[Coin-ipopt] constraints with divisions

[Berry, Jonathan]

I have a set of constraints that looks like:

f[i]/f_orig[i] <= 1 + epsilon

With small datasets, this type of constraint
works well with ipopt in my model -- good
solution in about 200 iterations.

However, in bigger datasets we might encounter
an f_orig[i] value of 0.  Modifying the constraints:

f[i] <= f_orig[i] + epsilon*f_orig[i]

ipopt gives the same solution on the small dataset,
but now it takes 5,000
iterations rather than 200.  When I apply this
type of change to enough instances of this
type of constraint, it doesn't finish in 100,000
iterations.

By Andreas' very useful suggestion, I have
been running with ISCALE=4.

Any ideas?

thanks,
   Jon