inputs} = {network outputs}. Auto-associative networks are one of the classic Artificial Neural Network (ANN) architectures used
commonly in robotics, machine learning, and signal processing. They have been used for a wide variety of pattern processing problems
such as cleaning up noisy pictures and recognizing known pictures when partially occluded1. Some of the known applications that
auto-associative networks are typically used in are: noise reduction, replacement of missing sensor values, gross error detection and
correction, and signal processing.
The purpose of training a highly-parameterized, nonlinear network in these areas is that feed-forward networks trained on the identity
function can perform several useful data screening tasks with appropriate internal architectures2. In other words, this particular type of
network is trained to reproduce its inputs and its output(s). The network is forced to represent the input patterns in fewer dimensions,
creating a compressed representation. These compressed representations may reveal interesting generalizations about the data. A typical
architecture of an auto-associative network contains 3 hidden layers, which are, respectively, called mapping layer, bottle neck layer,
and de-mapping layer3. This approach has been used by some researchers 4, 5, 6 to reduce the dimensionality (# of nodes) of the hidden
layer in ANNs for the commonly used applications listed above. The auto-associative network approach has been used in some
engineering areas for about two decades. However, it has not been explored in civil engineering, where artificial intelligence is mostly
referred to as a function approximation method. In this study, the auto-associative network approach was explored by using seven
engineering databases that contain both categorical and continuous variables. For this reason, model development of the autoassociative
network was considered with only one hidden layer. More than one hidden layer combined with an insufficient number of
databases may cause the network to memorize the data in the training phase7. Consequently, models were developed with only one
hidden layer to maintain the generalization capability of the network.
Auto-associative network is based on mapping n input variables into n output variables. In order to obtain predictions from this network,
an initial estimate of the controlled variable (i.e. output) has to be included as an input. For this reason, each database was utilized to
develop an ANN prediction as an input in the model development and then applied to auto-associative network approach. The four
sequential training stages for all seven databases and their desired criteria to choose the optimal network structures of auto-associative
network models are explained in the following sections. Even though the developed models are optimized on both inputs and output,
in this study only the output variable was evaluated in terms of statistical accuracy measures. Therefore, results presented in this study
are limited to output variables.