Development of Integrated Choice and Latent Variable (ICLV) Models Using Matrix-Based Analytic Approximation and Automatic Differentiation Methods on TensorFlow Platform
Table 1
BVN and TVN evaluation results for the different methods.
Dim
Measure
Numerical integration methods
Analytic approximation methods (TensorFlow)
Gauss
TF
ME
BME
TVBS
K = 2 (BVN)
MAE
4.8∗e − 10
1.9∗e − 15
0.0037
—
—
%MAE > 0.005
0.0
0.0
18.8
—
—
MAPE
8.6∗e − 8
1.3∗e − 12
1.28
—
—
%MAPE > 2
0.0
0.0
15.2
—
—
Time (s)
0.001
0.002
0.003
—
—
K = 3 (TVN)
MAE
8.8∗e − 10
6.1∗e − 11
0.0027
0.0018
—
%MAE > 0.005
0.0
0.0
18.1
12.4
—
MAPE
1.9∗e − 7
2.9∗e − 8
1.34
1.00
—
%MAPE > 2
0.0
0.0
16.7
11.0
—
Time (s)
0.005
0.005
0.006
0.003
—
Note. “—“ indicates that the corresponding analytic approximation methods cannot calculate the values of BVN or TVN.