The Approximation Trainer adapter allows you to configure and train an approximation based on
a job that has previously run. Output variables from the job and a selected approximation type
are used to train a new approximation.
You use approximations to visualize the behavior of your data (data trends) and to predict
new values of your output parameters based on a specified combination of values of your input
parameters, where that combination does not already exist in your data set.
For each output parameter, Optimization
Process Composer uses the selected approximation technique to predict the value of the output variable. The
following approximation techniques are available:
- Response Surface Model (RSM)
- The response surface model
(default)
uses a polynomial combination of vectors representing the input parameters. The order of
the polynomial regression model depends on the number of data points in your data set.
By default, Optimization
Process Composer uses a polynomial with up to tenth-order uni-variate terms and fifth-order cross
terms for the model. However, if the number of data points is small, Optimization
Process Composer chooses the order of the terms based on the number of parameters and the number of
alternatives.
The response surface model uses simple equations that quickly and
easily fit the data; however, it is valid only for simple smooth functions (linear
functions with limited noise) or in local regions. In addition, the length of time to
fit an approximation is dependent on the number of points, and the response surface
model can be slow if you have a large number of data points. To increase performance,
you can limit the maximum number of data points
to
be used by the approximation, and you can reduce the polynomial
order of the terms.
- Radial Basis Function (RBF)
- The radial basis function model is a type of neural network employing a hidden layer
of radial units and an output layer of linear units. The radial basis function model has
a short initialization time and is generally faster than the response surface model for
a large number of data points. In addition, the radial basis function model is
preferable when you know all of the inputs to be independent and when all of the inputs
are of equal importance. The radial basis function model is also preferable to the
response surface model when your data are nonlinear and when the data fall into
specified categories, such as strings defining the model or the manufacturing type.
- Universal Kriging
-
The Universal Kriging model is an interpolation method that converts partial
observations of a spatial field to predictions of that field at unobserved locations.
The model is useful in predicting temporally and spatially correlated data and
typically creates a good approximation in cases with a small number of data points.
The Kriging model is very flexible and allows you to choose between a wide range of
correlation functions for building the model. Depending on your choice of the
correlation function, the model can either honor the data (providing an exact
interpolation of the data) or smooth the data (providing an inexact interpolation).
Depending on the number of input parameters, the number of design points, and the
number of responses (outputs) of the Kriging model, the process of building the model
can be very time consuming. As the size of the matrices increases, the amount of CPU
power required for manipulating the matrices grows exponentially. Therefore,
generating a good Kriging model that uses many design points can take a substantial
amount of time even after all the data points are analyzed.
You can also create an approximation using Results Analytics. For more information, see the
Results Analytics User Guide: Approximations.