Nonlinear constraints in ga and gamultiobj with binary variables.

Does the population type 'bitstring' allow for the use of nonlinear constraints in ga and/or gamultiobj?

I have been attempting to use it in ga and failed. If nonlinear constraints cannot be used in these cases, then my question is: I can easily change my objective function for a penalty function (following the same references used in matlab documentation for the augmented lagrangian genetic algorithm), but the penalty parameters require general information about the ga (current iteration, best fitness, etc). I think some of this information may be stored using global variables in my objective function, but I was wondering if there is some way I can retrieve the genetic algorithm variables used by the ga() to my objective function workspace while ga is running.

Thanks.

Answer 1

[Walter Roberson]

Does the population type 'bitstring' allow for the use of nonlinear constraints in ga and/or gamultiobj?

No.

The way to proceed was to provide your own custom functions for initial population and mutation and crossover which always return integer results according to your needs; those can be used with non-linear constraints because you will not tell ga that you are running with integer constraints.

[Alan Weiss]

To quote from the documentation regarding the bitstring population type: "You cannot use a Hybrid function, and ga ignores all constraints, including bounds, linear constraints, and nonlinear constraints."

Alan Weiss

MATLAB mathematical toolbox documentation

[Walter Roberson]

Right. In the solution I gave you would not use bitstrings, you would use numeric values with lower bound 0, upper bound 1, and you would use custom functions so that you never generate non-integers.

[Opt User]

Yes, it worked using a continuous population and constraining all variables to integers between 0 and 1. But I wonder if there is any convergence advantage to this approach. Originally I was setting all my constraints as penalty terms of my objective function and using the bitstrign setting and it seems like I was obtaining faster convergence.

[Walter Roberson]

If you do not use the custom crossover and mutation and initial population, if you just round() the values in the function as you use them, then the ga will wander around trying (for example) 0.82 and 0.87 trying to find places that give better results, not understanding that you need to go to 0.499* to change the effective value. bitstring and intcons only use the exact values, but they do so by dropping in integer-constrained hybrid and mutation and population functions.