The RATS Software Forum

NLLS has a METHOD=GENETIC option.
I assume this is a genetic algorithm, and it works well on a small problem (4 declared NONLIN variables with 6 observations)
where the partial derivative of the squared errors is discontinuous in one of the four variables
but all partial derivatives exist for the other three variables. METHOD=GAUSS fails, of course, due to its reliance on derivatives.
METHOD=SIMPLEX also works well, but seems to require a higher value for ITERS than GENETIC.

I have several questions about how RATS implements GENETIC.
(1) Are the variables coded as a single concatenated binary string? What length are the concatenated strings?
(2) What is the probability of crossover?
(3) What is the probability of mutation?
(4) What is the population size?
(5) Can any of these be changed by the user?
(6) If I "know" the upper and lower bound for my one ill-behaved variable value, can I input this information somehow?

Thanks,
Ken Cogger.

METHOD=GENETIC is described on page UG-117 of the RATS Version 8 User's Guide.

You can use the REJECT option (page UG-120) to bound the values. That makes the function value "NA" for any set of parameters which gives a "true" value for the rejection formula.