4 ms·
Hmm, so what is the theory of the mutation function? Is it just determined ad-hoc from looking the problem and throwing some standard examples at it? It seems
by joe_the_user 7y ago
Hmm, so what is the theory of the mutation function? Is it just determined ad-hoc from looking the problem and throwing some standard examples at it?
It seems like you could implement effectively any function with a complicated enough combination of fitness, selection and mutation functions but without a theory of which to use, progress would be a bit hard.
- thephyber 7y agoI think it's usually a function of what data structure your "genotype" is stored in. In my experiments with GAs, defining a coefficient for mutation that is either far too high or far too low means you search search for too much of the problem search space or far too little, respectively. Testing parameters such as the frequency of mutation are part of the iterative process of tuning your GA.
- jacques_chester 7y agoSome schemes add those parameters to the genome input itself. IIRC it's not a big win, but I don't recall at the moment why.
- richk449 7y agoVery true - there are lots of what the AI people call hyper parameters to tweak. And they do affect the balance between fast convergence and avoidance of local minima traps, as you would expect. Once you have used them for a while, you start to get a feel - if the optimization is rapidly converging to a poor solution, you need more variation - more mutation, bigger population, etc. if the solution is converging very slowly, you need Less variation. I like to set up the functions as Gaussian, then the parameter controlling the variation is a well understood sigma.