12.9
Other Algorithms
The preceding sections list only a few of the many algorithms
that have been proposed. The list is not exhaustive by any means so we encourage
the reader to explore further for a more complete survey. Genetic algorithms,
for example, have been proposed to both generate the network structure and find
the appropriate weights. Some of the weight initialization techniques mentioned
in chapter 7 construct networks based on solutions provided by
other method, for example, decision trees (section 7.2.5) or rule-based knowledge (section 7.2.6). A polynomial time algorithm using clustering
and linear programming techniques to generate classifier networks is described
in [276]. Projection pursuit regression [129], [185], a well-known statistical procedure, creates a system
similar to a single-hidden-layer network with a linear output node. It is
constructive in the sense that it adds projection directions (corresponding to
hidden units) sequentially until the error is sufficiently
small.