Chapter 11
An Evaluation of Methods for the Selection of
Inputs for an Artificial Neural Network Based
River Model
GJ. Bowden . G.c. Dandy· H.R. Maier
11.1
Introduction
Artificial Neural Network (ANN) models are highly flexible function
approximators, wh ich have shown their utility in a broad range of ecological
modelling applications. The rapid emergence of ANN applications in the field of
ecological modelling can be attributed to their advantages over standard statistical
approaches. Such flexibility provides a powerful tool for forecasting and
prediction, however, the large number of parameters that must be selected only
serves to complicate the design process. In most practical circumstances, the
design of an ANN is heavily based on heuristic trial-and-error processes with only
broad rules of thumb to guide along the way.
The main steps in the development of an ANN model inc1ude choice of
performance criteria, division of data, data pre-processing, determination of model
inputs, determination of network architecture, optimisation (training) and model
validation (Maier and Dandy 2000a). One of the most important steps in this
developmental process is the determination of the significant input variables.
Where the potential number of input variables to an ANN is large and litde a
priori knowledge is available to suggest which subset of variables to inc1ude, the
selection process is inherendy difficult. In this paper, the step involving the
determination of model inputs is considered in detail and a number of different
methods are evaluated. As far as possible, all other steps in the ANN modelling
process are held constant so that the various input determination techniques can be
compared.
In the majority of ANN applications, practitioners give little attention to the
task of input selection (Maier and Dandy 2000b). This is largely because ANNs
belong to the c1ass of data driven approaches, whereas conventional statistical
methods are model driven (Chakraborty et al. 1992). In the latter, the model's
structure is determined first by using empirical or analytical approaches, before
An Evaluation of Methods for the Selection of
Inputs for an Artificial Neural Network Based
River Model
GJ. Bowden . G.c. Dandy· H.R. Maier
11.1
Introduction
Artificial Neural Network (ANN) models are highly flexible function
approximators, wh ich have shown their utility in a broad range of ecological
modelling applications. The rapid emergence of ANN applications in the field of
ecological modelling can be attributed to their advantages over standard statistical
approaches. Such flexibility provides a powerful tool for forecasting and
prediction, however, the large number of parameters that must be selected only
serves to complicate the design process. In most practical circumstances, the
design of an ANN is heavily based on heuristic trial-and-error processes with only
broad rules of thumb to guide along the way.
The main steps in the development of an ANN model inc1ude choice of
performance criteria, division of data, data pre-processing, determination of model
inputs, determination of network architecture, optimisation (training) and model
validation (Maier and Dandy 2000a). One of the most important steps in this
developmental process is the determination of the significant input variables.
Where the potential number of input variables to an ANN is large and litde a
priori knowledge is available to suggest which subset of variables to inc1ude, the
selection process is inherendy difficult. In this paper, the step involving the
determination of model inputs is considered in detail and a number of different
methods are evaluated. As far as possible, all other steps in the ANN modelling
process are held constant so that the various input determination techniques can be
compared.
In the majority of ANN applications, practitioners give little attention to the
task of input selection (Maier and Dandy 2000b). This is largely because ANNs
belong to the c1ass of data driven approaches, whereas conventional statistical
methods are model driven (Chakraborty et al. 1992). In the latter, the model's
structure is determined first by using empirical or analytical approaches, before