224
G.J. Bowden . G.c. Dandy . H.R. Maier
sampies for compiling training, testing and validation sets that are statistically
similar.
When computationally intensive supervised input determination
techniques such as the GA-ANN are used, compact training, testing and validation
sets are advantageous as they increase processing speed. Since the SOM data
division technique allows training, testing and validation sets to be selected that
are statistically representative of the same population, a fair comparison of the
input selection techniques can be made whilst providing the most rigorous test of
each method.
A full description of the use of the SOM for dividing data into statistically
similar subsets is provided in Bowden et al. (2000).
11.4.3
Determination of Model Inputs
Six ANN models were developed, each using a different combination of
unsupervised and supervised input determination methods (Figure 11.1). After
applying the unsupervised techniques to the initial data set, the inputs were
reduced to the subsets displayed in Table 11.1.
The supervised input
determination techniques were performed using the subsets shown in Table 11.1.
11.5
Results and Discussion
Details of each of the models developed are shown in Table 11.2. It can be seen
that a variety of different inputs have been selected in each of the models. In
general, the models that utilise a GA-ANN for the supervised input determination
included a wider variety of input variables than the models developed using the
stepwise modelling procedure. The stepwise modelling procedure tended to
produce more compact models since the process is terminated once the addition of
any extra variables fails to improve the model performance.
GAs are stochastic processes and as such repetition of experimental treatments
was performed to evaluate the effect of stochasticity on the set of inputs identified
for each GA-ANN model. It was found that by performing multiple runs for each
GA-ANN, there were only slight differences in the optimal subset of inputs
identified. However, most importantly, it was found that each of the models
identified by the GA-ANN consistently outperformed the models developed by the
stepwise ANN modelling procedure when measured on the validation set.
G.J. Bowden . G.c. Dandy . H.R. Maier
sampies for compiling training, testing and validation sets that are statistically
similar.
When computationally intensive supervised input determination
techniques such as the GA-ANN are used, compact training, testing and validation
sets are advantageous as they increase processing speed. Since the SOM data
division technique allows training, testing and validation sets to be selected that
are statistically representative of the same population, a fair comparison of the
input selection techniques can be made whilst providing the most rigorous test of
each method.
A full description of the use of the SOM for dividing data into statistically
similar subsets is provided in Bowden et al. (2000).
11.4.3
Determination of Model Inputs
Six ANN models were developed, each using a different combination of
unsupervised and supervised input determination methods (Figure 11.1). After
applying the unsupervised techniques to the initial data set, the inputs were
reduced to the subsets displayed in Table 11.1.
The supervised input
determination techniques were performed using the subsets shown in Table 11.1.
11.5
Results and Discussion
Details of each of the models developed are shown in Table 11.2. It can be seen
that a variety of different inputs have been selected in each of the models. In
general, the models that utilise a GA-ANN for the supervised input determination
included a wider variety of input variables than the models developed using the
stepwise modelling procedure. The stepwise modelling procedure tended to
produce more compact models since the process is terminated once the addition of
any extra variables fails to improve the model performance.
GAs are stochastic processes and as such repetition of experimental treatments
was performed to evaluate the effect of stochasticity on the set of inputs identified
for each GA-ANN model. It was found that by performing multiple runs for each
GA-ANN, there were only slight differences in the optimal subset of inputs
identified. However, most importantly, it was found that each of the models
identified by the GA-ANN consistently outperformed the models developed by the
stepwise ANN modelling procedure when measured on the validation set.