Neuro Evolution of Augmented Topologies

by charlieirish • 5 years ago

95 9 5 years ago

MrQuincle• 5 years ago

There's also for example HyerNEAT. I'm out of the field now. Has recent progress been made recently with these technologies?

FlyingSaucer• 5 years ago

I'm by no means an expert in the field but I do find it exceptionally interesting so i try and keep tabs on some of the research done by people who originated from the same group as Kenneth Stanley.

Seems like many from this group now pursue open-endedness in AI and view evolution as a way towards this goal (or lack thereof).

A very interesting evolution (ha!) of these ideas was presented in POET[0] towards evolution of agents in evolving environments.

There is also an interesting paper about accelerating neural architecture search when generating fake training data in generative teacher networks[1].

Lastly, a paper that i find very very interesting but might not be as relevant but still is 'First return, then explore'[2]

[0] https://eng.uber.com/poet-open-ended-deep-learning/

[1] http://proceedings.mlr.press/v119/such20a.html

[2] https://arxiv.org/pdf/2004.12919.pdf

pkoperek• 5 years ago

Apart from NEAT/HyperNEAT there are also other approaches to neuroevolution (I think in this context it is referred to as "Evolutionary Neural Architecture Search" [0]). Evolution in general can be applied in different ways (e.g. optimizing the architecture, replacing training gradient descent etc.).

A while ago I co-authored a paper in this space [1] and released some code for interested folks [2].

[0] https://arxiv.org/pdf/2008.10937.pdf

[1] https://arxiv.org/abs/1801.00119

[2] https://gitlab.com/pkoperek/pytorch-dnn-evolution/-/tree/mas...

Der_Einzige• 5 years ago

I also find the related ideas of neuroevolution of the weights of a neural networks to be fascinating in its own right.

I've implemented "cooperative coevolution" which felt like magic when I considered how good it is (on some tasks like continuous control RL problems) relative to known good methods like anything involving gradients.

I wish that this stuff was explored a bit more. Seems we are leaving the paradigm of evolutionary methods...

optimalsolver• 5 years ago

For the logical endpoint of this approach, check out FreeWire:

https://github.com/noahtren/Freewire

You can experiment with freely wired neural networks without traditional layers.

eggy2• 5 years ago

They seem to be similar to Gene Sher's TWEANNS - Topology and Weight Evolving Neural Networks - that I learned about in his 2012 book, "Handbook of Neuroevolution Through Erlang" (sure, it didn't pick up because Erlang ;))

Gene's specific implementation is DXNN (Discover and eXplore Neural Network) implemented in 2010 [1]

[1] https://arxiv.org/abs/1008.2412

dkislyuk• 5 years ago

Sher's TWEANN work cites NEAT and HyperNEAT quite extensively — Stanley's original work here is pretty influential in the neuroevolution space!

eggy2• 5 years ago

I missed that, thanks! I'll have to take another look.

When I think of neural networks that evolve changes in their topology and the weights between those connections, over and over towards optimizing a solution to function, I get the vision of the recursive Life in Life video[1] where Conway's Game of Life is emulated in Game of Life.

[1] https://www.youtube.com/watch?v=xP5-iIeKXE8

charles_f• 5 years ago

Author there, I'm curious how you found this post that was not indexed yet? :) (and not finished then)

Neuro Evolution of Augmented Topologies

Related Stories