Learning, Observing and Outputting

Learning in the PDP system consists of building up associations so that particular input patterns will elicit specific output results. At all times connections exist between the input and output (albeit these connections are frequently indirect). As learning takes place, the relevant connections become stronger and the irrelevant ones become weaker. Typically, a very large number of input patterns must be observed by the system for learning to occur. 

Analyze, test, and revise the model. This task, analyzing a model and learning from it, should be the most time consuming and demanding one. We have to make sure that the model is implemented correctly, observe model output, compare it to data, and test how changes in the model assumptions affect the model s behavior. Finally, we can also try to deduce new predictions for validation Does the model predict phenomena or patterns that we did not know and use in some way for model development and calibration ... 

Two generally different scenarios can be found for applications of machine learning technology so-called supervised and unsupervised learning. The difference is the presence or absence of observation of the desired output on a training data set. 

In unsupervised learning, some data is given and the cost function to be minimized, which can be any fiinction of the data and the network s output,/. The cost function is dependent on the task (what we are trying to mtxlel) and our a priori assumptions (the implicit properties of our mtxlel, its parameters, and the observed variables). 

We consider the continuous output value case first. PMs for these problems are usually a function of a residual or error term (O, - D,), where O, is the observed value of an output PE for the ith case and D, is the desired value of the same output PE for the rth case. Observed values are actual output PE values generated by the ANN. Desired values are correct values (i.e., values the ANN is trying to learn) and are sometimes called target values. Be warned that different authors use the same terminology to mean different things some use observed... 

Three predictive models, namely BP-ANN, GA-ANN and AGA-ANN, were developed in MATLAB. Network parameters, i.e. adaptive learning rate for BP-ANN, crossover and mutation probability for GA-ANN, adaptive crossover site and mutation probability for proposed AGA-ANN, were chosen optimally. For each input pattern, the predicted value of the output variable (the ratio of damage area to hole area) has been compared with the respective experimental value for different networks in Figure 6.12a. The network was tested with the 20 test data points that had not been used for the purpose of training. From the test results, it was observed that the predicted values are very close and follow almost the same trend as the experimental values for AGA-ANN... 

Table 21.4 shows the topology and the results for the five best neural networks implemented. To evaluate the accuracy of the ANN models developed, we have used the RMSEs in validation (RMSE ). As it can be observed in Table 21.4, the neural networks with lower RMSEv is, in this case, the neural network with topology 5-(4)j-l. We choose this ANN because it presents a lower RMSEv and lower APD The best topology developed, 5-(4)j-l, consists in five input neurons, one middle layer with four neurons and one output neuron in the output layer. To train the ANNj j model, a maximum number of 750 training cycles was established the learning rate was set at 0.60 and the momentum value at 0.80. 

---