Adaptive Modeling and Identification of Nonlinear MIMO Channels Using Neural Networks - Adaptive Signal Processing in Wireless Communications - page 141

Table 5.2

Sigmoid
Transforms Parameters Multiplications Additions

Proposed NN 34 10 24 26

MLP 50 10 40 50

The coupling matrix simulated in this example was taken as






1 . 3

H = .

 

03 . 1



In this example, the proposed block-oriented NN scheme was composed of two blocks
of N = 5 neurons each. The total number of neurons is then equal to 10. We have com-
pared this scheme to an MLP structure composed of ten neurons. The erf function has
been taken for the activation function.
The complexity of the algorithms is displayed in Table 5.2.
As can be seen in the table, the number of multiplications and additions in the MLP
structure is almost twice that of the proposed block structure.
We have tested both algorithms for a range of μ values belonging to the interval
[10 -5 10 -2 ] under various initial conditions. The proposed block structure has always
outperformed the MLP.
Figure 5.6 shows the learning curves for the block structure and the MLP for μ =
0.005. The proposed approach shows lower MSE and faster convergence speed.

10 -1

10 -2

10 -3

MLP

10 -4

Proposed NN

10 -5

0 20 40 60

× 5,000 Iterations

80 100

FIgure 5.6 Smoothed MSE curves (an averaging window of 1,000 samples has been used). μ =

0.005.