This paper gives a classification system for the kinds of learning machines based on several
dimensions, whether learning is parametric or structural, the kinds of training information
utilized, and so forth. Various machines have been categorized using this system and are
discussed in Part I of this paper, here, or in other surveys. The second portion of this paper
discusses relationships between the sizes and distributions of the target and learnable
classes of behaviors and the error levels and rates of learning. If the learnable class is large …

Computer science has historically required programmers of systems to anticipate every
possible behavior that could be desired and to program in advance all the knowledge and
mechanisms needed to achieve it. Unfortunately, it has been found that such extensive and
explicit programing is expensive and it still, in many cases, does not achieve the range of
behaviors that might be needed. The only alternative is to have the machines program
themselves to acquire the knowledge they need to function satisfactorily. This chapter has …