I am working on a CBT project that used neural networks in MATLAB as a universal approximation of functions, and obtained very good results: more than 70% of the samples present a relative error of less than 10% in relation to the real value, and 60 % of the total presents an error of less than 5% in relation to the real value.
However, I know what values the results should take, are some discrete values, all known. I would like to improve my work, using some method to approximate the values obtained by the network to the actual values, is this possible? Is there any algorithm, or any statistical analysis I can use?
For example: if I create a vector with all possible values (171 possible values), and compare the values found by the network with the possible values, and get closer to the nearest one, I can solve this problem for those they present low relative error, which gives about 70% of the total. But assuming I do not know the relative error, is there any method that I can increase the efficiency of my work?