“Recurrent neural networks (RNN) attract considerable interest in computational intelligence because of their superior power in processing spatio-temporal data and time- varying signals. In this research, we analyze how the architecture and training algorithm affect the dynamic properties of RNN. Proofs are presented to show that different transfer functions and biases can influence the system convergence to desired outputs in certain models. We demonstrate the advantages and limitations of gradient-based training algorithms, Recurrent Back Propagation (RBP) and Back Propagation through Time (BPTT). Evolutionary Algorithm (EA) is discussed for its global search process and non-gradient adaptation. A cellular Simultaneous Recurrent Network (SRN) computer Go player design illustrates the unique ability of function approximation for dynamic programming"--Abstract, page iii.
Wunsch, Donald C.
Pottinger, Hardy J., 1944-
St. Clair, Daniel C.
Electrical and Computer Engineering
M.S. in Computer Engineering
University of Missouri--Rolla
viii, 38 pages
© 2002 Xindi Cai, All rights reserved.
Thesis - Restricted Access
Neural networks (Computer science)
Print OCLC #
Link to Catalog Record
Electronic access to the full-text of this document is restricted to Missouri S&T users. Otherwise, request this publication directly from Missouri S&T Library or contact your local library.http://merlin.lib.umsystem.edu/record=b4820017~S5
Cai, Xindi, "Recurrent neural networks: Architecture, algorithms and applications" (2002). Masters Theses. 2136.
Share My Thesis If you are the author of this work and would like to grant permission to make it openly accessible to all, please click the button above.