Where communities thrive


  • Join over 1.5M+ people
  • Join over 100K+ communities
  • Free without limits
  • Create your own community
People
Activity
lagleki
@lagleki
Have you checked other libs implementing neat? Presumably python ones?
Manuel Raimann
@raimannma

Have you checked other libs implementing neat? Presumably python ones?

We are using the implementation from code bullet

Manuel Raimann
@raimannma
correct
we are also looking at his games where he uses that lib
lagleki
@lagleki
Any plans on finalizing the lib? Or no ultimate goal?
Manuel Raimann
@raimannma
I think there are many more steps we can go with NEAT
We have algorithms like DeepNEAT and CoDeepNEAT, but also we can construct GAN architectures and more
Automatic Hyperparameter tuning .......
lagleki
@lagleki
Gating is taken from neataptic? the original NEAT doesnt seem to have it
Manuel Raimann
@raimannma
Yes, correct
Actually, gating is already implemented in synaptic
lagleki
@lagleki
Do you have tests? Is lstm structure naturally appear with certain datasets?
Manuel Raimann
@raimannma
very rarely with the instinct algorithm
i expect it to appear with full neat, but currently we don't have gates at the full neat part integrated. First we need to finish the standard full-neat.
lagleki
@lagleki
Natural birth of lstm is a very noble goal. Thanks for letting me know that instinct doesn't produce lstm
lagleki
@lagleki
Hi. Are parametrized activation functions any good? Like relu with learnable cap value instead of just fixed relu6?
Manuel Raimann
@raimannma
I have a question, why is the library called "carrot" ?
@lagleki I don't think so, because our evolution process doesn't use any learning rate. For backpropagation though it could be very useful, but i think our main goal should be improving evolution.
Manuel Raimann
@raimannma

There is a new library we are working on: https://github.com/raimannma/MatrixNEAT

We develop on gitlab, but cloning it automatically to github.
If you like to contribute on gitlab, just message me.

There is a whole new data structure we want to use and currently it works perfect and is much easier to handle.

Instead of adjacency lists, as carrot, neataptic, synaptic, ... use. We want to use Adjacency matrices, see here: https://en.wikipedia.org/wiki/Adjacency_matrix

This takes a little bit more memory space, but is much faster.
We can save a whole network in just one NxN matrix, where N is the number of Nodes.

Christian Echevarria
@chrisgereina
It's pretty impressive how quick it runs and how small the code footprint is. It's going to be really interesting to see how it grows into its potential
Manuel Raimann
@raimannma
and it should be much easier to make use of gpu.js
This message was deleted
or other performance optimiztion libraries
Manuel Raimann
@raimannma
@chrisgereina We could rewrite the topological sort algorithm of the matrices so that it can give us information about which nodes can be activated in parallel, and as activation of nodes is just a Dotproduct it should be easily parallelizable
currently we are searching for only one column with zeros in the matrix, but we could search for all and activate them in parallel
Christian Echevarria
@chrisgereina

@chrisgereina We could rewrite the topological sort algorithm of the matrices so that it can give us information about which nodes can be activated in parallel, and as activation of nodes is just a Dotproduct it should be easily parallelizable

I think this is a really good idea, but we should stay disciplined and not try to optimize prematurely

My own premature optimization is at least part of what made working on the full-neat branch so painful hahaha
@raimannma I had a thought today, since we meet every sunday to do a development session I thought maybe we could start recording them for the community
DM me on what you think about it
Manuel Raimann
@raimannma
We could choose another platform for our meetings, so it's simple to join for everyone else

@raimannma I had a thought today, since we meet every sunday to do a development session I thought maybe we could start recording them for the community

We could do that, but I don't think out community is so interested into that.

We could change to discord or to google meet

My own premature optimization is at least part of what made working on the full-neat branch so painful hahaha

Yeah, first of all we should get a fully functional version of NEAT, and enough tests, that we cannot break something with our optimizations

Manuel Raimann
@raimannma
Here is the new library: https://gitlab.com/my_neat/matrix-neat
lagleki
@lagleki
@raimannma anything like that in python?
Manuel Raimann
@raimannma
No, I can't find something like that already exisiting
Maybe we can port this over to python once it is functional
Manuel Raimann
@raimannma
Maybe if we have someone here who is good at visualization stuff, would be great to have a library that takes in an adjacency matrix and outputs an image of the corrosponding graph. This would make debugging much easier.
Or maybe there is already a library that supports that.
Manuel Raimann
@raimannma

@lagleki Here you have a link to our meeting: https://meet.google.com/ewu-ecrb-mcy

We will start at 7 pm in germany.
I don't know your timezone, but for San Fracisco it is at 10 AM.
We would be happy if you would come by

lagleki
@lagleki
@raimannma ok so where did you get this algorithm? Any papers?
Manuel Raimann
@raimannma
My own thinking :-)
it's just the same algorithm as the original NEAT paper
but with another way to express graphs

There are two main ways how to store a graph with an object oriented programming language.

Adjacency lists
Adjacency matrices

other libraries (and carrot) use lists, containing objects which are referenced to each other
@lagleki If you want to know more about adjacency matrices you can watch some videos or just read the article on wikipedia
Christian Echevarria
@chrisgereina

Hi everyone, we set up a recurring meeting event on Zoom (you can use this any Sunday):

chrisgereina is inviting you to a scheduled Zoom meeting.

Join Zoom Meeting
https://us02web.zoom.us/j/88387393479?pwd=Mk5tSzVXTDBTZ1Q2Mm84WnAwcjlLUT09

Meeting ID: 883 8739 3479
Passcode: NeatMatrix
One tap mobile
+16699009128,,88387393479#,,,,4251945189# US (San Jose)
+13462487799,,88387393479#,,,,
4251945189# US (Houston)

Dial by your location
+1 669 900 9128 US (San Jose)
+1 346 248 7799 US (Houston)
+1 253 215 8782 US (Tacoma)
+1 301 715 8592 US (Washington D.C)
+1 312 626 6799 US (Chicago)
+1 646 558 8656 US (New York)
Meeting ID: 883 8739 3479
Passcode: 4251945189
Find your local number: https://us02web.zoom.us/u/keIQcsbGjt

Christian Echevarria
@chrisgereina

Hey everyone -- we're moving to Discord to make it easier to message the community: https://discord.gg/P4FJG8rEYC

Please come find us on the new server!