Is the connectionist architecture really so different?
(It is possible to view the creation of a connectionist network as a form of problem space search...
States are particular settings of connection weights & one or more example I/O pairs (the initial state is just the a priori weight settings).
An operator is the application of the learning algorithm (e.g. backprop) in the presence of one or more examples.
The desired state is the set of weights that correctly maps the inputs to the outputs.
The search is through the space of weight states given a particular sequence of learning trials.
…however connectionists don’t like to).