Swarm Intelligence

Disadvantages

Schaffer, Whitley, and Eshelman (1992) state: “Using a genetic algorithm

as a replacement for back propagation does not seem to be competitive

with the best gradient methods (e.g., quickprop).” GAs are known to per-

form global search quite well, but to be relatively inefficient in fine-

tuned local search (Yao, 1995).

Evolution of network topology is generally done in ways that re-

sult in discontinuities in the search space. Examples include removing

and inserting connections (weights), discrete changes in connections

(weights), from 1 to −1, for example, and removing and inserting PEs.

These discontinuities usually require retraining