How can I avoid negative values in the output of a feedforward net? Note that I do not have negative values in the inputs.
.
it is always possible to enforce positive values on the output by using a positive transfer function (say, logistic sigmoid) and a simple max normalization of the output data (so that they fit in [0, 1])
.
.
it is always possible to enforce positive values on the output by using a positive transfer function (say, logistic sigmoid) and a simple max normalization of the output data (so that they fit in [0, 1])
.
It will be helpful, if you explain why you need them all to be positive. If the target values are all positive, then before training, you can scale up target values of the samples proportionally with a positive value, so the negative output of the net will be less likely to be negative after training. An alternative is to use a transform function to transform all outputs to a positive range, say [0,1]. You can also consider all negative outputs as predefined positive values, depending on your task.
Than you Homayoun for your response. In my study I used set of building characteristic data to predict savings from the adoption of individual measures – based upon actual building data. I used feedforward net toolbox to create the net. I set the actual data for energy consumption of the houses as target and the building characteristic as the input to the net. I got really nice fit between the actual data(Target) and neural net work outputs. The question is when I use this net to forecast the energy consumption for the buildings by making some changing in the neural net inputs I did not get good results and there are some negative value in the outputs.
- Badges
- Science topic
- Similar topics
- Engineering
- Materials Engineering