Class FitController

Overrides: object.__init__

(inherited documentation)

Reset the bounds on parameter p.

Range can be::

None: let p vary in (-inf,inf) value: fix p at the value (low,high): vary p in range (low,high)

Reset the fit population so that parameter p varies uniformly within its range.

This is not a sensible option for all optimizers.

Run a local optimizer at point p.

The result will be sent to the FitMonitor. The user may opt to introduce this new individual back into the fit population.

If no p is given, then run from the best current value in the fit.

If no optimizer is given, then run a quasi-Newton optimizer such as secant or dog-leg for functions without derivatives or tnc for functions with derivatives.

Note: it is theoretically possible to split the minimization into D0 and D1 portions, using different algorithms for each.

Introduce an individual to the fit population.

The suggested value may come from a local optimization from point p or from the user playing with the parameters in the foreground while the global optimization is running in the background.

Lamarkian evolution.

Improve the fittest individual in the population using a local optimizer and reintroduce it into the population. This strategy is meaningful for a number of the stochastic optimizers including simulated annealing since it allows the user to control the goal directed behaviour.

This is a server-side combination of localopt and suggest, without the need to cycle the results through FitMonitor.

Search near p.

Sigma is the size of the attractor relative to the entire search space.

This is a user supplied heuristic telling the stochastic search to prefer to look near point p when proposing new members of the population. This will not affect the calculated objective value

The attractor translates into a scaled gaussian peak.

Avoid p. Avoid a particular region of the fit space when choosing new individuals in the population.

Having identified a bad local optimum in the fit space, have the stochastic optimizers avoid the this region using a penalty function based