I haven't been doing math in years and I'm facing the following problem. I'm trying to implement in Java a linear regression under a set of inequality constraints. Sorry in advance for all the approximations while trying to formulate this problem...
More precisely, I have a linear function $f_\Theta:R^n->R$ defined as such: $f_\Theta(X) = \theta_0 + \theta_1x_1+ \theta_0x_2+... \theta_nx_n$, with $\Theta$ a (n+1) vector.
My aim is to $\min_\Theta \sum_{i}||f_\Theta(X_i)-Y_i||^2$ under the constraint that $f_\Theta(X_i)-Y_i < 0$ for a maximum of $i$ (I wish I could avoid to take into account "obvious" outliers in my observations)
So, bottom line, I'm looking for the linear regression that will act as the "best fit" providing a lower bound for most/all the observations.
How can I do that ? I've been looking for a Java Library that would implement such a thing. I known the Apache Common Math package and I also found this Flanagan library, but it does not seem to address this.
Any help greatly appreciated. Thanks