2 RNN Network Structure
Recurrent neural network is a kind of network model for processing sequential data,
which is widely used in text processing, speech recognition and image tracking. The
network design pattern adopted in this paper is the recurrent neural network expansion
about time, which has a single output at the end of the sequence, as shown in Fig. 1.
The basic RNN network structure includes input layer, hidden layer and output
layer. The parameters sharing mechanism of the recurrent neural network makes the
calculation graph of the network can be extended continuously and adapt to longer
sequence samples of different lengths. Figure 1 shows the RNN network model
adopted in this paper.
In Fig. 1, the right network figure shows the recurrent neural network expansion
about time, s is timestep. L
ðtÞ is the value of loss at time t, x
ðtÞ
¼ ðx
ðtÞ
1 ; x
ðtÞ
2 ; . . .; x
ðtÞ
k Þ is
the input vector at time t, which has a size of 1 Â k. j is the number of cells in hidden
layer, which has a full connect with input layer. The output vector is
h
ðtÞ
¼ ðh
ðtÞ
1 ; h
ðtÞ
2 ; . . .; h
ðtÞ
j Þ, which has a size of 1 Â j. The output layer has the same
dimension as the input layer, outputs a vector o
ðtÞ
¼ ðo
ðtÞ
1 ; o
ðtÞ
2 ; . . .; o
ðtÞ
k Þ. The previous
hidden layer connects with the hidden layer at next time point by matrix W and forms a
recurrent network.
Network propagation model is the mathematical representation of network structure, including forward propagation model and back propagation model.
Forward propagation model as follow:
h
ðtÞ
¼ tanhðW Á h
ðtÀ1Þ
þ U Á x
ðtÞ
þ b i Þ
ð 2:1Þ
o
ðtÞ
¼ V Á h
ðtÞ
þ b o
ð2:2Þ
U
W
x
h
o
L
y
( 1)
t
x
−
( 1)
t
h
−
( )
x
τ
( )
t
h
( )
t
o
( )
t
y
( )
t
L
( )
h
L
( )
h
L
W
W
W
U
U
V
Expand
V
1
2
j
k
k
1
1
1
k
( )
x
L
U
Fig. 1. RNN model
Maritime Target Trajectory Prediction Model Based on the RNN …
335