This appears to be a classification problem; NOT a clustering problem.
1. 200 data points define a 199 dimensional space. Using a dimension of 1000 is not reasonable. You can reduce the dimensions using PCA. However, since your problem is classification instead of regression, LDA or PLS may be better for reducing dimensionality.
2. For an I-H-O node topology (I <= 199, O = 2), the number of unknown weights is Nw = (I+1)*H+(H+1)*O = 2 + 202*H for I = 199 and the number of training equations is Neq = Ntrn*O = 200*2 = 400.
3. For accurate and stable weight estimates that will generalize well to nontraining data;
Neq >> Nw or H << (Neq-O)/(I+O+1) = 398/200 ~ 2
Therefore, it looks like you need to substantially reduce your input dimensions below 199 and/or substantially increase the number of training examples above 200.
Hope this helps.
Greg
