Improve optimization algorithm for finding thresholds

[auduno]

The current implementation of optimization to find the thresholds at given levels, is a mocked up gradient descent algorithm where simulation is used to find the gradient descent steps. While it works decently, it is very slow and is not guaranteed to converge.

Any suggestions for improvements are welcome, but an algorithm that might work better is the one described in this paper : http://arxiv.org/pdf/1107.1744v2.pdf