Training Patterns - Weighting

Qnet has a unique case weighting system that allows you to optionally specify weight factors that can make some patterns (cases) "more important" than others. One example of where this option could be useful is in financial market modeling. Assume that one wishes to make the training cases that represent the largest market moves more important than those that represent smaller moves. Qnet's weighting factors can be used to do this. During training, each factor is applied to the response error for that case. This means that "missing" a highly weighted target becomes more costly, causing it to drive the training response more than lesser weighted cases. By default, weighting factors are 1, making all training cases equal. When implemented, case factors greater than 1 promote an improved training response. Factors below 1 decrease training sensitivity (0 eliminates a case completely). In addition, factors are relative. If one sets all weighting factors to 2 (or any other constant value), then each case would remain equally important.

Weighting factors must be included with the training target data file if activated. See the Qnet Reference Section, Training Data Dialog for how this option is specified.