Under what conditions do the fitness values in the GA solver (genetic algorithm) not fall monotonically over iterations?

Using MATLAB R2014a, but get similar results for MATLAB 2013b. Global Optimization Toolbox. Options are the following:

optParams.gaOpt = gaoptimset(... 'Display', 'iter', ... 'Generations', 10, ... 'InitialPopulation', {[]}, ... 'PopulationSize', 16, ... 'TimeLimit', 5*60*60,... 'StallGenLimit', 100, ... 'TolFun', 1e-16, ... 'PlotFcns', @gaplotbestf, ... 'MutationFcn', @mutationadaptfeasible, ... 'EliteCount', 1,... 'CrossoverFraction', 0.2, ... 'Vectorized', 'on', ... 'UseParallel', 'never' ... );

I expect that the plotted fitness value from @gaplotbestf to fall monotonically for non-zero EliteCount.

Here's a Mathworks link where the 2nd and 3rd figures show non-monotonic fitness function values:

http://www.mathworks.com/help/gads/examples/genetic-algorithm-options.html

The problem is when my optimization terminates the result I get is worse than the best possible value found over all iterations. I also worry that the best population member is not carried forward and so does not contribute to improving the optimization.

Best..

[Rakesh Kumar]

The new version (R2014a) seems to have a bug (I will investigate more) with default EliteCount value (it is set to zero for smaller population). This is the reason for the example plot to show non-monotone best fitness function value.

Your options setting for EliteCount will not have this bug in R2013b or R2014a.

However, I have some questions about the options settings you posted. You are specifying an initial population as {[]}, a cell array is used to indicate custom population. And at the same time, you are using @mutationadaptfeasible which is for 'doubleVector' population. These options will result in error. Am I missing something here?

Can you post more details about this problem?

Regards, Rakesh

[JMB]

Thanks for the tip Rakesh. I actually set the InitialPopulation to null further down in the code before running the ga. Sorry for the misunderstanding. Everything seems to be working now.