Training epoch

Steps 2 and 3 are performed for all input objects. When all data points have been fed into the network one training epoch has been achieved. A network is usually trained in several training epochs, depending on the size of the network and the number of data points. 

Training is a computer controlled process that once started, it might be stopped when either of the next conditions is fulfilled maximum number of training epochs has elapsed, training and testing errors have reached an acceptable level or there is no more improvement in learning with further iterations. 

Cumulative Delta Rule This is a variant on the delta rule. In the delta rule, the weights are updated after the presentation of each input/output pair of data. In the cumulative delta rule, the weight changes are accumulated and then applied all at one, typically at the end of an integer multiple of the training epoch. Mathematically, we write ... 

Finally, a stopping criterion must be decided. We usually state two criteria to prevent the net from being stacked in a local minimum and save time the number of training epochs and a maximum level of error, given as MSE [eqn (6.8)]. 

LIBS and ANNs were applied to quantitatively evaluate pollution by heavy metals in soils.The ANNs were used to process the spectral lines and a new algorithm (that considered weight iteration within the ANN) was developed to reduce the number of training epochs. As a matter of example, the limits of detection for Cu and Cd were 42 and 5 ppm, respectively. 

In order to determine the training algorithm that can provide the best classification performance, the actions that had been taken consist of 2 main steps namely analysis of number of training epochs and analysis of number of hidden nodes. Here, the optimum number of training epochs and hidden nodes are obtained when the MLP network achieved the highest classification performance. 

At the beginning, when the number of training epochs is equal to zero, the p value extends over the entire network (p = /inet). while at the end of training, when Hep = /itot, P is equal to zero (p = 0). This means that in the last epoch only the weights of the excited neuron are corrected. 

The configuration of the ANN is shown in Figure 10.3. There are two nodes on the input layer corresponding to the amplitudes for the vessel s size (the dead weight (dwt)) and age of typical bulk carriers. After performing a series of experiments on the effects of number of hidden neurons on training epoch, ten nodes are selected on the hidden layer to allow for the nonlinearity of the problem. The output layer has one node corresponding to the amplitude for hull failure rate. Table 10.1 outlines the major neural network characteristics. 

---