230
G.J. Bowden . G.c. Dandy . H.R. Maier
representative of the same population, despite the use of the SOM data division
technique. Due to the nature of the Anabaena spp. data, it is difficult to compile
training, testing and validation sets that are statistically similar.
The results in Table 11.3 also confirm that the GA-ANN performed better than
the stepwise ANN procedure when measured on the validation set. However, the
GA-ANN and stepwise ANN procedure produced very similar results when
measured on the test set. Table 11.2 shows that each of the six models developed
contained different network architectures. The network architecture of each
model was optimised using the NGO and was found to be very dependent on the
input subset used. The GA-ANN technique usually identified more input
variables than the stepwise ANN procedure, resulting in larger models.
11.6
Conclusions
The results of this study show that the combination of apriori knowledge and a
hybrid GA-ANN provided the most effective means of identifying the significant
input variables. The resulting model (modell) was able to forecast the onset and
duration of the major incidence of Anabaena spp. in the validation set with a good
level of accuracy. The inputs found to be important in this model include nutrient
levels (total phosphorus, soluble phosphorus and TKN) as weIl as turbidity,
colour, temperature, flow, pH and previous concentrations of Anabaena (Table
11.2).
Of the analytical unsupervised techniques, the PCA ANN models and the SOM
ANN models had identical performance when measured on the independent
validation set. Both unsupervised methods provide a suitable means of reducing
input dimensionality.
The models (1, 3 and 5) wh ich utilised the GA-ANN as the supervised input
determination outperformed the models developed using the stepwise ANN
modelling procedure. The GA-ANN had the ability to efficiently evaluate a very
large number of networks and consequently, it was able to find synergistic
combinations of inputs resulting in superior forecasts for the validation set.
Acknowledgements
Financial support for this research was provided by an Australian Postgraduate
Award (Industry) in conjunction with United Utilities Australia Pty Ltd. This
support is gratefully acknowledged. The authors would like to thank Mike Burch
of the Australian Water Quality Centre for the many helpful discussions, which
have been of great benefit to the research.
G.J. Bowden . G.c. Dandy . H.R. Maier
representative of the same population, despite the use of the SOM data division
technique. Due to the nature of the Anabaena spp. data, it is difficult to compile
training, testing and validation sets that are statistically similar.
The results in Table 11.3 also confirm that the GA-ANN performed better than
the stepwise ANN procedure when measured on the validation set. However, the
GA-ANN and stepwise ANN procedure produced very similar results when
measured on the test set. Table 11.2 shows that each of the six models developed
contained different network architectures. The network architecture of each
model was optimised using the NGO and was found to be very dependent on the
input subset used. The GA-ANN technique usually identified more input
variables than the stepwise ANN procedure, resulting in larger models.
11.6
Conclusions
The results of this study show that the combination of apriori knowledge and a
hybrid GA-ANN provided the most effective means of identifying the significant
input variables. The resulting model (modell) was able to forecast the onset and
duration of the major incidence of Anabaena spp. in the validation set with a good
level of accuracy. The inputs found to be important in this model include nutrient
levels (total phosphorus, soluble phosphorus and TKN) as weIl as turbidity,
colour, temperature, flow, pH and previous concentrations of Anabaena (Table
11.2).
Of the analytical unsupervised techniques, the PCA ANN models and the SOM
ANN models had identical performance when measured on the independent
validation set. Both unsupervised methods provide a suitable means of reducing
input dimensionality.
The models (1, 3 and 5) wh ich utilised the GA-ANN as the supervised input
determination outperformed the models developed using the stepwise ANN
modelling procedure. The GA-ANN had the ability to efficiently evaluate a very
large number of networks and consequently, it was able to find synergistic
combinations of inputs resulting in superior forecasts for the validation set.
Acknowledgements
Financial support for this research was provided by an Australian Postgraduate
Award (Industry) in conjunction with United Utilities Australia Pty Ltd. This
support is gratefully acknowledged. The authors would like to thank Mike Burch
of the Australian Water Quality Centre for the many helpful discussions, which
have been of great benefit to the research.