... employed¹

Most likely because the cost of devising and implementing learning systems is prohibitive, especially for the somewhat limited and disjoint problem domains.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... intelligent²

The other school of thought is that initial input into the design of a system does not constitute a homunculi, and that systems without independent intelligence require a homunculus to be present at all times. Both of these viewpoints are debatable.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... classifications³

In other words, only supervised learning was considered, not unsupervised learning.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... bad⁴

For example, one of the initial rule sets constructed had a large error on the training set of ~ 70%, yet had an error of ~ 50% on the validation set. This could be coincidence but could also be valid, and this is the whole problem - when faced with a training set error of 70%, how is the human to know if this reflects bad rules or bad luck?

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.