International Journal of Advance Research and Innovation

Article Info 
Article history: 
Received 20 July 2014 
Received in revised form 
10 August 2014 
Accepted 20 August 2014 
Available online 15 September 2014 
Keywords 
Surface Roughness,
Turning,
Artificial Neural Network,
Parametric Analysis

---

Volume 2, Issue 3 (2014) 676-683 
ISSN 2347 - 3258
International Journal of Advance Research and Innovation 
677 
IJARI 
a set of nodes (units) connected together via directed links. 
Each node in the network has a numeric activation level 
associated with it. The overall pattern vector of activation 
represents the current state of the network. Activation rule is 
a local procedure that each node follows in updating its 
activation level in the context of input from neighbouring 
nodes. Learning rule is a local procedure that describes how 
the weights on connections should be altered as a function 
of time. Types of activation functions include: threshold 
function; Piecewise-linear function, and sigmoid function. 
The sigmoid function, whose graph is s-shaped graph, is by 
far the most common form of activation function used in the 
construction of neural networks. 
A simple process element of the artificial neural 
network has three layers; the input, hidden, and output 
layers. The input and output layers are defined as nodes, and 
the hidden layer provides a relation between the input and 
output layers. Initially, the weights of the nodes are random 
and the network has not any knowledge. For a given input 
pattern, the network produces an associated output pattern. 
Its learning and update procedure is based on a relatively 
simple concept: the network is provided with both a set of 
patterns to be learned and the desired system response for 
each pattern. If the network generates the wrong answer, 
then the weights are updated to be less error. Finally, future 
responses of the network are more likely to be correct. 
Regarding the number of layers, the only certainty is that 
there should be an input and an output layer so as to be able 
to present and obtain data to and from the ANN, 
respectively. The number of neurons in each of these two 
layers is specified by the number of input and output 
parameters that are used to model each problem so it is 
readily determined. Therefore, the objective is to find the 
number of hidden layers and the number of neurons in each 
hidden layer. Unfortunately, it is not possible to 
theoretically determine how many hidden layers or neurons 
are needed for each problem. The activation functions are 
chosen based on the kind of data that are available (binary, 
bipolar, decimal, etc.) and the type of layer. For instance, 
the identity function is almost always used in the input 
layer, while continuous non-linear sigmoid functions are 
used in the hidden layers (usually the hyperbolic tangent 
function). The training algorithm influences to a far greater 
extent the training speed, performance of an ANN (training 
error) or the necessary computing power rather than the 
architecture itself. 
 
Fig: 1. The Mathematical Model of Neuron 
The basic rules are that neurons are added when 
training is slow or when the mean squared error is larger 
than a specified value, and that neurons are removed when a 
change in a neuron‟s value does not correspond to a change 
in the network‟s response or when the weight values that are 
associated with this neuron remain constant for a large 
number of training epochs. 
Learning is the process by which the free parameters of 
a neural network get adapted through a process of 
stimulation by the environment in which the network is 
embedded. The type of learning is determined by the 
manner in which the parameter changes take place. The set 
of well-defined rules of the solution of a learning problem is 
called a learning algorithm. Each learning algorithm differ 
from the other in the way in which the adjustment to a 
synaptic weight of a neuron is formulated. Also, the 
manner in which a neural network is made up of inter-
connected neurons relating to its environment, is also to be 
considered. There are various learning rules. Hebb‟s 
learning rule is the oldest and most famous of all learning 
rules. It states that, “when an axon of cell A is near enough 
to excite a cell B and repeatedly or persistently takes part in 
firing it, some growth process or metabolic changes take 
place in one or both cells such that A‟s efficiency as one of 
the cells firing B, is increased”. This learning can also be 
called correlational learning. This statement may be split 
into a two-part rule: 1. If two neurons on either side of a 
synapse are activated simultaneously, then the strength of 
that synapse is selectively increased. 2. If two neurons on 
either side of a synapse are activated simultaneously, then 
that synapse is selectively weakened or eliminated. This 
type of synapse is called hebbian synapse. The four key 
mechanisms that characterize a hebbian synapse are 
dependent mechanism, local mechanism, interactive 
mechanism and correlational mechanism. 
For the perception learning rule, the learning rule, the 
learning signal is the difference between the desired and 
actual neuron‟s response. This type of learning is 
supervised. The fact that the weight vector is perpendicular 
to the plane separating the input patterns during the learning 
processes, can be used to interpret the degree of difficulty of 
training a perceptron for different types of input. There is a 
perceptron learning rule convergence theorem which states, 
`` if there is a weight vector w* such that f (x (p) w*) = t (p) 
for all p, then for any starting vector w
1
the perceptron 
learning rule will converge to a weight vector that gives the 
correct response for all training patterns, and this will be 
done in a finite number of steps”.