Sensitivity Analysis

This type of analysis can be used to do Feature Selection for a Neural Network by looking at how big the influence of a tiny change in one input variable has on the output of the very first prediction.

$\frac{\partial y_t}{\partial u_{\tau \le t}}$

There are four possibley types of relations

• linear with constant first derivative
• monotone, can be used in one dimensional control
• non-monotone, can only be used in multi-dimensional control
• near zero, useless for modelling and control