All learning algorithms have certain aspects which may be labelled as ``tweaks'' -- factors that significantly affect the behaviour of the algorithm for which there is no obvious or straightforward way to set. For example, in HMM's the set of states and transitions of the the HMM is difficult to arrive at. For neural nets, the number of neural nets in each layer, output encoding, neuron interconnection and error function need to be considered. In instance based learning techniques, deciding on the similarity function is a major difficulty if high performance is required. In symbolic learning there are also variable such as the pruning level, assessing the usefulness of particular attributes and so on.
However, instance-based learning and symbolic learning appear to be the least complex of these.
Furthermore, because of the speed of learning compared with
other techniques, effective state-space searches (eg hill-climbing,
simulated annealing or just good old exhaustive search) can be
accomplished in a feasible time
.