Hacker News Reader

Comment by josephg | Hacker News Reader

This is really clever, but I bet there's all sorts of wacky ways that a 3d version of this could be optimized.

It would be really interesting to throw this design into a learning algorithm with an attached fluid simulation to estimate effectiveness. I wonder if anyone's done this. What learning algorithm would be most effective? And I'd love to see what a learning algorithm would come up with. I suspect an optimized version of this would be really organic looking and have lots of weird spirals and things in non-obvious ways.

View on HN

Akin to this: https://en.wikipedia.org/wiki/Evolved_antenna

And that in turn reminds me of the obscure zig-zaggy convoluted shapes in molecular biology: https://www.youtube.com/watch?v=B_zD3NxSsD8

These are also born out of an evolutionary optimisation process, and probably has a form that is intrinsically linked to its function - we just haven't remotely uncovered yet.

Continuing this line of thought and i am reminded of steganography, hiding complexity in simplicity or objects with several functions. Also encoding by moving things in relation to each other and the mysterious field of emergence. Especially in biology where complexity or patterns can emerge seemingly from nowhere and hide properties that has to do with relations between points instead of the points themselves.

Alogorithms like what you are describing (minus the learning part) are a very active area of research, its quite exciting! https://en.wikipedia.org/wiki/Topology_optimization

It would have to balance multiple objective functions, 1) does it impede fluid flow 2) is it buildable, can it be constructed 3) is it sturdy, will it last, will it perform under the required criteria. Train an AI to turn knobs that drives the parameters (hyper) that generates networks that solve problems across multiple competing objective functions.

Don't generate networks, generate schools that make networks.

Wouldn't surprise me if you would also find a lot of bugs and inaccuracies with the computational fluid dynamics module through that. That is the algorithm would find an optimal solution in simulation and when tested in real life it would turn out not to be nearly as good.

Reply on news.ycombinator.com