superimposed fields at truss top-opt

Comment

Comment by Martin D. on February 1, 2015 at 12:27pm

Hey Robert,
thanks for your answer! Haven't heard of CPPNs yet, but as far as I saw it seems to be really helpful for evolutionary solvers, because of its diverse but small search space.
Furthermore I see great potential for the low amount of parameters, as you described it. I probably could have used it in some of my earlier projects...too late :D
But I think, what is also really interesting in this approach is definetly its aesthetics evolving out of the algorithm. In your video its not only the symmetry, but also that the members of the truss are not placed randomly. They seem to follow a pattern even when you pause the video and look at a still. That also counts for the examples of Dan Richards in my opinion.

Thanks,
Martin

Comment by Robert Vier on January 18, 2015 at 12:03pm

Hi Martin, thank you.

It's been evolutionary, done with octopus. The idea of superimposing fields or functions comes from CPPNs which I implemented for GH recently, so did Daniel Richards in Java to see on http://www.performativeassemblies.com/

The GH-magnetic fields are one way for a flexible parameterization [to generate different typologies of bearing structures, hence maybe more a 'true' search instead of a quantitative optimization; here, diversity preservation is quite important if you do not want to have all those different typologies just at the beginning of the process] which at the same time does not need too many parameters (as convergence of an evolutionary algorithm strongly depends on that). Each charge point was driven by X|Z coordinates, controlled by octopus. Just in the animation the charge-points' position was morphed between four solutions I think. The distribution function of the points along the middle axis was also controlled by octopus, but this was formulated as a mathematical expression.

Best

Robert

Comment by Martin D. on January 18, 2015 at 7:52am

Hey Robert,
interesting approach! I think it is quite smart, to use the flow lines as sperators of the truss. Could you share some experiences on how efficient this method is? Meaning, do you think the solution space is quite big and the best fitness easily to find?
And how did you set up the Genome/moving of the points? Do they move on a fixed curve and if so why constrain it like that?
The process does not really seem to be evolutionary right?

Sorry for all these questions, but I think it is always interesting to see, how to set up an non-linear optimization process. :-)

Greetings!