 |
Opti++
0.1.0
Optimization framework in pure C++
|
|
Opti++
0.1.0
Optimization framework in pure C++
|
Opti++ supplies a given set of analysis exploration techniques out of the box. In this pages there is an overall description.
Design Of Experiment (DOE) is a set of techiniques devoted to exploration of variables spaces. Such strategy can be used when an exploration of design space is required or when any optimization strategy could not be applied (i.e. when evaluation of objective function is too much time consuming). Opti++ has two DOE techniques: Full Factorial and Latin hypercube.
Full factorial is the simplest possible available DOE techinique. It just creates a grid on each variable. Than a matrix is creating crossing each node of the grid with any other. Such technique is appliable when it is possible performing a medium or large set of analyses. An example image is available in below image. In such example there are two variables, X1 and X2. X1 domain interval is splitted into 14 nodes. X2 is domain interval is splitted into 6 nodes. Final matrix of full factorial DOE analysis is made by 14x6= 84 samples. For each variable it is possible to define grid elements both via equispatial splitting of domain interval or with an explicit list of values.
Latin Hypercube is always based on discretization of each variables domain interval. But it is preliminary chosen the number of total samples (let's say N). Later each variable's domain interval is split into N intervals (i.e. N+1 nodes). Finally it is randomly chosen a set of sub intervals. Choices are performed in oreder to avoid repetition of any variable's interval. Such technique is appliable when it is not possible performing a medium or large set of analyses and a wide exploration is required. Taking as an example the image below, it was set N=8. For each variable was created a set of 9 grid elements. Finally it was chose a set of 9 variables sets in order to respect constraint of not repetition.
1.8.15