shows great potential for processing unstructured data such as text, image, audio, video
and time series. New types of neural network structures such as convolutional neural
network, recurrent neural network, and generative adversarial network have sprung up.
BP neural network is currently the most commonly used neural network. Its
algorithm could be described as follows:
(1) There are N sample data:
X k ; Y k
ð
Þ; X k ¼ ðx 1k ; x 2k ; . . .; x nk Þ; Y k ¼ ðy 1k ; y 2k ; . . .; y mk Þ; k ¼ 1; 2; . . .; N:
(2) There are L layers in neural network, L ! 3:
(3) There are n
l nodes in layer l; l ¼ 1; 2; . . .; L. The input layer has X k
j j ¼ n nodes,
and the output layer has Y k
j j ¼ m nodes.
(4) f ðxÞ indicates the activation function of nodes.
(5) The learning rate and error precision of neural network are denoted as g 2 ð0; 1Þ
and e.
(6) The number of iterations t is initialized as 1.
(7) X k is input into neural network, the output of every node is O
l
jk :
I
l
jk ¼
X
n
ðlÀ1Þ
i¼1
w
ðlÀ1Þ
ij
O
ðlÀ1Þ
ik
O
l
jk ¼ f ðI
l
jk Þ
(8) I
l
jk and O
l
jk are the input and output of node j in layer l, w
l
ij is the weight between
node i in layer l and node j in layer l þ 1.
(9) The output layer error is E jk ; j ¼ 1; 2; . . .; m
E jk ¼
1
2
y jk À O
L
jk
2
(10) The weights are modified as follows:
w
ðlÀ1Þ
ij
h
i à ¼ w
ðlÀ1Þ
ij
À g
@E
l
jk
@w
ðlÀ1Þ
ij
(11) k ¼ k þ 1, the iteration number t ¼ t þ 1, iterations continue until the output layer
error or iteration number meets the requirements.
2 Current Research on Performance Evaluation of Satellite
Communication System
Performance evaluation of satellite communication system could be divided into three
categories: analytical method, software simulation method and index system method.
240
S. Cheng et al.