Design File1

Settings for DNNs of route 2 (Q and V trained separately)...
1. Q's DNN:
20 3x3 kernels, unit padding, 1x2 avg pooling, BN. 40 3x3 kernels, 1x2 avg pooling, BN. FC1 240 neurons, FC2 120 neurons, FC3 50 neurons.

2. V's DNN:
20 3x3 kernels, unit padding, 1x2 avg pooling, BN. 40 3x3 kernels, 1x2 avg pooling, BN. FC1 240 neurons, FC2 160 neurons, FC3 100 neurons, FC4 50 neurons. 




Hyperparameters for DNNs of route 2 (Q and V trained separately)...

1. Q's DNN:
number of epochs = 300, training batch size = 64, test batch size = 100, stochastic gradient descent (SGD) is selected for weight backpropagation with learning rate = 0.01 and momentum = 0.5, and L2 regularization is not added. Shuffle is done at each epoch to increase the randomness of the training data for better generalization capacity [41]. Momentum is used in the SGD optimizer to avoid gradient plateaus or non-optimal solutions [41, 43]. Finally, an adaptive (i.e. decaying) learning rate was employed to gain robustness against gradient noise and generate a smoother convergence [41, 44].

2. V's DNN:
number of epochs = 300, training batch size = 64, test batch size = 100, stochastic gradient descent (SGD) is selected for weight backpropagation with learning rate = 0.01 and momentum = 0.9, and L2 regularization factor = 0.0005 is added. Shuffle is done at each epoch to increase the randomness of the training data for better generalization capacity [41]. Momentum is used in the SGD optimizer to avoid gradient plateaus or non-optimal solutions [41, 43]. L2 regularization is used in the backpropagation to reduce model overfitting [41]. Finally, an adaptive (i.e. decaying) learning rate was employed to gain robustness against gradient noise and generate a smoother convergence [41, 44].

References are included in the main text.