Measuring Software Reliability: A Trend Using Machine Learning Techniques - Complex System Modelling and Control Through Intelligent Soft Computations - page 805

Fig. 7 b versus t

that minimum error is incurred. The estimated and tuned values of all the param-
eters are then used in the testing process to predict the failure for the next week
given the failure in the current week, and the relative error for the predicted output
is computed.
The graph in Fig. 7 shows the variation of error of prediction with time and is
generated for the testing portion of the data set (from 22nd to 28th week) using the
parameter values obtained in the training process. The decreasing amplitude of this
graph implies that with the availability of more data, the heuristics become stronger
and the prediction becomes more accurate giving lesser error. Figure 8 gives the
graph of relative error versus time for 11th 28th week, and this also shows

decreasing amplitude, validating the fact as mentioned above. However, steep
fl

-

fluctuations are visible in the graph owing to the fact that random values are chosen
for Lagrange multipliers (Fig. 9 ).

5.3 Discussion

Software reliability analysis has been arranged using two methods; one is a sta-
tistical method while the other is based on machine learning. Above results clearly
reveal that support vector regression overcomes the limitation of NHPP modeling.
The parameters in SVR can be appropriately tuned through learning over the data
set itself therefore giving no space to non-existent or unstable parameters. This is
evident from the fact that 11th and 12th week GO model parameters could not be
estimated by MLE but using SVR, learning parameters could be appropriately
estimated for these weeks. The plot produced below is concentrating on the sample
error normalization strategy as to quantify minimum rate of error under each