What are the design issues in an Artificial Neural Network?

An artificial neural network is a system based on the functions of biological neural networks. It is a simulation of a biological neural system. The feature of artificial neural networks is that there are several structures, which required several methods of algorithms, but regardless of being a complex system, a neural network is easy.

These networks are between the specific signal-processing sciences in the director’s toolbox. The area is hugely interdisciplinary, but this approach will restrict the view to the engineering perspective.

In engineering, neural networks produce two essential functions as pattern classifiers and as non-linear adaptive filters. An artificial neural network is dynamic, it provides a non-linear system that learns to execute a function (an input/output map) from data. Adaptive represents that the system parameters are changed during operation, frequently called the training phase.

After the training phase, the artificial neural network parameters are fixed and the system is start to solve the problem at hand (the testing phase). The artificial neural network is produced with a systematic step-by-step procedure to improve a performance test or to follow some definite internal constraint, which is usually described as the learning rule.

There are the following design issues that must be considered which are as follows −

The several nodes in the input layer must be decided. It can be created an input node for each mathematical or binary input variable. If the input variable is categorical, it can create one node for each categorical value or encrypt the k-ary variable using [log2k] input nodes.

The multiple nodes in the output layer must be created. For a two-class problem, it is adequate to need a single output node. For a k-class problem, there are k output nodes.

The network topology such as the number of hidden layers and hidden nodes, and feed-forward or recurrent network structure should be selected. The target function description is based on the weights of the connection, the multiple hidden nodes and hidden layers, biases in the nodes, and the type of activation function. It is discovering the right topology is not a simple task.

One method is to start from a fully connected network with an adequately huge number of nodes and hidden layers, and then repeat the model-building structure with a smaller number of nodes. This method can be moderate. Alternatively, instead of repeating the model-building structure, it can delete several nodes and repeat the model evaluation process to choose the right model complexity.

The weights and biases should be initialized. Random assignments are generally adequate.

Training instances with missing values must be deleted or restored with the most likely values.