32 Prediction of Width and Thickness of Injection Molded …
463
32.4.4 Multilayer Perceptron
The last algorithm used in this experiment is the Multilayer Perceptron (MLP)
ANN. It is one of the classic ANN models and is based on the sequence of layers
of neurons interconnected between each other with weights. Every time a new
sample of data is given to the network, those weights are adjusted accordingly. In
addition to that, “layer-to-layer mapping is activated with a non-linear function”
(Ogorodnyk et al. 2018). In this study, the sigmoid function is used as an activation
function and the number of neurons in the hidden layer of the network is calculated
as (number_of_parameters + number_of_output_neurons)/2. In this experiment, the
MP was built using a different number of parameters as suggested by different feature
selection algorithms (7, 8, 18 and 25).
32.5 Results
Predictive models for the thickness and width of the injection molded dogbone specimens have been built using different machine learning algorithms presented in the
previous section. Depending on the algorithm, parameter configurations were varied
to see how they affect the quality of the resulting model. Figure 32.2 shows the
dependence of the correlation coefficient of the measured and predicted values of the
part’s dimensions and different algorithm parameters that were tested. The quality
of the models was estimated using the 10-folds cross-validation procedure.
Table 32.2 shows the performance of tested algorithms using the correlation coefficient, mean absolute error (MAR) and root square mean error (RSME). The numbers
are shown for the best algorithm parameter configurations. For the REPTree the best
minimum number of instances in a leaf for thickness1, thickness2, and width2 is
equal to 2, while for width1 it is 5. In the case of the random forest algorithm, the
number of trees that gives better results for thickness1 and width1 is 100, for thickness2 and width2 it is 200. At the same time for the kNN method the best numbers
of neighbors are 5 and 1, the first number shows better results for the thickness1
and width1, while the second number for thickness2 and width2. For the MLP, on
the other hand, 8 features have shown the best result for the width2 model, 25 for
thickness2 and 18 for thickness1 and width1 models.
32.5.1 Interpretation
The only algorithm out of those used in this study that is easy to interpret for a
human being is REPTree decision tree algorithm. Figure 32.3 shows the structure of
a decision tree with the minimum number of instances in a leaf equal to 2 that has
shown the best results for prediction of the width of dogbone 1.
463
32.4.4 Multilayer Perceptron
The last algorithm used in this experiment is the Multilayer Perceptron (MLP)
ANN. It is one of the classic ANN models and is based on the sequence of layers
of neurons interconnected between each other with weights. Every time a new
sample of data is given to the network, those weights are adjusted accordingly. In
addition to that, “layer-to-layer mapping is activated with a non-linear function”
(Ogorodnyk et al. 2018). In this study, the sigmoid function is used as an activation
function and the number of neurons in the hidden layer of the network is calculated
as (number_of_parameters + number_of_output_neurons)/2. In this experiment, the
MP was built using a different number of parameters as suggested by different feature
selection algorithms (7, 8, 18 and 25).
32.5 Results
Predictive models for the thickness and width of the injection molded dogbone specimens have been built using different machine learning algorithms presented in the
previous section. Depending on the algorithm, parameter configurations were varied
to see how they affect the quality of the resulting model. Figure 32.2 shows the
dependence of the correlation coefficient of the measured and predicted values of the
part’s dimensions and different algorithm parameters that were tested. The quality
of the models was estimated using the 10-folds cross-validation procedure.
Table 32.2 shows the performance of tested algorithms using the correlation coefficient, mean absolute error (MAR) and root square mean error (RSME). The numbers
are shown for the best algorithm parameter configurations. For the REPTree the best
minimum number of instances in a leaf for thickness1, thickness2, and width2 is
equal to 2, while for width1 it is 5. In the case of the random forest algorithm, the
number of trees that gives better results for thickness1 and width1 is 100, for thickness2 and width2 it is 200. At the same time for the kNN method the best numbers
of neighbors are 5 and 1, the first number shows better results for the thickness1
and width1, while the second number for thickness2 and width2. For the MLP, on
the other hand, 8 features have shown the best result for the width2 model, 25 for
thickness2 and 18 for thickness1 and width1 models.
32.5.1 Interpretation
The only algorithm out of those used in this study that is easy to interpret for a
human being is REPTree decision tree algorithm. Figure 32.3 shows the structure of
a decision tree with the minimum number of instances in a leaf equal to 2 that has
shown the best results for prediction of the width of dogbone 1.