In chapter 5 a set of features were illustrated that appeared to, individually, be good discriminants between signs. None of them, however, was sufficiently accurate to use alone. In such a situation, a logical continuation is to combine such feature sets together.
Of course, we cannot expect that if we have one feature set which is 50 per cent accurate, and another which is also 50 per cent accurate, that when we combine them, we will have a 100 per cent accurate feature set; because there is an intersection in the input data of samples that would be correctly classified by either -- in other words, there are ``easy'' samples that would be classified correctly by either feature set. In fact, for the above to occur, one set of features would have to get every instance that the other one got wrong right and vice versa. Also it is possible that adding a non-useful feature set to a useful feature set may result in a higher error rate than the useful feature set alone.
Another thing that needs to be investigated is how GRASP behaves as certain aspects change. For example, we may wish to consider some of the following:
Finally, we consider ways of optimising the system to obtain higher accuracies.
In the previous sections, we saw that IBL1 was unlikely to have worst performance than IBL2 and IBL3, and that IBL2 and IBL3 were consistently somewhere between 2 and 7 per cent worse in accuracy than IBL1. We will thus focus on IBL1 and C4.5 from now on.