Calling GA() with 10 generations 10x vs GA() with 100 generations.

Question

Simon Perales 2024-4-17

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/2108251-calling-ga-with-10-generations-10x-vs-ga-with-100-generations

评论： John D'Errico 2024-4-17

I'm running GA optimization on a problem, and I want a way to save intermediate results because of potential crashes and early exiting if desired.

I've been running GA like this:

options = optimoptions('ga','PlotFcn', @gaplotbestf,'Display','iter','PopulationSize',150,'MaxGenerations',10);
[x,fval,exitflag,output,final_pop,final_scores] = ga(fun,nvars,[],[],[],[],lb,ub,[],[],options);
for idx = 1:9
        file_name = 'dummy';% Excluded file naming code for brevity
        save(file_name,'final_pop','final_scores');
        options = optimoptions('ga','PlotFcn', @gaplotbestf,'Display','iter','PopulationSize',150,'MaxGenerations',10,'InitialPopulationMatrix',final_pop,'InitialScoresMatrix',final_scores);
        [x,fval,exitflag,output,final_pop,final_scores] = ga(fun,nvars,[],[],[],[],lb,ub,[],[],options);
end

Clarification edit: The options with each call after the first gives ga() the final population and scores from the previous ga() run with the option in optimoptions(): 'InitialPopulationMatrix',final_pop,'InitialScoresMatrix',final_scores

Is there any difference in the results GA will give (besides due to different seeding) vs running with 100 generations?

options = optimoptions('ga','PlotFcn', @gaplotbestf,'Display','iter','PopulationSize',150,'MaxGenerations',100,'UseParallel',true,'SelectionFcn',{@selectiontournament,2},'EliteCount',13);
[x,fval,exitflag,output,final_pop,final_scores] = ga(fun,nvars,[],[],[],[],lb,ub,[],[],options);

0 个评论
显示 -2更早的评论隐藏 -2更早的评论

请先登录，再进行评论。

请先登录，再回答此问题。

Answer 1

John D'Errico 2024-4-17

0
链接

此回答的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/2108251-calling-ga-with-10-generations-10x-vs-ga-with-100-generations#answer_1443221

编辑：John D'Errico 2024-4-17

Is there any difference? Of course! Make it more extreme yet. Suppose you were to stop after 1 iteration. ONLY 1. But do it 100 times. Each call will not come even remotely close to convergence. So you will get random crapola 100 times. Surely that is different when compared to allowing the optimizer to run to the point where it has converged, but doing it only once?

Even were you to take the average of 100 sets of randomly unconverged crap will still probably not be very good.

2 个评论
显示无隐藏无

Simon Perales 2024-4-17

I should have clarified, the 9 calls after the original should continue right where it left off by taking the inputs: 'InitialPopulationMatrix',final_pop,'InitialScoresMatrix',final_scores

It does converge, but I was wondering if there is anything underlying about the ga algorithm that makes continuing in a new ga() call any different than doing the full 100 generations in 1 call

John D'Errico 2024-4-17

Ok. That is different.

GA does not generate anything in the form of a hessian matrix, and that is how a memory would be employed. As such, I think the two cases would now be similar, that is as long as you can start out with an initial population that is identical to where GA ended up last. I don't know they would be identical of course. The random seed state would influence things, so if that seed managed to get touched between calls, all bets are off.

请先登录，再进行评论。