, N = 5, SNR = 60 db ) Fig. 4.9: Improved NN model - HPA A/P characteristic of g 2 (x): true curve and normalized NN models, (o) and (*) represent the three 6-QA amplitudes and their corresponding outputs for LS-NGBP and LS-BP ( = L = 2, N = 5, SNR = 60 db ) Fig. 4.0: Improved NN model Time varying case: SE vs. µ : comparison between LS-BP and LS-NGBP algorithms ( = L = 2, N = 5, SNR = 60 db, 6-QA) Fig 4.: Improved NN model - Time varying case: Smoothed SE curve: comparison between LS-BP and LS-NGBP algorithms (f d = 0.000, = L = 2, N = 5, SNR = 60 db, 6- QA) Fig 4.2 Improved NN model - Identification of the time-varying channel: comparison between LS-BP and LS-NGBP algorithms (f d = 0.000, = L = 2, N = 5, SNR = 60 db, 6- QA) Fig. 4.3: Improved NN model - Time varying case: HPA A/A characteristic of g (x): (o) and (*) represent the three 6-QA amplitudes and their corresponding outputs for LS-NGBP and LS-BP (f d = 0.000, = L = 2, N = 5, SNR = 60 db )... 7 Fig. 4.4: Improved NN model - Time varying case: HPA A/A characteristic of g 2 (x): (o) and (*) represent the three 6-QA amplitudes and their corresponding outputs for LS-NGBP and LS-BP (f d = 0.000, = L = 2, N = 5, SNR = 60 db )... 7 Fig. 4.5: Improved NN model - Time varying case: HPA A/P characteristic of g (x): (o) and (*) represent the three 6-QA amplitudes and their corresponding outputs for LS-NGBP and LS-BP (f d = 0.000, = L = 2, N = 5, SNR = 60 db ) Fig. 4.6: Improved NN model - Time varying case: HPA A/P characteristic of g 2 (x): (o) and (*) represent the three 6-QA amplitudes and their corresponding outputs for LS-NGBP and LS-BP (f d = 0.000, = L = 2, N = 5, SNR = 60 db ) Fig 4.7: Improved NN model - Time varying case: Smoothed SE curve: comparison between LS-BP and LS-NGBP algorithms (f d = 0.000, = L = 2, N = 5, SNR = 60 db, 64- QA) Fig 4.8: Improved NN model - Identification of the time-varying channel: comparison between LS-BP and LS-NGBP algorithms (f d = 0.000, = L = 2, N = 5, SNR = 60 db, 64- QA) Fig. 4.9: Improved NN model - Time varying case: HPA A/A characteristic of g (x): (o) and (*) represent the three 64-QA amplitudes and their corresponding outputs for LS-NGBP and LS-BP (f d = 0.000, = L = 2, N = 5, SNR = 60 db ) Fig. 4.20: Improved NN model - Time varying case: HPA A/A characteristic of g 2 (x): (o) and (*) represent the three 64-QA amplitudes and their corresponding outputs for LS-NGBP and LS-BP (f d = 0.000, = L = 2, N = 5, SNR = 60 db ) Fig. 4.2: Improved NN model - Time varying case: HPA A/P characteristic of g (x): (o) and (*) represent the three 64-QA amplitudes and their corresponding outputs for LS-NGBP and LS-BP (f d = 0.000, = L = 2, N = 5, SNR = 60 db ) ix

12 Fig. 4.22: Improved NN model - Time varying case: HPA A/P characteristic of g 2 (x): (o) and (*) represent the three 64-QA amplitudes and their corresponding outputs for LS-NGBP and LS-BP (f d = 0.000, = L = 2, N = 5, SNR = 60 db ) Fig. 4.23: Improved NN model Smoothed SE curves: comparison between LS-BP and LS- NGBP for 6-QA and 64-QA Fig. 4.24: Improved NN model Smoothed SE curves: performance comparison between f d = and f d = 0.00 for 64-QA Fig. 5.: Simplified nonlinear IO communication system Fig. 5.2: Neural networ based V-BLAST receiver detection part Fig. 5.3: LS-NN equalizer Fig. 5.4: NN predistorter Fig. 5.5: Training of the nonlinearity inverter Fig. 5.6: g (x) - HPA A/A: Amplitude linearization Fig. 5.7: g 2 (x) - HPA A/A: Amplitude linearization Fig. 5.8: g (x) - HPA A/P: Phase shift cancellation Fig. 5.9: g 2 (x) - HPA A/P: Phase shift cancellation Fig. 5.0: Rectangular 6-QA constellation Fig. 5.: g (x) Output distorted 6-QA constellation Fig. 5.2: g 2 (x) Output distorted 6-QA constellation Fig. 5.3: g (x) Output 6-QA constellation after predistortion Fig. 5.4: g 2 (x) Output 6-QA constellation after predistortion... 9 Fig. 5.5: Rectangular 64-QA constellation Fig. 5.6: g (x) output distorted 64-QA constellation Fig. 5.7: g 2 (x) output distorted 64-QA constellation Fig. 5.8: g (x) Output 64-QA constellation after predistortion Fig. 5.9: g 2 (x) Output 64-QA constellation after predistortion Fig. 5.20: Frame structure Fig. 5.2: BER vs. SNR (normalized f d = 0-5, 6-QA) Fig. 5.22: BER vs. SNR (normalized f d = 0.000, 6-QA) Fig. 5.23: BER vs. SNR (normalized f d = 0-5, 32-QA) Fig. 5.24: BER vs. SNR (normalized f d = 0.000, 32-QA) Fig. 5.25: BER vs. SNR (normalized f d = 0-5, 64-QA) Fig. 5.26: BER vs. SNR (normalized f d = 0.000, 64-QA) x

13 List of Tables Table 3.: SE and scaling factors for the different algorithms Table 5.: SNR (db) needed to reach 0-4 BER for the different proposed detection techniques, (- ) indicates that the 0-4 cannot be achieved, (f d = 0-5 ) Table 5.2: SNR (db) needed to reach 0-4 BER for the different proposed detection techniques, (- ) indicates that the 0-4 BER cannot be achieved, (f d = 0.000) xi

14 Summery of Abbreviations and Symbols Abbreviations A/A A/P AWGN BER BP CSI db DD FI FIR HPA LS IO SE SE NG NGBP NN NNCE QA RLS Amplitude to Amplitude Conversions Amplitude to Phase Conversions Additive White Gaussian Noise Bit Error Rate Bac Propagation Channel State Information Decibel Decision Directed Fisher Information atrix Finite Impulse Response High Power Amplifier Least ean Squares ultiple-input ultiple-output inimum ean-squared Error ean Squared Error Natural Gradient Natural Gradient Bac-Propagation Neural Networ Neural Networ Channel Estimator Quadrature Amplitude odulation Recursive Least Squares xii

15 SNR TS ZF ZF-VBLAST Signal to Noise Ratio Training Sequence Zero-Forcing Zero-Forcing Vertical Bell Laboratories Layered Space- Time Symbols A (r ) A/A response of the HPA th th a Input weight of the i neuron of the bloc i th th b Output weight of the i neuron of the bloc i th th c Bias weight of the i neuron of the bloc i e j The j th Error e j The j th Error: imaginary part I e j The j th Error: real part R f (). Neural networ activation function fd Normalized Doppler frequency G Inverse of the Fisher information matrix (). g System s th memory-less nonlinearity H Linear combiner matrix H estimate Estimation of the channel matrix H + ( H estimate ) oore-penrose pseudo-inverse of H estimate ( estimate ) i H The th column of H xiii

16 ( estimate ) i H atrix obtained by zeroing the columns,, 2 of H estimate ( H ) + estimate Pseudo-inverse of ( H estimate ) i i J Cost function L N N j NN G ( r ) Number of antennas at the transmitter side Number of antennas at the receiver side Number of neurons in each neural networ bloc The j th white noise The th neural networ bloc gain output th NN (.) The neural networ bloc NN P ( r ) The th neural networ bloc phase output P (r ) A/P response of the HPA Q (). A quantizer to the nearest constellation point r s j Amplitude of the th input signal Neural networ s j th output x System s th input W Neural networ weight matrix ( W i ) j The j th row of Wi y j System s j th output α, α, β, β Parameters of the th HPA G P G P θ System parameter vector for th neural networ bloc µ Learning rate xiv

17 θ Ordinary gradient with respect to matrix θ λ γ RLS forgetting factor Scaling factor φ Phase of the th input signal xv

18 Chapter Introduction Nonlinear IO systems are being increasingly applied in several engineering fields, including satellite communications [23, 35], system control [42, 44], and control of Underwater Vehicles [4, 9, 25].. otivation The quest for higher data rates in wireless communications is significantly increasing at rapid paces. To achieve this goal, IO systems have been well looed at and investigated as a mean of increasing channel capacity []. The IO concept can offer significantly high data rate and capacity, with no additional bandwidth, when the channel exhibits rich scattering and its variation can be accurately traced [2]. In order to achieve higher date rate and fulfill the power requirement at the same time, IO communication systems may be equipped with high power amplifiers (HPAs). However, HPAs introduce nonlinearity to the system when operating near their nonlinear saturation regions. This nonlinearity imposes a major restriction on the modulation scheme the communication system can use. For instance, ulti-level modulation schemes, such as -ary (>4) Quadrature Amplitude odulation (QA), are sensitive to nonlinear distortion due to their large envelope fluctuation; hence, the system is restricted to simple modulation schemes such as Binary Phase Shift Keying (BPSK) [3, 5, 23, 26, 35].

19 In addition to the nonlinear behavior of HPAs there is the further challenge that lies in the variations of the wireless fading channel. The parameters of the time-varying channel are not directly observed and thus should be accurately traced. Therefore, to reach maximum throughput; nowledge of accurate and timely channel state information (CSI) is extremely important in wireless IO communication systems. For example, the receiver needs to now the channel for accurate detection and data-demodulation [27]. However, modeling the channel time-varying parameters and the HPA nonlinearities is a highly challenging tas, especially when both the nonlinearity and fading parameters are unnown. The goal of this thesis is to improve a previously proposed neural networ channel estimator (NNCE) for nonlinear IO channels which are composed of inputs, memory-less nonlinearities, a linear combiner, and L outputs. The improved NNCE is expected to adaptively identify the overall IO input-output transfer function and characterize each component of the channel. The NNCE is then applied to the design of efficient detection techniques for the type of nonlinear IO channels under study. These detection techniques include: a Zero-Forcing Vertical Bell Laboratories Layered Space-Time (ZF V-BLAST) receiver, an LS-NN equalizer, and a NN based predistorter. 2. Bacground and Literature Review Power efficiency represents the ability of a system to reliably transmit information at the lowest practical power level; while spectral efficiency demonstrates the ability of a system (e.g. modulation schemes) to accommodate data within an allocated bandwidth 2

20 [20]. High power efficiency can be achieved through the employment of HPAs within the communication chain. However, HPAs cause nonlinear distortion to the transmitted signals which becomes significant when high-level modulations schemes are used and thus restricting the system to simple and spectrally inefficient modulation schemes. High level modulation schemes are particularly susceptible to the distortion caused by HPAs because of their large envelope fluctuation. Therefore, the tas of achieving power and spectral efficient communication systems remains highly challenging particularly when considering the nonlinearity problem caused by the use of HPAs. One way to address this challenge is to accurately identify the channel parameters..2. Neural Networ Neural networs have been widely applied for modeling and identification of nonlinear IO channels. This is mainly due to their universal approximation, learning and adaptation abilities, which mae them powerful modeling tools for nonlinear dynamic systems [4, 5, 9, 2, 22, 28]. When adapting a bloc-oriented structure, the NN model is capable of characterizing not only the overall input-output transfer function but also individual components of the channel. The bloc oriented approach copies the general physical structure of the channel to be identified [7]. In other words, for each component in the channel there is a corresponding component in the bloc-oriented NN model. Several different ways exist for training NNs. One such way is supervised learning, where a set of training data is presented to the neural networ. The training set consists of pairs of input and desired output. The networ parameters are then continuously adjusted 3

21 under the affect of the training data and the error signal; the error signal is defined as the difference between the actual output of the networ and desired output [4, 7]. One of the well-nown adaptive learning algorithms used for training NNs is the bacpropagation algorithm. BP, however, has two major limitations: first, it has slow convergence speed and second it can get trapped in a local minima resulting in suboptimal approximation. Natural gradient (NG) learning, on the other hand, has been shown to have better convergence speed than classical BP because it taes into account the geometry of the coordinate system in which the NN parameters evolve. This maes NG learning better in escaping the plateau regions, which are typical of the classical BP [8, 20]..2.2 V-BLAST Detection Algorithm Several detection algorithms have been proposed in order to exploit the high spectral capacity offered by IO channels. One such algorithm is the V-BLAST algorithm, where the data streams are independently encoded and transmitted from each transmit antenna simultaneously, and detected at the receiver by a nulling and successive cancellation scheme [0, 40]. Thus, the specifics of the detection process depend on the method used to perform the nulling cancellation, the most common choices being zeroforcing (ZF) and minimum mean-squared error (SE). The detection algorithm described in this thesis will be with respect to the ZF criterion due to its simplicity. 4

22 .2.3 HPA Linearization Two approaches have been proposed to overcome the distortion caused by HPAs: predistortion [3, 32, 39] and equalization [6, 29, 38]. Predistortion is performed at the transmitter side prior to the amplification stage as it aims at pre-canceling the nonlinear effects via modeling the inverse of the amplifier characteristics. On the other hand, equalization is performed at the receiver side by post-canceling the amplifier nonlinear distortions..3 Thesis Contribution Our first contribution is to examine a previously proposed NN scheme for nonlinear IO channel modeling and identification. We investigate the performance of the NN scheme under natural gradient descent learning and compare it to that obtained under ordinary gradient descent learning. Indeed, NG descent learning significantly outperforms ordinary gradient descent learning in terms of convergence speed, mean squared error (SE) performance, and nonlinearity approximation. As a result, NG descent based NN scheme allows better characterization of the different parts of the unnown IO channel as well as excellent adaptive identification of the overall IO input-output transfer function. In addition, the NN scheme seems to exhibit better tracing capabilities under the NG descent learning. Our second contribution is to modify the already existing proposed NN scheme to account for complex-valued signals. The improved scheme will be able to model the amplitude distortion as well as the phase shift caused by the HPAs, which was not possible with the previous NN scheme. 5

23 Our third contribution is to propose an adaptive NN based V-BLAST receiver for nonlinear IO systems. The receiver is composed of the proposed neural networ channel estimator and a ZF V-BLAST detection algorithm. The ZF V-BLAST algorithm has been modified so that the channel nonlinearity is taen into consideration. Computer simulations show that in slow time-varying environments the performance of our receiver is close to the ideal V-BLAST receiver in which the channel is perfectly nown. Our forth contribution is to present a NN based linearization technique for HPAs used in IO communication systems under study. We apply our proposed NN scheme to approximate the inverse of the nonlinearities caused by HPAs, which can then be used for adaptive data predistortion at the transmitter side or adaptive nonlinear equalization at the receiver side. Simulation results show that, when higher modulation schemes (> 6- QA) are used, the nonlinear distortion caused by the use of HPAs is greatly minimized by our proposed NN predistorter and the performance of the communication system is significantly improved..4 Thesis Outline The rest of this thesis is organized as follows: In Chapter 2, we will state the class of nonlinear IO channels to be identified. Then, we will present the already proposed bloc-oriented NN identification structure. Finally we survey the different learning algorithms that will be used to train the proposed NN scheme. The algorithms to be studied are a combination of nonlinear adaptive algorithms, ordinary Bac-Propagation (BP) and Natural Gradient Bac-Propagation (NGBP), and 6

24 linear adaptive algorithms, Least ean Squares (LS) and Recursive Least Squares (RLS). In Chapter 3, we test the NN scheme by simulating a 2 2 IO system. The scheme performance is examined and compared under each of the algorithms discussed in Chapter 2. The comparison is done in terms of SE, convergence speed, and nonlinearity approximation. In addition, by simulating a 3 3 IO system, we illustrate that the NN identification structure can be extended to model IO systems with higher number of antennas. Finally, we apply the NN scheme to the tracing of time-varying nonlinear IO channels. In Chapter 4, we modify the improved NN model to mae it more suitable for complexvalued signals. The new proposed NN model will be able to model the phase shift caused by the HPAs. In Chapter 5, we introduce a neural networ V-BLAST receiver for nonlinear IO channels. The receiver is composed of a neural networ channel estimator and a ZF V- BLAST detection algorithm. We also propose a NN based linearization technique for HPAs, which can be used either at the transmitter or receiver side. In Chapter 6, we will summarize the conclusion and mae suggestions for future wor. 7

25 Chapter 2 Nonlinear IO Channel odel and Neural Networ Scheme In this chapter, we state the nonlinear IO channel model that will be used throughout this thesis. Then, we delineate the NN scheme used for modeling the nonlinear channel. Finally, we present several learning algorithms used to train the proposed NN structure. 2. Nonlinear IO Channel odel x () n g (.) h N () n y () n x 2 () n g 2 (.) h N 2 () n y 2 () n x () n g (.) h L N L () n y L () n h L Fig. 2.: Nonlinear IO channel Fig. 2. shows the nonlinear IO channel considered in this thesis [5]. The system consists of uncorrelated inputs x (, K, ) transformed by a memory-less nonlinearity g (.) then linearly combined by L matrix H = [ ]. = each of which is nonlinearly h j. The outputs of these nonlinearities are 8

26 The th j output can be expressed as: j h ji g i ( xi ) + N j, j = L ( 2. y L ) = i= where, N j is a white noise. The system output-input relation can be expressed in a matrix form as: y y y 2 L g = g H g ( x ) ( x ) 2 2 ( x ) + N N N 2 L ( 2.2) Depending on the application, the linear combiner H could be static or time varying. Our modeling approach assumes that only the structure of the nonlinear IO channel is nown. In other words, we now that the IO channel is composed of nonlinear memory-less blocs and a linear combining matrix, but we do not now the values or the behavior of those components. 2.2 Neural Networ Scheme The neural networ scheme proposed for modeling the nonlinear IO channel is shown in Fig. 2.2 [7]. It is composed of NN blocs, modeling the channel nonlinearities, followed by adaptive L linear combiner, modeling the linear component of the channel. 9

27 b Bloc a f c x () n f NN w s a N b N c N w f b 2 Bloc 2 a 2 f c 2 x 2 () n f NN 2 s 2 a 2N b 2N c 2N f b Bloc a f c w L x () n f NN w L s L c N a N b N f Fig. 2.2: NN identification structure 0

28 Each bloc has a scalar input x ( =, L, ), N neurons and a scalar output: () C N NN = ci f ( ai x + bi ), =, L, ( 2. 3) i= where f. is the NN activation function, a, b, c represent respectively, the input th th weight, output weight, and bias weight of the i neuron of the bloc. The output i i i NN th th of the bloc is connected to the j output of the system through weight w. j The system th j output is then expressed as: j w j NN, j =,, L ( 2.4) s L = = The decoupling constraint C is introduced to avoid error propagation ambiguity (which part of the nonlinear channel caused the error), where vector output weight set, C = c, K, c ], of blocs. [ N C represents the The NN model output-input relation can be expressed in a matrix form as: s s s 2 L NN = NN W NN ( x ) ( x ) 2 2 ( x ) ( 2.5) where W [ w j ] ( j =, L, L; =, L, = ) is the weight matrix. 2.3 Learning Algorithms The NN model uses supervised learning during the training process to update the weight parameters. The unnown IO channel and the proposed NN model are fed with the same input vector as shown in Fig At each iteration, the NN parameters are updated

29 in order to minimize a cost function, which is taen here to be the sum of the squared errors between the nown channel outputs and the corresponding model outputs: where e y s j =. j j J ( n ) e ( n ) 2 ( j ) ( 2.6 ) = L j = x x 2 x () n Unnown nonlinear IO channel () n N N 2 N L y L () n y () n y 2 () n x () n x 2 x () n () n Adaptive NN scheme s s 2 s L e L () n e 2 e Learning algorithm Fig. 2.3: Adaptive system diagram The tas of training the NN scheme can be split into two parts Nonlinear training: this part corresponds to the training of the memory-less NN blocs and it uses nonlinear adaptive algorithms to update the NN weights, a, c, b ( =, K, ; i =, K N ). i i i, Linear training: this part uses linear adaptive algorithms to update the coefficients of the linear combiner matrix W [ w j ] ( j =, K, L; =, K, ) =. 2

30 2.3. Nonlinear Adaptive Algorithms The system parameter vector for NN bloc will be denoted by θ, which includes all NN parameters to be updated in bloc θ = [ a K, a, b, K, b, c, K c ] t, N N N Ordinary Gradient Descent Bac-Propagation (BP) Algorithm [7] The BP algorithm updates the weight by following the ordinary gradient descent of the error surface: θ ( n + ) = θ µ J ( 2.7) θ where µ is a small positive constant and respect to matrix θ, which is expressed as: θ represents the ordinary gradient with θ J L = 2 e j θ s j ( 2.8) j= where θ s j cx c N x c = cn f f f f ( a x + b ) f ( an x + bn ) f ( a x + b ) w ( an x + bn ) ( a x + b ) w ( a x + b ) w N N w j j j w j w j j Natural Gradient Descent Bac-Propagation (NGBP) Algorithm The ordinary gradient is the steepest descent direction of a cost function if the space of parameters is an orthonormal coordinate system. It has been shown [, 20, 4, 43] that 3

31 in the case of multilayer nets, the steepest descent direction of the loss function is actually given by: where ~ θ J θ = G J ( 2.9) G is the inverse of the Fisher information matrix (FI): G = g i, j = g i, j = E J θ i J θ j ( ) ( 2.0) n Therefore, the neural networ weights will be updated as follows: θ ~ ( n + ) = θ µ J ( 2.) θ The calculation of the expectation in the expression of G requires the probability ( ) ( ) distribution of the inputs x n =, L,, which is unnown in most cases. oreover, the inversion of is computationally costly. To obtain directly, a G Kalman filter technique is used [2, 20]: ( n ) = ( + ε ) G ε G s s G t ~ ( ) G ( 2.2) ~ ~ ~ G + n n θ j θ A search-and-converge schedule will be used for ε n in order to obtain a good trade-off between convergence speed and stability: j ε n ε 0 + c n ε = τ 2 + c n ε ε n τ 0 + τ ( 2.3) such that small n corresponds to a search phase ( ε n is close to ε 0 ) and large n corresponds to a convergence phase ( ε n is equivalent to c ε for large n), where ε n 0, cε and τ are positive real constants. Using this online Kalman filter technique, the update of the weights (Eqn. 2.) becomes: θ ~ ( n + ) = θ µ G J ( 2.4 ) θ 4

32 2.3.2 Linear Adaptive Algorithms Least-ean-Square (LS) Algorithm [3] The LS algorithm is one of the most widely used adaptive filtering algorithms. It is a stochastic gradient descent method in that the weights are only updated based on the error signal at the current time. The error is defined to be the difference between the desired signal and the actual signal. The goal is to minimize the underlying cost function, which is achieved by minimizing the error signal. The LS algorithm may be described in words as follows: updatedvalue old value learning tap error of tap weight = of tap weight + rate input signal vector vector parameter vector Therefore, updating the elements of matrix W is done as follows: s e j j w w NN ( 2.5) = = y j s j ( 2.6) ( n + ) = w + 2µ e NN ( 2.7) j = j j The LS algorithm is simple to implement, however, its major limitation is a relatively slow rate of convergence. j Recursive Least-Square (RLS) Algorithm [3, 6] The RLS algorithm may be described in words as follows: updatedvalue old value Kalman innovation of the = of the + gain vector state state Hence, we update the elements of matrix W as follows: K ( n + ) pˆ = λ + NN NN Pˆ NN ( 2.8) 5

33 s e w w NN ( 2.9) j j = = y j s j ( 2.20) ( n + ) = w + K e ( 2.2) j = j j ( n + ) = Pˆ K NN Pˆ j ( ) ( 2.22) ˆ P λ The RLS forgetting factor is λ = and the initial value of P is P(0) = I, where I is the identity matrix. The RLS algorithm is capable of realizing a rate of convergence that is, in general, much faster than the LS algorithm, because the RLS algorithm utilizes all the information contained in the input data from the start of the adaptation up to the present. However, this is attained at the expanse of extensive computational complexity. The learning curve is defined as the evolution of the mean squared error during the learning process. The SE error starts to decrease until the system reaches a steady state and only slight changes in the weights are observed. The learning process may then be stopped if the modeled IO channel is static, however, if the IO channel is time varying, then the learning process should continuous in order to eep tracing of the time-varying parameters. 6

34 Chapter 3 Neural Networ odeling and Identification of Nonlinear IO Channels In this chapter we evaluate, through simulation, the performance of the NN scheme described in Chapter 2. The algorithms: LS-BP, LS-NGBP, RLS-BP, and RLS- NGBP are tested and examined in training the NN scheme where the performance of the scheme under each algorithm is compared in terms of convergence speed, SE performance and nonlinearity approximation. 3. Study Case For the purpose of conducting the simulation we choose a study case which we thin is comprehensive and good for illustrating the results. In this chapter, the inputs are Zeromean white Gaussian processes with unit variance. A 2 2 IO system is considered α x + β x with the unnown nonlinearities to be identified chosen to be: g( x) = 2 2 β x α, where α = β2 = 2, α 2 = β =. These nonlinear functions are and ( ) 2 g = exp 2 2 x 2 x reasonable models for amplitudes conversions for nonlinear high power amplifiers used in digital communications. For this chapter, the noise is taen as white Gaussian with varianceσ = Later in this chapter the simulation will be extended to include 3 3 IO systems. Each bloc in the proposed NN scheme is composed of N = 5 neurons. The NN activation function will be taen as the erf function. The RLS forgetting factor is taen to 2 7

35 be λ = We have tested the NN model for a range of values of µ belonging to the interval [ ]. 3.2 odeling and Identification of Nonlinear IO Channels In this section we consider a fixed IO channel where the linear combiner matrix is static. In other words, the matrix H is assigned fixed values and it does not change during 0.3 simulation. For illustration purposes we choose H to be: H =. 0.3 The SE vs. µ curves of BP and NG (at 50,000 iterations) are shown in Fig. 3.. It can be seen that µ = represents an optimal value for the SE performance of LS-BP and RLS-BP, while the optimal SE value of LS-NGBP and LS-NGBP is achieved when µ = Thus, when the learning rate is small the SE curve converges slowly and the SE value remains high, as a minimum is not yet reached. On the other hand, when the learning rate is large the NN performance will experience an increase in the SE value, moving away from the minimum, and eventually will become unstable. Therefore, to achieve a near optimum performance, one must carefully choose the learning rate which results in the lowest SE. 8

36 SE 0-4 RLS-BP 0-5 LS-BP LS-NGBP RLS-NGBP µ Fig. 3.: SE vs. µ: comparison between BP and NG based algorithms ( = L = 2, N = 5, SNR = 60 db, Zero-mean white Gaussian input) Fig. 3.2, Fig. 3.6, Fig. 3.0, and Fig. 3.4 show the learning curves of the LS-BP, RLS- BP, LS-NGBP and RLS-NGBP algorithms, respectively. Table 3. lists the SE for the different algorithms. The SE values are calculated by averaging over the last 5,000 iterations of the SE curves. Fig. 3.3, Fig. 3.7, Fig.3., and Fig. 3.5 demonstrate the transient behavior of matrix W weights. The memory-less nonlinearities have been successfully identified by the NN approach, (see Fig. 3.4, Fig. 3.5, Fig.3.8, Fig. 3.9, Fig.3.2, Fig. 3.3, Fig.3.6, and Fig. 3.7). It is important to note that each of the nonlinearities as well as the coupling matrix H have been identified to within scaling factors. That is: NN ( x) has converged to ( x) γ, NN 2 ( x) has converged to γ ( x) g 2 g 2, and 9

37 W has converged to γ 0 0 H γ 2, where γ and γ 2 are real constants. The scaling factors are determined by comparing the average gain of the unnown channel nonlinearities and the average gain of the corresponding NN blocs. For instance, if the average gain of the unnown nonlinearity, (.) g is giving by: gaverage output power g Average Gain = and the average gain of the corresponding NN bloc, g average input power NN (). is given by: NNaverage output power NN Average Gain =, then the scaling factor, γ is given NN average input power Average Gain by: γ NN = g. Average Gain Table 3.: SE and scaling factors for the different algorithms Scaling factors Algorithm SE γ γ 2 LS-BP RLS-BP LS-NGBP RLS-NGBP

38 SE Iteration, n x 0 4 Fig. 3.2: Smoothed SE curve (LS-BP, = L = 2, µ = 0. 09, N = 5, SNR = 60 db, Zeromean white Gaussian input).2 h, h w w w 2 h 2, h w Iteration, n x 0 4 Fig. 3.3: Evolution of the normalized weights (LS-BP, = L = 2, µ = 0. 09, N = 5, SNR = 60 db, Zero-mean white Gaussian input) 2

39 .4.2 Nonlinearity approximation True nonlinearity Output, g (x) Input, x Fig. 3.4: HPA nonlinearity g (x) and normalized NN (x) (LS-BP, = L = 2, µ = 0. 09, N = 5, SNR = 60dB, Zero-mean white Gaussian input) True nonlinearity Nonlinearity approximation Output, g 2 (x) Input, x Fig. 3.5: HPA nonlinearity g 2 (x) and normalized NN 2 (x) (LS-BP, = L = 2, µ = 0. 09, N = 5, SNR = 60 db, Zero-mean white Gaussian input) 22

40 SE Iteration, n x 0 4 Fig. 3.6: Smoothed SE curve (RLS-BP, = L = 2, µ = 0. 09, N = 5, SNR = 60 db, Zeromean white Gaussian input).2 h, h 22 w 0.8 w h 2, h w 2 w Iteration, n (x50 iterations) Fig 3.7: Evolution of the normalized weights (RLS-BP, = L = 2, µ = 0. 09, N = 5, SNR = 60 db, Zero-mean white Gaussian input) 23

41 .4.2 Nonlinearity approximation True nonlinearity Output, g (x) Input, x Fig. 3.8: HPA nonlinearity g (x) and normalized NN (x) (RLS-BP, = L = 2, µ = 0. 09, N = 5, SNR = 60 db, Zero-mean white Gaussian input) Nonlinearity approximation 0.3 True nonlinearity Output, g 2 (x) Input, x Fig. 3.9: HPA nonlinearity g 2 (x) and normalized NN 2 (x) (RLS-BP, = L = 2, µ = 0. 09, N = 5, SNR = 60 db, Zero-mean white Gaussian input) 24

42 SE Iteration, n x 0 4 Fig. 3.0: Smoothed SE curve (LS-NGBP, = L = 2, µ = , N = 5, SNR = 60 db, Zero-mean white Gaussian input) Fig. 3.: Evolution of the normalized weights (LS-NGBP, = L = 2, µ = , N = 5, SNR = 60 db, Zero-mean white Gaussian input) 25

43 Output, g (x) 0.6 True nonlinearity, Nonlinearity approximatioin Input, x Fig. 3.2: HPA nonlinearity g (x) and normalized NN (x) (LS-NGBP, = L = 2, µ = , N = 5, SNR = 60 db, Zero-mean white Gaussian input) True nonlinearity, Nonlinearity approximation Output, g 2 (x) Input, x Fig. 3.3: HPA nonlinearity g 2 (x) and normalized NN 2 (x) (LS-NGBP, = L = 2, µ = , N = 5, SNR = 60 db, Zero-mean white Gaussian input) 26

44 SE Iteration, n x 0 4 Fig. 3.4: Smoothed SE curve (RLS-NGBP, = L = 2, µ = , N = 5, SNR = 60 db, Zero-mean white Gaussian input) Fig. 3.5: Evolution of the normalized weights (RLS-NGBP, = L = 2, µ = , N = 5, SNR = 60 db, Zero-mean white Gaussian input) 27

45 Output, g (x) 0.6 True nonlinearity, Nonlinearity approximation Input, x Fig. 3.6: HPA nonlinearity g (x) and normalized NN (x) (RLS-NGBP, = L = 2, µ = , N = 5, SNR = 60 db, Zero-mean white Gaussian input) True nonlinearity, Nonlinearity approximation Output, g 2 (x) Input, x Fig. 3.7: HPA nonlinearity g 2 (x) and normalized NN 2 (x) (RLS-NGBP, = L = 2, µ = , N = 5, SNR = 60 db, Zero-mean white Gaussian input) 28

46 Discussion From the figures above it is evident that the NN scheme has successfully modeled and identified the IO channel under testing. Yet, the performance varies depending on the learning algorithm used for training the NN. For instance, Fig. 3.7 exhibits evolution of the normalized weights at a faster convergence speed than it is in Fig This should not be surprising, since the rate of convergence of the RLS algorithm is typically an order of magnitude faster than that of the LS algorithm [3]. Also, Table 3. shows that ordinary BP descent based algorithms tend to converge to an SE value of approximately 4 0-5, while NG descent based algorithms tend to converge to an SE value of approximately This can be explained by that fact that the SE is mostly controlled by the channel nonlinearities, which taes more time to converge. Nonlinearities are trained using nonlinear adaptive algorithms (i.e. BP and NGBP). Furthermore, simulation results show that NG descent based algorithms outperform BP descent based algorithms in terms of convergence speed, SE performance and nonlinear approximation. NG has better capabilities of avoiding the plateau phenomena, which is typical of BP learning curves, because it accounts for the space parameters in which the NN weights evolve [20]. This yields faster convergence speed and lower SE performance. Fig. 3.8 demonstrates the considerable convergence speed and SE improvement obtained by employing NG instead of BP. We can see that while BP SE curves converge to a general value of 4 0-5, NG SE curves tend to converge to a general value of resulting in lower SE. In addition, Fig. 3.8 shows that, in order to achieve an SE of 0-4 the NG needs less than 5,000 iterations, whereas the BP needs more than 0,000 iterations. 29

12.4 Known Channel (Water-Filling Solution)

12.4 Known Channel (Water-Filling Solution) ECEn 665: Antennas and Propagation for Wireless Communications 54 2.4 Known Channel (Water-Filling Solution) The channel scenarios we have looed at above represent special cases for which the capacity

More information

Multilayer Perceptron

Multilayer Perceptron Outline Hong Chang Institute of Computing Technology, Chinese Academy of Sciences Machine Learning Methods (Fall 2012) Outline Outline I 1 Introduction 2 Single Perceptron 3 Boolean Function Learning 4

More information

Lecture 7 MIMO Communica2ons

Lecture 7 MIMO Communica2ons Wireless Communications Lecture 7 MIMO Communica2ons Prof. Chun-Hung Liu Dept. of Electrical and Computer Engineering National Chiao Tung University Fall 2014 1 Outline MIMO Communications (Chapter 10

More information

Lecture 8: MIMO Architectures (II) Theoretical Foundations of Wireless Communications 1. Overview. Ragnar Thobaben CommTh/EES/KTH

Lecture 8: MIMO Architectures (II) Theoretical Foundations of Wireless Communications 1. Overview. Ragnar Thobaben CommTh/EES/KTH MIMO : MIMO Theoretical Foundations of Wireless Communications 1 Wednesday, May 25, 2016 09:15-12:00, SIP 1 Textbook: D. Tse and P. Viswanath, Fundamentals of Wireless Communication 1 / 20 Overview MIMO

More information

Lecture 16: Introduction to Neural Networks

Lecture 16: Introduction to Neural Networks Lecture 16: Introduction to Neural Networs Instructor: Aditya Bhasara Scribe: Philippe David CS 5966/6966: Theory of Machine Learning March 20 th, 2017 Abstract In this lecture, we consider Bacpropagation,

More information

Supplementary Figure 1: Scheme of the RFT. (a) At first, we separate two quadratures of the field (denoted by and ); (b) then, each quadrature

Supplementary Figure 1: Scheme of the RFT. (a) At first, we separate two quadratures of the field (denoted by and ); (b) then, each quadrature Supplementary Figure 1: Scheme of the RFT. (a At first, we separate two quadratures of the field (denoted by and ; (b then, each quadrature undergoes a nonlinear transformation, which results in the sine

More information

An artificial neural networks (ANNs) model is a functional abstraction of the

An artificial neural networks (ANNs) model is a functional abstraction of the CHAPER 3 3. Introduction An artificial neural networs (ANNs) model is a functional abstraction of the biological neural structures of the central nervous system. hey are composed of many simple and highly

More information

SPEECH ANALYSIS AND SYNTHESIS

SPEECH ANALYSIS AND SYNTHESIS 16 Chapter 2 SPEECH ANALYSIS AND SYNTHESIS 2.1 INTRODUCTION: Speech signal analysis is used to characterize the spectral information of an input speech signal. Speech signal analysis [52-53] techniques

More information

ADAPTIVE FILTER THEORY

ADAPTIVE FILTER THEORY ADAPTIVE FILTER THEORY Fourth Edition Simon Haykin Communications Research Laboratory McMaster University Hamilton, Ontario, Canada Front ice Hall PRENTICE HALL Upper Saddle River, New Jersey 07458 Preface

More information

Signal Processing for Digital Data Storage (11)

Signal Processing for Digital Data Storage (11) Outline Signal Processing for Digital Data Storage (11) Assist.Prof. Piya Kovintavewat, Ph.D. Data Storage Technology Research Unit Nahon Pathom Rajabhat University Partial-Response Maximum-Lielihood (PRML)

More information

Adaptive Filtering Part II

Adaptive Filtering Part II Adaptive Filtering Part II In previous Lecture we saw that: Setting the gradient of cost function equal to zero, we obtain the optimum values of filter coefficients: (Wiener-Hopf equation) Adaptive Filtering,

More information

Lessons in Estimation Theory for Signal Processing, Communications, and Control

Lessons in Estimation Theory for Signal Processing, Communications, and Control Lessons in Estimation Theory for Signal Processing, Communications, and Control Jerry M. Mendel Department of Electrical Engineering University of Southern California Los Angeles, California PRENTICE HALL

More information

LECTURE # - NEURAL COMPUTATION, Feb 04, Linear Regression. x 1 θ 1 output... θ M x M. Assumes a functional form

LECTURE # - NEURAL COMPUTATION, Feb 04, Linear Regression. x 1 θ 1 output... θ M x M. Assumes a functional form LECTURE # - EURAL COPUTATIO, Feb 4, 4 Linear Regression Assumes a functional form f (, θ) = θ θ θ K θ (Eq) where = (,, ) are the attributes and θ = (θ, θ, θ ) are the function parameters Eample: f (, θ)

More information

AdaptiveFilters. GJRE-F Classification : FOR Code:

AdaptiveFilters. GJRE-F Classification : FOR Code: Global Journal of Researches in Engineering: F Electrical and Electronics Engineering Volume 14 Issue 7 Version 1.0 Type: Double Blind Peer Reviewed International Research Journal Publisher: Global Journals

More information

ADAPTIVE INVERSE CONTROL BASED ON NONLINEAR ADAPTIVE FILTERING. Information Systems Lab., EE Dep., Stanford University

ADAPTIVE INVERSE CONTROL BASED ON NONLINEAR ADAPTIVE FILTERING. Information Systems Lab., EE Dep., Stanford University ADAPTIVE INVERSE CONTROL BASED ON NONLINEAR ADAPTIVE FILTERING Bernard Widrow 1, Gregory Plett, Edson Ferreira 3 and Marcelo Lamego 4 Information Systems Lab., EE Dep., Stanford University Abstract: Many

More information

MMSE Decision Feedback Equalization of Pulse Position Modulated Signals

MMSE Decision Feedback Equalization of Pulse Position Modulated Signals SE Decision Feedback Equalization of Pulse Position odulated Signals AG Klein and CR Johnson, Jr School of Electrical and Computer Engineering Cornell University, Ithaca, NY 4853 email: agk5@cornelledu

More information

Various Nonlinear Models and their Identification, Equalization and Linearization

Various Nonlinear Models and their Identification, Equalization and Linearization Various Nonlinear Models and their Identification, Equalization and Linearization A THESIS SUBMITTED IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF Master of Technology in Telematics and

More information

Channel Estimation with Low-Precision Analog-to-Digital Conversion

Channel Estimation with Low-Precision Analog-to-Digital Conversion Channel Estimation with Low-Precision Analog-to-Digital Conversion Onkar Dabeer School of Technology and Computer Science Tata Institute of Fundamental Research Mumbai India Email: onkar@tcs.tifr.res.in

More information

The Sorted-QR Chase Detector for Multiple-Input Multiple-Output Channels

The Sorted-QR Chase Detector for Multiple-Input Multiple-Output Channels The Sorted-QR Chase Detector for Multiple-Input Multiple-Output Channels Deric W. Waters and John R. Barry School of ECE Georgia Institute of Technology Atlanta, GA 30332-0250 USA {deric, barry}@ece.gatech.edu

More information

Expectation propagation for signal detection in flat-fading channels

Expectation propagation for signal detection in flat-fading channels Expectation propagation for signal detection in flat-fading channels Yuan Qi MIT Media Lab Cambridge, MA, 02139 USA yuanqi@media.mit.edu Thomas Minka CMU Statistics Department Pittsburgh, PA 15213 USA

More information

Data Detection for Controlled ISI. h(nt) = 1 for n=0,1 and zero otherwise.

Data Detection for Controlled ISI. h(nt) = 1 for n=0,1 and zero otherwise. Data Detection for Controlled ISI *Symbol by symbol suboptimum detection For the duobinary signal pulse h(nt) = 1 for n=0,1 and zero otherwise. The samples at the output of the receiving filter(demodulator)

More information

Revision of Lecture 4

Revision of Lecture 4 Revision of Lecture 4 We have discussed all basic components of MODEM Pulse shaping Tx/Rx filter pair Modulator/demodulator Bits map symbols Discussions assume ideal channel, and for dispersive channel

More information

ADAPTIVE CHANNEL EQUALIZATION USING RADIAL BASIS FUNCTION NETWORKS AND MLP SHEEJA K.L.

ADAPTIVE CHANNEL EQUALIZATION USING RADIAL BASIS FUNCTION NETWORKS AND MLP SHEEJA K.L. ADAPTIVE CHANNEL EQUALIZATION USING RADIAL BASIS FUNCTION NETWORKS AND MLP SHEEJA K.L. Department of Electrical Engineering National Institute of Technology Rourkela Rourkela-769008, Orissa, India. i ADAPTIVE

More information

Digital Predistortion Using Machine Learning Algorithms

Digital Predistortion Using Machine Learning Algorithms Digital Predistortion Using Machine Learning Algorithms Introduction CS229 ; James Peroulas ; james@peroulas.com Wireless communications transmitter The power amplifier (PA) is the last stage of a wireless

More information

Adaptive Filter Theory

Adaptive Filter Theory 0 Adaptive Filter heory Sung Ho Cho Hanyang University Seoul, Korea (Office) +8--0-0390 (Mobile) +8-10-541-5178 dragon@hanyang.ac.kr able of Contents 1 Wiener Filters Gradient Search by Steepest Descent

More information

AI Programming CS F-20 Neural Networks

AI Programming CS F-20 Neural Networks AI Programming CS662-2008F-20 Neural Networks David Galles Department of Computer Science University of San Francisco 20-0: Symbolic AI Most of this class has been focused on Symbolic AI Focus or symbols

More information

Lecture 14 October 22

Lecture 14 October 22 EE 2: Coding for Digital Communication & Beyond Fall 203 Lecture 4 October 22 Lecturer: Prof. Anant Sahai Scribe: Jingyan Wang This lecture covers: LT Code Ideal Soliton Distribution 4. Introduction So

More information

Cooperative Spectrum Prediction for Improved Efficiency of Cognitive Radio Networks

Cooperative Spectrum Prediction for Improved Efficiency of Cognitive Radio Networks Cooperative Spectrum Prediction for Improved Efficiency of Cognitive Radio Networks by Nagwa Shaghluf B.S., Tripoli University, Tripoli, Libya, 2009 A Thesis Submitted in Partial Fulfillment of the Requirements

More information

Square Root Raised Cosine Filter

Square Root Raised Cosine Filter Wireless Information Transmission System Lab. Square Root Raised Cosine Filter Institute of Communications Engineering National Sun Yat-sen University Introduction We consider the problem of signal design

More information

EM-algorithm for Training of State-space Models with Application to Time Series Prediction

EM-algorithm for Training of State-space Models with Application to Time Series Prediction EM-algorithm for Training of State-space Models with Application to Time Series Prediction Elia Liitiäinen, Nima Reyhani and Amaury Lendasse Helsinki University of Technology - Neural Networks Research

More information

Engineering Part IIB: Module 4F10 Statistical Pattern Processing Lecture 5: Single Layer Perceptrons & Estimating Linear Classifiers

Engineering Part IIB: Module 4F10 Statistical Pattern Processing Lecture 5: Single Layer Perceptrons & Estimating Linear Classifiers Engineering Part IIB: Module 4F0 Statistical Pattern Processing Lecture 5: Single Layer Perceptrons & Estimating Linear Classifiers Phil Woodland: pcw@eng.cam.ac.uk Michaelmas 202 Engineering Part IIB:

More information

WIRELESS COMMUNICATIONS AND COGNITIVE RADIO TRANSMISSIONS UNDER QUALITY OF SERVICE CONSTRAINTS AND CHANNEL UNCERTAINTY

WIRELESS COMMUNICATIONS AND COGNITIVE RADIO TRANSMISSIONS UNDER QUALITY OF SERVICE CONSTRAINTS AND CHANNEL UNCERTAINTY University of Nebraska - Lincoln DigitalCommons@University of Nebraska - Lincoln Theses, Dissertations, and Student Research from Electrical & Computer Engineering Electrical & Computer Engineering, Department

More information

Pattern Classification

Pattern Classification Pattern Classification All materials in these slides were taen from Pattern Classification (2nd ed) by R. O. Duda,, P. E. Hart and D. G. Stor, John Wiley & Sons, 2000 with the permission of the authors

More information

Hand Written Digit Recognition using Kalman Filter

Hand Written Digit Recognition using Kalman Filter International Journal of Electronics and Communication Engineering. ISSN 0974-2166 Volume 5, Number 4 (2012), pp. 425-434 International Research Publication House http://www.irphouse.com Hand Written Digit

More information

Stochastic Analogues to Deterministic Optimizers

Stochastic Analogues to Deterministic Optimizers Stochastic Analogues to Deterministic Optimizers ISMP 2018 Bordeaux, France Vivak Patel Presented by: Mihai Anitescu July 6, 2018 1 Apology I apologize for not being here to give this talk myself. I injured

More information

Digital Band-pass Modulation PROF. MICHAEL TSAI 2011/11/10

Digital Band-pass Modulation PROF. MICHAEL TSAI 2011/11/10 Digital Band-pass Modulation PROF. MICHAEL TSAI 211/11/1 Band-pass Signal Representation a t g t General form: 2πf c t + φ t g t = a t cos 2πf c t + φ t Envelope Phase Envelope is always non-negative,

More information

Lecture 7: Wireless Channels and Diversity Advanced Digital Communications (EQ2410) 1

Lecture 7: Wireless Channels and Diversity Advanced Digital Communications (EQ2410) 1 Wireless : Wireless Advanced Digital Communications (EQ2410) 1 Thursday, Feb. 11, 2016 10:00-12:00, B24 1 Textbook: U. Madhow, Fundamentals of Digital Communications, 2008 1 / 15 Wireless Lecture 1-6 Equalization

More information

EE4512 Analog and Digital Communications Chapter 4. Chapter 4 Receiver Design

EE4512 Analog and Digital Communications Chapter 4. Chapter 4 Receiver Design Chapter 4 Receiver Design Chapter 4 Receiver Design Probability of Bit Error Pages 124-149 149 Probability of Bit Error The low pass filtered and sampled PAM signal results in an expression for the probability

More information

Revision of Lecture 4

Revision of Lecture 4 Revision of Lecture 4 We have completed studying digital sources from information theory viewpoint We have learnt all fundamental principles for source coding, provided by information theory Practical

More information

The interference-reduced energy loading for multi-code HSDPA systems

The interference-reduced energy loading for multi-code HSDPA systems Gurcan et al. EURASIP Journal on Wireless Communications and Networing 2012, 2012:127 RESEARC Open Access The interference-reduced energy loading for multi-code SDPA systems Mustafa K Gurcan *, Irina Ma

More information

Fast Near-Optimal Energy Allocation for Multimedia Loading on Multicarrier Systems

Fast Near-Optimal Energy Allocation for Multimedia Loading on Multicarrier Systems Fast Near-Optimal Energy Allocation for Multimedia Loading on Multicarrier Systems Michael A. Enright and C.-C. Jay Kuo Department of Electrical Engineering and Signal and Image Processing Institute University

More information

Recursive Least Squares for an Entropy Regularized MSE Cost Function

Recursive Least Squares for an Entropy Regularized MSE Cost Function Recursive Least Squares for an Entropy Regularized MSE Cost Function Deniz Erdogmus, Yadunandana N. Rao, Jose C. Principe Oscar Fontenla-Romero, Amparo Alonso-Betanzos Electrical Eng. Dept., University

More information

DETECTING PROCESS STATE CHANGES BY NONLINEAR BLIND SOURCE SEPARATION. Alexandre Iline, Harri Valpola and Erkki Oja

DETECTING PROCESS STATE CHANGES BY NONLINEAR BLIND SOURCE SEPARATION. Alexandre Iline, Harri Valpola and Erkki Oja DETECTING PROCESS STATE CHANGES BY NONLINEAR BLIND SOURCE SEPARATION Alexandre Iline, Harri Valpola and Erkki Oja Laboratory of Computer and Information Science Helsinki University of Technology P.O.Box

More information

Maximum Likelihood Diffusive Source Localization Based on Binary Observations

Maximum Likelihood Diffusive Source Localization Based on Binary Observations Maximum Lielihood Diffusive Source Localization Based on Binary Observations Yoav Levinboo and an F. Wong Wireless Information Networing Group, University of Florida Gainesville, Florida 32611-6130, USA

More information

Diffusion LMS Algorithms for Sensor Networks over Non-ideal Inter-sensor Wireless Channels

Diffusion LMS Algorithms for Sensor Networks over Non-ideal Inter-sensor Wireless Channels Diffusion LMS Algorithms for Sensor Networs over Non-ideal Inter-sensor Wireless Channels Reza Abdolee and Benoit Champagne Electrical and Computer Engineering McGill University 3480 University Street

More information

Multilayer Perceptron = FeedForward Neural Network

Multilayer Perceptron = FeedForward Neural Network Multilayer Perceptron = FeedForward Neural Networ History Definition Classification = feedforward operation Learning = bacpropagation = local optimization in the space of weights Pattern Classification

More information

LECTURE 16 AND 17. Digital signaling on frequency selective fading channels. Notes Prepared by: Abhishek Sood

LECTURE 16 AND 17. Digital signaling on frequency selective fading channels. Notes Prepared by: Abhishek Sood ECE559:WIRELESS COMMUNICATION TECHNOLOGIES LECTURE 16 AND 17 Digital signaling on frequency selective fading channels 1 OUTLINE Notes Prepared by: Abhishek Sood In section 2 we discuss the receiver design

More information

PHASE RETRIEVAL OF SPARSE SIGNALS FROM MAGNITUDE INFORMATION. A Thesis MELTEM APAYDIN

PHASE RETRIEVAL OF SPARSE SIGNALS FROM MAGNITUDE INFORMATION. A Thesis MELTEM APAYDIN PHASE RETRIEVAL OF SPARSE SIGNALS FROM MAGNITUDE INFORMATION A Thesis by MELTEM APAYDIN Submitted to the Office of Graduate and Professional Studies of Texas A&M University in partial fulfillment of the

More information

Artificial Neural Network : Training

Artificial Neural Network : Training Artificial Neural Networ : Training Debasis Samanta IIT Kharagpur debasis.samanta.iitgp@gmail.com 06.04.2018 Debasis Samanta (IIT Kharagpur) Soft Computing Applications 06.04.2018 1 / 49 Learning of neural

More information

Analysis of Finite Wordlength Effects

Analysis of Finite Wordlength Effects Analysis of Finite Wordlength Effects Ideally, the system parameters along with the signal variables have infinite precision taing any value between and In practice, they can tae only discrete values within

More information

Multilayer Perceptrons (MLPs)

Multilayer Perceptrons (MLPs) CSE 5526: Introduction to Neural Networks Multilayer Perceptrons (MLPs) 1 Motivation Multilayer networks are more powerful than singlelayer nets Example: XOR problem x 2 1 AND x o x 1 x 2 +1-1 o x x 1-1

More information

Timing Recovery at Low SNR Cramer-Rao bound, and outperforming the PLL

Timing Recovery at Low SNR Cramer-Rao bound, and outperforming the PLL T F T I G E O R G A I N S T I T U T E O H E O F E A L P R O G R ESS S A N D 1 8 8 5 S E R V L O G Y I C E E C H N O Timing Recovery at Low SNR Cramer-Rao bound, and outperforming the PLL Aravind R. Nayak

More information

Improved Multiple Feedback Successive Interference Cancellation Algorithm for Near-Optimal MIMO Detection

Improved Multiple Feedback Successive Interference Cancellation Algorithm for Near-Optimal MIMO Detection Improved Multiple Feedback Successive Interference Cancellation Algorithm for Near-Optimal MIMO Detection Manish Mandloi, Mohammed Azahar Hussain and Vimal Bhatia Discipline of Electrical Engineering,

More information

Utilizing Correct Prior Probability Calculation to Improve Performance of Low-Density Parity- Check Codes in the Presence of Burst Noise

Utilizing Correct Prior Probability Calculation to Improve Performance of Low-Density Parity- Check Codes in the Presence of Burst Noise Utah State University DigitalCommons@USU All Graduate Theses and Dissertations Graduate Studies 5-2012 Utilizing Correct Prior Probability Calculation to Improve Performance of Low-Density Parity- Check

More information

The Modeling and Equalization Technique of Nonlinear Wireless Channel

The Modeling and Equalization Technique of Nonlinear Wireless Channel Send Orders for Reprints to reprints@benthamscience.ae The Open Cybernetics & Systemics Journal, 4, 8, 97-3 97 Open Access The Modeling and Equalization Technique of Nonlinear Wireless Channel Qiu Min

More information

Computation of Bit-Error Rate of Coherent and Non-Coherent Detection M-Ary PSK With Gray Code in BFWA Systems

Computation of Bit-Error Rate of Coherent and Non-Coherent Detection M-Ary PSK With Gray Code in BFWA Systems Computation of Bit-Error Rate of Coherent and Non-Coherent Detection M-Ary PSK With Gray Code in BFWA Systems Department of Electrical Engineering, College of Engineering, Basrah University Basrah Iraq,

More information

EFFECTS OF ILL-CONDITIONED DATA ON LEAST SQUARES ADAPTIVE FILTERS. Gary A. Ybarra and S.T. Alexander

EFFECTS OF ILL-CONDITIONED DATA ON LEAST SQUARES ADAPTIVE FILTERS. Gary A. Ybarra and S.T. Alexander EFFECTS OF ILL-CONDITIONED DATA ON LEAST SQUARES ADAPTIVE FILTERS Gary A. Ybarra and S.T. Alexander Center for Communications and Signal Processing Electrical and Computer Engineering Department North

More information

Lecture 4. Capacity of Fading Channels

Lecture 4. Capacity of Fading Channels 1 Lecture 4. Capacity of Fading Channels Capacity of AWGN Channels Capacity of Fading Channels Ergodic Capacity Outage Capacity Shannon and Information Theory Claude Elwood Shannon (April 3, 1916 February

More information

Condensed Table of Contents for Introduction to Stochastic Search and Optimization: Estimation, Simulation, and Control by J. C.

Condensed Table of Contents for Introduction to Stochastic Search and Optimization: Estimation, Simulation, and Control by J. C. Condensed Table of Contents for Introduction to Stochastic Search and Optimization: Estimation, Simulation, and Control by J. C. Spall John Wiley and Sons, Inc., 2003 Preface... xiii 1. Stochastic Search

More information

8. Lecture Neural Networks

8. Lecture Neural Networks Soft Control (AT 3, RMA) 8. Lecture Neural Networks Learning Process Contents of the 8 th lecture 1. Introduction of Soft Control: Definition and Limitations, Basics of Intelligent" Systems 2. Knowledge

More information

ARTIFICIAL NEURAL NETWORK PART I HANIEH BORHANAZAD

ARTIFICIAL NEURAL NETWORK PART I HANIEH BORHANAZAD ARTIFICIAL NEURAL NETWORK PART I HANIEH BORHANAZAD WHAT IS A NEURAL NETWORK? The simplest definition of a neural network, more properly referred to as an 'artificial' neural network (ANN), is provided

More information

Unit III. A Survey of Neural Network Model

Unit III. A Survey of Neural Network Model Unit III A Survey of Neural Network Model 1 Single Layer Perceptron Perceptron the first adaptive network architecture was invented by Frank Rosenblatt in 1957. It can be used for the classification of

More information

NUMERICAL COMPUTATION OF THE CAPACITY OF CONTINUOUS MEMORYLESS CHANNELS

NUMERICAL COMPUTATION OF THE CAPACITY OF CONTINUOUS MEMORYLESS CHANNELS NUMERICAL COMPUTATION OF THE CAPACITY OF CONTINUOUS MEMORYLESS CHANNELS Justin Dauwels Dept. of Information Technology and Electrical Engineering ETH, CH-8092 Zürich, Switzerland dauwels@isi.ee.ethz.ch

More information

Statistical and Adaptive Signal Processing

Statistical and Adaptive Signal Processing r Statistical and Adaptive Signal Processing Spectral Estimation, Signal Modeling, Adaptive Filtering and Array Processing Dimitris G. Manolakis Massachusetts Institute of Technology Lincoln Laboratory

More information

KNOWN approaches for improving the performance of

KNOWN approaches for improving the performance of IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS II: EXPRESS BRIEFS, VOL. 58, NO. 8, AUGUST 2011 537 Robust Quasi-Newton Adaptive Filtering Algorithms Md. Zulfiquar Ali Bhotto, Student Member, IEEE, and Andreas

More information

ECE 564/645 - Digital Communications, Spring 2018 Homework #2 Due: March 19 (In Lecture)

ECE 564/645 - Digital Communications, Spring 2018 Homework #2 Due: March 19 (In Lecture) ECE 564/645 - Digital Communications, Spring 018 Homework # Due: March 19 (In Lecture) 1. Consider a binary communication system over a 1-dimensional vector channel where message m 1 is sent by signaling

More information

The Viterbi Algorithm EECS 869: Error Control Coding Fall 2009

The Viterbi Algorithm EECS 869: Error Control Coding Fall 2009 1 Bacground Material 1.1 Organization of the Trellis The Viterbi Algorithm EECS 869: Error Control Coding Fall 2009 The Viterbi algorithm (VA) processes the (noisy) output sequence from a state machine

More information

Expected Error Based MMSE Detection Ordering for Iterative Detection-Decoding MIMO Systems

Expected Error Based MMSE Detection Ordering for Iterative Detection-Decoding MIMO Systems Expected Error Based MMSE Detection Ordering for Iterative Detection-Decoding MIMO Systems Lei Zhang, Chunhui Zhou, Shidong Zhou, Xibin Xu National Laboratory for Information Science and Technology, Tsinghua

More information

Improved Detected Data Processing for Decision-Directed Tracking of MIMO Channels

Improved Detected Data Processing for Decision-Directed Tracking of MIMO Channels Improved Detected Data Processing for Decision-Directed Tracking of MIMO Channels Emna Eitel and Joachim Speidel Institute of Telecommunications, University of Stuttgart, Germany Abstract This paper addresses

More information

Constellation Shaping for Communication Channels with Quantized Outputs

Constellation Shaping for Communication Channels with Quantized Outputs Constellation Shaping for Communication Channels with Quantized Outputs Chandana Nannapaneni, Matthew C. Valenti, and Xingyu Xiang Lane Department of Computer Science and Electrical Engineering West Virginia

More information

Lattice Reduction Aided Precoding for Multiuser MIMO using Seysen s Algorithm

Lattice Reduction Aided Precoding for Multiuser MIMO using Seysen s Algorithm Lattice Reduction Aided Precoding for Multiuser MIMO using Seysen s Algorithm HongSun An Student Member IEEE he Graduate School of I & Incheon Korea ahs3179@gmail.com Manar Mohaisen Student Member IEEE

More information

Blind MIMO communication based on Subspace Estimation

Blind MIMO communication based on Subspace Estimation Blind MIMO communication based on Subspace Estimation T. Dahl, S. Silva, N. Christophersen, D. Gesbert T. Dahl, S. Silva, and N. Christophersen are at the Department of Informatics, University of Oslo,

More information

Analysis of Receiver Quantization in Wireless Communication Systems

Analysis of Receiver Quantization in Wireless Communication Systems Analysis of Receiver Quantization in Wireless Communication Systems Theory and Implementation Gareth B. Middleton Committee: Dr. Behnaam Aazhang Dr. Ashutosh Sabharwal Dr. Joseph Cavallaro 18 April 2007

More information

Lecture 12. Block Diagram

Lecture 12. Block Diagram Lecture 12 Goals Be able to encode using a linear block code Be able to decode a linear block code received over a binary symmetric channel or an additive white Gaussian channel XII-1 Block Diagram Data

More information

19. Channel coding: energy-per-bit, continuous-time channels

19. Channel coding: energy-per-bit, continuous-time channels 9. Channel coding: energy-per-bit, continuous-time channels 9. Energy per bit Consider the additive Gaussian noise channel: Y i = X i + Z i, Z i N ( 0, ). (9.) In the last lecture, we analyzed the maximum

More information

EE6604 Personal & Mobile Communications. Week 13. Multi-antenna Techniques

EE6604 Personal & Mobile Communications. Week 13. Multi-antenna Techniques EE6604 Personal & Mobile Communications Week 13 Multi-antenna Techniques 1 Diversity Methods Diversity combats fading by providing the receiver with multiple uncorrelated replicas of the same information

More information

A FEASIBILITY STUDY OF PARTICLE FILTERS FOR MOBILE STATION RECEIVERS. Michael Lunglmayr, Martin Krueger, Mario Huemer

A FEASIBILITY STUDY OF PARTICLE FILTERS FOR MOBILE STATION RECEIVERS. Michael Lunglmayr, Martin Krueger, Mario Huemer A FEASIBILITY STUDY OF PARTICLE FILTERS FOR MOBILE STATION RECEIVERS Michael Lunglmayr, Martin Krueger, Mario Huemer Michael Lunglmayr and Martin Krueger are with Infineon Technologies AG, Munich email:

More information

ADVERSARY ANALYSIS OF COCKROACH NETWORK UNDER RAYLEIGH FADING CHANNEL: PROBABILITY OF ERROR AND ADVERSARY DETECTION. A Dissertation by.

ADVERSARY ANALYSIS OF COCKROACH NETWORK UNDER RAYLEIGH FADING CHANNEL: PROBABILITY OF ERROR AND ADVERSARY DETECTION. A Dissertation by. ADVERSARY ANALYSIS OF COCKROACH NETWORK UNDER RAYLEIGH FADING CHANNEL: PROBABILITY OF ERROR AND ADVERSARY DETECTION A Dissertation by Tze Chien Wong Master of Science, Wichita State University, 2008 Bachelor

More information

MIMO Broadcast Channels with Finite Rate Feedback

MIMO Broadcast Channels with Finite Rate Feedback IO Broadcast Channels with Finite Rate Feedbac Nihar Jindal, ember, IEEE Abstract ultiple transmit antennas in a downlin channel can provide tremendous capacity ie multiplexing gains, even when receivers

More information

Neural networks: Associative memory

Neural networks: Associative memory Neural networs: Associative memory Prof. Sven Lončarić sven.loncaric@fer.hr http://www.fer.hr/ipg 1 Overview of topics l Introduction l Associative memories l Correlation matrix as an associative memory

More information

The Optimality of Beamforming: A Unified View

The Optimality of Beamforming: A Unified View The Optimality of Beamforming: A Unified View Sudhir Srinivasa and Syed Ali Jafar Electrical Engineering and Computer Science University of California Irvine, Irvine, CA 92697-2625 Email: sudhirs@uciedu,

More information

New Recursive-Least-Squares Algorithms for Nonlinear Active Control of Sound and Vibration Using Neural Networks

New Recursive-Least-Squares Algorithms for Nonlinear Active Control of Sound and Vibration Using Neural Networks IEEE TRANSACTIONS ON NEURAL NETWORKS, VOL. 12, NO. 1, JANUARY 2001 135 New Recursive-Least-Squares Algorithms for Nonlinear Active Control of Sound and Vibration Using Neural Networks Martin Bouchard,

More information

Introduction to Wireless & Mobile Systems. Chapter 4. Channel Coding and Error Control Cengage Learning Engineering. All Rights Reserved.

Introduction to Wireless & Mobile Systems. Chapter 4. Channel Coding and Error Control Cengage Learning Engineering. All Rights Reserved. Introduction to Wireless & Mobile Systems Chapter 4 Channel Coding and Error Control 1 Outline Introduction Block Codes Cyclic Codes CRC (Cyclic Redundancy Check) Convolutional Codes Interleaving Information

More information

Neural networks. Chapter 20. Chapter 20 1

Neural networks. Chapter 20. Chapter 20 1 Neural networks Chapter 20 Chapter 20 1 Outline Brains Neural networks Perceptrons Multilayer networks Applications of neural networks Chapter 20 2 Brains 10 11 neurons of > 20 types, 10 14 synapses, 1ms

More information

On the Optimality of Multiuser Zero-Forcing Precoding in MIMO Broadcast Channels

On the Optimality of Multiuser Zero-Forcing Precoding in MIMO Broadcast Channels On the Optimality of Multiuser Zero-Forcing Precoding in MIMO Broadcast Channels Saeed Kaviani and Witold A. Krzymień University of Alberta / TRLabs, Edmonton, Alberta, Canada T6G 2V4 E-mail: {saeed,wa}@ece.ualberta.ca

More information

Least Squares Estimation Namrata Vaswani,

Least Squares Estimation Namrata Vaswani, Least Squares Estimation Namrata Vaswani, namrata@iastate.edu Least Squares Estimation 1 Recall: Geometric Intuition for Least Squares Minimize J(x) = y Hx 2 Solution satisfies: H T H ˆx = H T y, i.e.

More information

Principles of Communications

Principles of Communications Principles of Communications Chapter V: Representation and Transmission of Baseband Digital Signal Yongchao Wang Email: ychwang@mail.xidian.edu.cn Xidian University State Key Lab. on ISN November 18, 2012

More information

Part 8: Neural Networks

Part 8: Neural Networks METU Informatics Institute Min720 Pattern Classification ith Bio-Medical Applications Part 8: Neural Netors - INTRODUCTION: BIOLOGICAL VS. ARTIFICIAL Biological Neural Netors A Neuron: - A nerve cell as

More information

Adaptive Space-Time Shift Keying Based Multiple-Input Multiple-Output Systems

Adaptive Space-Time Shift Keying Based Multiple-Input Multiple-Output Systems ACSTSK Adaptive Space-Time Shift Keying Based Multiple-Input Multiple-Output Systems Professor Sheng Chen Electronics and Computer Science University of Southampton Southampton SO7 BJ, UK E-mail: sqc@ecs.soton.ac.uk

More information

Learning Vector Quantization (LVQ)

Learning Vector Quantization (LVQ) Learning Vector Quantization (LVQ) Introduction to Neural Computation : Guest Lecture 2 John A. Bullinaria, 2007 1. The SOM Architecture and Algorithm 2. What is Vector Quantization? 3. The Encoder-Decoder

More information

Statistical Machine Learning from Data

Statistical Machine Learning from Data January 17, 2006 Samy Bengio Statistical Machine Learning from Data 1 Statistical Machine Learning from Data Multi-Layer Perceptrons Samy Bengio IDIAP Research Institute, Martigny, Switzerland, and Ecole

More information

Need for Deep Networks Perceptron. Can only model linear functions. Kernel Machines. Non-linearity provided by kernels

Need for Deep Networks Perceptron. Can only model linear functions. Kernel Machines. Non-linearity provided by kernels Need for Deep Networks Perceptron Can only model linear functions Kernel Machines Non-linearity provided by kernels Need to design appropriate kernels (possibly selecting from a set, i.e. kernel learning)

More information

Data-aided and blind synchronization

Data-aided and blind synchronization PHYDYAS Review Meeting 2009-03-02 Data-aided and blind synchronization Mario Tanda Università di Napoli Federico II Dipartimento di Ingegneria Biomedica, Elettronicae delle Telecomunicazioni Via Claudio

More information

Decision Weighted Adaptive Algorithms with Applications to Wireless Channel Estimation

Decision Weighted Adaptive Algorithms with Applications to Wireless Channel Estimation Decision Weighted Adaptive Algorithms with Applications to Wireless Channel Estimation Shane Martin Haas April 12, 1999 Thesis Defense for the Degree of Master of Science in Electrical Engineering Department

More information

Approximate Best Linear Unbiased Channel Estimation for Frequency Selective Channels with Long Delay Spreads: Robustness to Timing and Carrier Offsets

Approximate Best Linear Unbiased Channel Estimation for Frequency Selective Channels with Long Delay Spreads: Robustness to Timing and Carrier Offsets Approximate Best Linear Unbiased Channel Estimation for Frequency Selective Channels with Long Delay Spreads: Robustness to Timing and Carrier Offsets Serdar Özen,SMNerayanuru, Christopher Pladdy, and

More information

VID3: Sampling and Quantization

VID3: Sampling and Quantization Video Transmission VID3: Sampling and Quantization By Prof. Gregory D. Durgin copyright 2009 all rights reserved Claude E. Shannon (1916-2001) Mathematician and Electrical Engineer Worked for Bell Labs

More information

Communication Engineering Prof. Surendra Prasad Department of Electrical Engineering Indian Institute of Technology, Delhi

Communication Engineering Prof. Surendra Prasad Department of Electrical Engineering Indian Institute of Technology, Delhi Communication Engineering Prof. Surendra Prasad Department of Electrical Engineering Indian Institute of Technology, Delhi Lecture - 41 Pulse Code Modulation (PCM) So, if you remember we have been talking

More information

Sub-Gaussian Model Based LDPC Decoder for SαS Noise Channels

Sub-Gaussian Model Based LDPC Decoder for SαS Noise Channels Sub-Gaussian Model Based LDPC Decoder for SαS Noise Channels Iulian Topor Acoustic Research Laboratory, Tropical Marine Science Institute, National University of Singapore, Singapore 119227. iulian@arl.nus.edu.sg

More information

Effect of number of hidden neurons on learning in large-scale layered neural networks

Effect of number of hidden neurons on learning in large-scale layered neural networks ICROS-SICE International Joint Conference 009 August 18-1, 009, Fukuoka International Congress Center, Japan Effect of on learning in large-scale layered neural networks Katsunari Shibata (Oita Univ.;

More information

Direct-Sequence Spread-Spectrum

Direct-Sequence Spread-Spectrum Chapter 3 Direct-Sequence Spread-Spectrum In this chapter we consider direct-sequence spread-spectrum systems. Unlike frequency-hopping, a direct-sequence signal occupies the entire bandwidth continuously.

More information