Parallel GA is Ignoring Time Limit?

When running a vectorized GA

options = gaoptimset('Vectorized', 'off', 'UseParallel', 'Always', 'TimeLimit', 300);

it will run indefinitely (left it running for more than two hours, and it still hadn't stopped or found an optimal solution.

But when I change it to

options = gaoptimset('Vectorized', 'on', 'TimeLimit', 300);

It runs just fine and stops after 5 minutes. Before running the parallel version, I run 'matlabpool 4'. Does the parallelization usually negate the time limit, or does someone have any clue why it would do so?

Answer 1

[Alan Weiss]

Thanks for reporting this behavior. I believe I understand what is happening, and will consider either making a change in the documentation or in the GA parallel behavior.

As documented, GA uses cputime to measure time for the purpose of stopping. When you run in parallel, the host CPU sends jobs to the workers, and spends essentially no CPU time while it waits for the results. When the results come back, it speedily computes a new population, then sends the jobs out to the workers again. So cputime does not advance much at all in each iteration, but wall clock time does.

So the parallel behavior of the TimeLimit option with GA is currently pretty useless.

If you have a lot of gumption, you can try editing the private GA isItTimeToStop function to use tic/toc instead of cputime. Or you can use an output function that stops GA after a period of time; this is probably much simpler and less error-prone.

You can also estimate how long each iteration takes, and set an iteration limit instead of a time limit.

Sorry about this behavior. We are working on either changing it, or documenting it so it won't surprise people.

Alan Weiss

MATLAB mathematical toolbox documentation

[Bryan]

Ah, thanks for the explanation!