In this case, the task is to predict the next value in the sequence (ergodic and chaotic) generated by the logistic map [Lapedes:87a], according to the recurrence relation:
We tried different architecture and obtained a speedup of 400-500, and slightly better generalization properties for the BFGS optimization method presented.
In general, we obtained a larger speedup for problems with high-precision requirements (using real values for inputs or outputs). See also [Battiti:89a].