Selecting the Parameters' Values

Suppose that we have $P$ parameters to be taken into account, namely ( ${\cal P} = \{\pi_1,\pi_2,\cdots,\pi_P\}$). Denote by $n_p$ the number of different selected values of the parameter $\pi_p$. If we choose the whole set of configurations of the $P$ parameters, the total number of samples may be very large ( $n_1 \times n_2 \times \cdots \times n_P$). Carrying out subjective quality tests for such a number of samples may be impractical. Fortunately, the NN does not need all these samples to capture the relationship between the quality and the parameters' values, it is its goal to predict the quality in the missed configurations ⁴¹. Therefore, we propose the following method to generate the minimum number of samples needed to train and test the neural networks. We should define default values for all the parameters, ${\cal P_{\circ}} = \{\pi_{01},\pi_{02},\cdots,\pi_{0P}\}$). These values could be, for instance, the most frequently observed values. We change the values of two parameters at a time and give the default values to the others. We repeat this step for all the parameters. In this way, a list of configurations is defined, containing many duplications. Once these duplicates are removed, the remaining ones constitute the minimum samples required to train and test the NN.