I am using a radial basis network with a thin plate spline as an activation function for the centers. I am finding the network parameters using Recursive orthogonal Least squares algorithm.
Scaling- I was unable to get proper training of the network with the original data(the data consists of inputs and outputs in different ranges.one in 1000's and another below 10) without scaling.(Divide the inputs by its maximum value so that the input ranges between 0-1 before giving it to network)
Even though i get decent rmse(
Are you sure the dataset is appropriate for model development?
You have to be!
How many inputs do you have? What’s the size of dataset? Have you apply any pretreatment method? Is dataset contaminated with noise?
I suggest you first normalizing your dataset and employ MATLAB toolbox (nntool) to develop RBF models. Check the model’s performances and find the best parameters for RBF training and then apply recursive orthogonal Least squares algorithm to your dataset.
Notice: the number of cluster centers should not be too much, because it may lead to over fitting problems. You can use the geometric pyramid rule to find the minimum number of cluster centers.
Hi,
I have 2 inputs and 2 outputs The dataset is simulated through a company software emulating the plant. I kinda have no choice rather than build a model with it. I did a sort of normalization by scaling and making all the inputs and outputs within 0-1 range.(If you mean normalization by zero mean and unit variance then i didnt do that.Should i?) .The problem arises when i find model parameters by OLS. Since the values are between 0 and 1 when i find inverse (by least squares) the parameters are large.I did encounter the problem of overfitting and i currently doing my overfitting analysis.Is there any scaling procedures that one particularly employ for RBF ??
- Badges
- Science topic
- Similar topics
- Geoscience
- Asia
- India