Training standard one hideen layer backprop for Mglass Data

[William Hsu]

Dear Researchers,

	I have been trying to train a standard backprop network
with one hidden layer with varying number of sigmoid units to
learn the Mglass (Chaotic Data ref. Platt, Moody) Time series.
I have not been able to acheive Normalized root mean square values
of < 0.4 for archiectures to 50 hidden units. The architectures I have looked
at includes 4-N-1. for various N. where the output unit is a linear
summation unit (without sigmoid). I would appreciate hearing from people
who have had similar experience or better still, people who have
successful experience with the above setup. Pls email directly to me.
Thank you very much.