Why can I not to see the convergence graphs?

Replies to This Discussion

Permalink Reply by Robert Vier on August 15, 2014 at 15:15

hi,

this is strange indeed.

can you post the definition or send it over?

best

Robert

Permalink Reply by Dan Hou on August 16, 2014 at 9:40

Hi Robert,

I have rerun the optimization and the convergence graph appeared. But I have some wonders about the optimization and I think they have great impacts on my results.

1. Whether the objectives must be conflicted. If not sure what are the relations among them, can I use MOEA? I see the pareto optimal frontier in my case is not like others in your example. I think the objectives which are selected lead to this appearance because not all of them have conflicts.

2. Whether the variables must be sliders. As my variables are dependent, the ranges of some variable are affected by others. I use remap component to automatically change the range, sometimes multiple values in slider correspond to the same new value. When taking this method into the optimization cases, this condition will lead to duplicate calculation and thus make the results unevenly distribute. Do you have some ideas to solve this problem?

3. Why do the upper and lower boundaries of the first convergence graph both vary? And what is the meaning of the second convergence graph with no variation?

4. What is the hypervolume graph like when the results converge?

Thank you for your help.

Dan

Attachments:

• g9.PNG, 171 KB
• g9-1.PNG, 71 KB
• g9-2.PNG, 77 KB

Permalink Reply by Dan Hou on August 16, 2014 at 9:45

These are the rest images.

Attachments:

• g9-3.PNG, 75 KB
• 1.PNG, 11 KB
• 2.PNG, 98 KB

Permalink Reply by Robert Vier on August 16, 2014 at 13:00

hello dan

1. a pareto front can have a lot of different shapes. convex / concave / discontinuous / lower-dimensional (the latter e.g. if some goal dimensions correlate 1:1)

in principle you can use any objectives, no matter what their relations are. the formulation of the fitness functions always is kind of crucial in evolutionary optimization, and sometimes quite tricky. there is hardly a general answer to this. but for instance, if your goals are not contradicting, the algorithm would converge much faster, opposed to contradicting goals where more diversity in a solution set is produced.

you could start with all the goals you can think of, then look at the resulting pareto-fronts and how the overall process goes, then filter out the necessary ones, and/or turn some of the goals into hard constraints.

2. the variables must be sliders or gene lists for octopus. the more variables there are, the harder the problem. the more you satisfy 'little change in variables causes little change in goals', the easier the problem is to solve. if you change sliders and nothing changes in your goals, you should probably rethink your problem definition - i cant tell if its really necessary in your case without having an example.

3. the manual tells you that these are the boundaries of the pareto-front (dark) and the elite (light) of the number of history generations set in the top left corner of the viewport. the lower boundary can go up again as a result of the filtering strategy (either hypervolumeContribution or spea2-archiving) - the set of solutions per generation converges to smaller goals over time, some 'bad' extreme solutions are truncated sometimes - hence the lower bound can also go up.

no variation means you have an ill-defined goal there

4. please refer to this paper for example, which is linked in octopus' examples page:

http://www.tik.ee.ethz.ch/sop/publicationListFiles/zitz2007a.pdf

or one of the following:

http://scholar.google.co.id/scholar?q=hypervolume+multiobjective+op...

Best

Robert

Permalink Reply by Dan Hou on August 17, 2014 at 11:04

Hi Robert,

Thanks a lot for your patient answers. They are very useful for me. I have read something about MOEA, but there are still many things hard for me. I will continue to learn and now I urgently need to solve the issues in my first case.

1. I see the pareto front of my case is unevenly distributed. Except for some errors of defining variables and objectives, is it also related with the features of them? As the variable is actually discrete, they lead to the uneven distribution of objective values and thus probably affect pareto front.

2. One of objectives is ill-defined and thus it is not converged, but I don’t know what is wrong with this objective. The recorded values of them are no problem. I upload the record of these values which can be opened in the Excel. The only one difference is that the generation of these values is later than other objective value because they need some time to calculate after simulation. Does this reason lead to non-convergence?

3. My wonder about defining constraints of variables is in the files. Hope you could check it. This file needs ladybug&honeybee to run.

4. I have read your suggested file. I’m sorry that I don’t really understand what is the function of mark preferred.

5. Why does the text export nothing?

Looking forward to your help. Thanks again.

Bests,

Dan

Permalink Reply by Dan Hou on August 17, 2014 at 11:05

Hi Robert,

Thanks a lot for your patient answers. They are very useful for me. I have read something about MOEA, but there are still many things hard for me. I will continue to learn and now I urgently need to solve the issues in my first case.

1. I see the pareto front of my case is unevenly distributed. Except for some errors of defining variables and objectives, is it also related with the features of them? As the variable is actually discrete, they lead to the uneven distribution of objective values and thus probably affect pareto front.

2. One of objectives is ill-defined and thus it is not converged, but I don’t know what is wrong with this objective. The recorded values of them are no problem. I upload the record of these values which can be opened in the Excel. The only one difference is that the generation of these values is later than other objective value because they need some time to calculate after simulation. Does this reason lead to non-convergence?

3. My wonder about defining constraints of variables is in the files. Hope you could check it. This file needs ladybug&honeybee to run.

4. I have read your suggested file. I’m sorry that I don’t really understand what is the function of mark preferred.

5. Why does the text export nothing?

Looking forward to your help. Thanks again.

Bests,

Dan

Attachments:

• uniformity performance, 7 KB

Permalink Reply by Dan Hou on August 17, 2014 at 11:07

Er...Sorry it duplicates. You can overlook the second reply.

Permalink Reply by Dan Hou on August 18, 2014 at 1:22

I also have a wonder when the optimization engine evaluate whether the variables or objectives meet the constraint. In other words, does it only generate the population within the constraints or judge them after the population formulate?