Mutual Information Computing
The tool computes the mutual Information (MI) between several input variables xi (features) and an output y time series (target).
For linear relations (e.g. by a linear mixing matrix) the mutual information serves as a mean to unmix the independent sources, see blind source seperation ICA.
Here, we have another question: Before analyzing the relations, what are the most important, essential features for predicting the output? For linear and non-linear functions y (x1,x2,..) the mutual information MI (y,x1,x2,..) is maximal for the essential inputs. In order to find them, the mutual information of order two (MI2) is computed between the target y(t) and
- each variable with different time delays or
- each input x in precondition on already selected, better features with
MI2( y(x) | x1*, x2*..) = max
Thus, the input features are ranked due to their importance.
A Fuzzy User Interface
This demo program uses a intuitive fuzzy interface for input of vague knowledge. It uses training data to update the fuzzy knowledge and reflects it as new fuzzy rules.
The demo data shows how it works as a trained medical diagnosis tool.
Download » Analyst1.exe
Neural Network C++ class libraries
There are two class libraries available: