Interpret surrogateoptplot

The surrogateoptplot plot function provides a good deal of information about the surrogate optimization steps. For example, consider the plot of the steps surrogateopt takes on the built-in test function rastriginsfcn. This function has a global minimum value of 0 at the point [0,0]. By giving asymmetric bounds, you encourage surrogateopt to search away from the global minimum.

Create a surface plot of rastriginsfcn.


Set asymmetric bounds of [-3,-3] and [9,10]. Set options to use the surrogateoptplot plot function, and then call surrogateopt.

lb = [-3,-3];
ub = [9,10];
options = optimoptions('surrogateopt','PlotFcn','surrogateoptplot');
rng default
[x,fval] = surrogateopt(@rastriginsfcn,lb,ub,options);

Surrogateopt stopped because it exceeded the function evaluation limit set by 

Begin interpreting the plot from its left side. For details of the algorithm steps, see Surrogate Optimization Algorithm.

  • The first points are light blue triangles, indicating quasirandom samples of the function within the problem bounds. These points come from the Construct Surrogate phase.

  • Next are black dots indicating the adaptive points, the points created in the Search for Minimum phase.

  • The thick green line represents the best (lowest) objective function value found. Shortly after evaluation number 30, surrogateopt is stuck in a local minimum with an objective function value near 5. Zoom in to see this behavior more clearly.

xlim([20 100])
ylim([0 10])

  • Near evaluation number 115, a vertical line indicates a surrogate reset. At this point, the algorithm returns to the Construct Surrogate phase.

  • The dark blue x points represent the incumbent, which is the best point found since the previous surrogate reset.

  • Near evaluation number 150, the incumbent improves on the previous best point by attaining a value less than 1. After this evaluation number, the best point slowly drops in value to nearly zero. Zoom in to see this behavior more clearly.

xlim([140 200])
ylim([0 6])

  • After evaluation number 180 or so, most adaptive points are near the incumbent, indicating that the scale of the search shrinks.

  • The optimization halts at evaluation number 200 because it is the default function evaluation limit for a 2-D problem.

See Also

Related Topics