Table 1 Description of the types of hunters used in the benchmark tests
Hunter class | Description |
|---|---|
EvaluationsUnmoving (call, tol) | Calculates the standard deviation of the last calls function evaluations. Returns true if this value is smaller than tol times the last function evaluation. Used to terminate an optimizer when its function evaluations are unchanging, i.e., when it has reached its focus phase. |
ValueAnnealing (med_kill_chance) | The probability of returning a kill signal follows an exponential distribution based on the difference in function value between two optimizers. The chance of killing an optimizer twice as large as the lowest optimizer is med_kill_chance. Optimizers which are exploring values which are close to one another are less likely to be killed than those far apart. Used as a way to save optimizers which are competitive and may become the best. |
BestUnmoving (calls, tol) | Kills an optimizer if it has not improved its best ever function evaluation by at least tol percent in calls. Used to terminate optimizers that explore for too long without focusing on to a point. |
ParameterDistance (relative_tolerance) | Kills optimizers which are exploring points in the domain which are separated by a distance less than relative_tolerance times the maximum distance between any two points within the bounded domain. Used to terminate optimizers in the same basin. |