Genetic Algorithm Fitting

The genetic algorithm generates random solutions and ``breeds'' and ``mutates'' them, trying to find the solutions with the highest fitness. The implimentation here differs from typical ones (c.f., Charbonneau, P., 1995, ApJSuppl, 101, 309) in that breeding and mutation are performed on the actual parameter values, and not some encoding of the parameters (usual done by turning the floating point numbers into strings, and then treating each digit as though it were an amino acid in a gene). Instead, breeding is defined as swapping parameter values between solutions, and mutation is defined as adding a normal deviate to a parameter value. With this technique, it is crucial that the use redefine the hard limits of parameters (using the new command).
Parameters:

nsamples

Number of random solutions to generate.

ntrials

Number of iterations (i.e., generations).

mode

N=``new'', meaing give initial solutions random param values (default)
P=``perturb'', meaning set initial solutions to random perturbations of current fit, essentially done by mutating the currnet fit nsamples times
N is more appropriate for fitting a model for the first time, P is intended for refining a fit.

cross_prob

Probability of breeding (parameters) between two solutions.

mut_prob

Probability of a mutation affecting a parameter.

mut_amp

Size of mutations (st. dev. of normal deviates) as a fraction of the allowed range of the parameter.

outfile

Append the current GA arguments and final fit statisic to the file outfile, mainly intended to be used for debugging and performing a grid search of the arguments to find optimal values.

In general, for fits involving only a few parameters and with good signal-to-noise, normal non-linear fitting techniques (i.e., the fit command) are more appropriate, particularly when used with error all which probes for relative minima. However, for models with many parameters and/or a bumpy error surface (i.e., lots of relative minima), the GA fitting technique can be much more efficient. In these cases, the NL fitting will converge quickly, but is very sensitive to the choice of initial parameters and the presence of relative minima.