genetic algorithm solver convergence
7 ビュー (過去 30 日間)
I am trying to maximize a function using ga solver. The value of the objective function that I get when the optimizer terminates is not the highest one (I can see some higher value in the command window). The function has four design variables and each of them is constrained between an upper value and a lower one. I'm using the following parameters: population size=20; creation function: use constraint dependent;scaling function=proportional; Selection function=Roulette; elite count=2; crossover fraction=0.8; mutation:uniform (probability=0.01);migration: forward, crossover:Heuristics (1.6)
Even the values of the design variables that I got are not the same as the optimum and those values vary according to the initial population (I'm sure it's a local maximum and not an absolute max.).
- * I know that the convergence is a problem dependent, but do you have some general comments/suggestions??**
Alan Weiss 2015 年 10 月 8 日
It is not easy to assess the reliability of a ga solution. First, I want to make sure that you are maximizing by taking the negative of your fitness function. If so, then the best advice I can give you is to use patternsearch instead of ga. patternsearch is faster, more reliable, and easier to tune. To search for a global optimum, start patternsearch from a variety of initial points. For example, if you have finite bounds on all components (and you should), then repeatedly try
x0 = lb + rand(size(lb)).*(ub - lb);
FYI, there is some documentation on how to assess the reliability of a solution, but this documentation is more aimed at Optimization Toolbox solvers than Global Optimization Toolbox solvers.
MATLAB mathematical toolbox documentation