Fusion of Neural Networks, Fuzzy Systems and Genetic Algorithms: Industrial Applications Fusion of Neural Networks, Fuzzy Systems and Genetic Algorithms: Industrial Applications
by Lakhmi C. Jain; N.M. Martin
CRC Press, CRC Press LLC
ISBN: 0849398045   Pub Date: 11/01/98
  

Previous Table of Contents Next


Results reported in Table 2 confirm the hypothesis that configuration Net0 independently minimizes the cost figure for each parameter outperforming NetA and NetB In the case of Net0, in fact, the weights of each TDNN are tuned independently toward a unique objective in contrast to NetA and NetB where 5 uncorrelated objectives must be met by the same, although larger structure.

Table 2 Performance comparison evaluated for each mouth articulatory parameter for the configurations Net0, NetA, and NetB by applying the training procedure T1 (1051-8215/97$10.00 © 1997 IEEE).
T1 PARAMETER
LM H W
Cost net0 netA netB net0 netA netB net0 netA netB
MSE 2.761 3.665 3.358 3.380 4.691 4.150 3.762 3.460 2.312
MAX 0.863 1.072 0.762 0.783 1.103 0.773 0.653 0.732 0.771
r 0.9435 0.9186 0.9269 0.9386 0.9060 0.9198 0.8586 0.7855 0.7777
T1 PARAMETER
Lup DW Average
Cost net0 netA netB net0 netA netB net0 netA netB
MSE 1.994 3.717 3.687 5.531 9.933 8.969 3.015 4.646 4.495
MAX 0.704 0.911 0.871 1.020 1.078 1.148 0.804 0.972 0.865
r 0.8090 0.5966 0.6031 0.7067 0.4448 0.4387 0.8513 0.7303 0.7330

Table 3 Performance comparison evaluated for each mouth articulatory parameter for the configurations Net0, NetA, and NetB after applying the training procedure T2 (1051-8215/97$10.00 © 1997 IEEE).
T2 PARAMETER
LM H W
Cost net0 netA netB net0 netA netB net0 netA netB
MSE 5.782 6.622 6.148 7.091 8.008 7.309 3.762 3.460 3.700
MAX 0.988 1.006 0.957 1.095 1.066 1.106 1.132 1.071 1.097
r 0.8599 0.8342 0.8489 0.8474 0.8260 0.8420 0.5619 0.5887 0.5655
T2 PARAMETER
Lup DW Average
Cost net0 netA netB net0 netA netB net0 netA netB
MSE 16.18 18.27 18.75 10.93 10.12 10.03 8.569 9.296 9.187
MAX 1.384 1.430 1.490 1.354 1.090 1.210 1.190 1.132 1.172
r 0.3645 0.2937 0.2858 0.2701 0.2928 0.3184 0.5807 0.5671 0.5721

NetB yields average performances higher than NetA confirming that the optimal size for the time-window on which the articulatory estimates are based equals 260 msec. while 380 msec time-windows (19 × 20 msec) are used in the case of NetA.


Previous Table of Contents Next

Copyright © CRC Press LLC