Neuroevolution of Augmenting Topologies (NEAT)

This video explains the NEAT algorithm! This algorithm (published in 2001) lays the groundwork for the evolution of neural network architectures/topologies. This groundwork includes the encoding of neural network topologies, modifications to enable crossover, and the importance of minimal initialization!
Thanks for watching! Please Subscribe!

[IBMboy]

I learned about NEAT watching SethBling video on MarI/O AI, it was mind blowing watching NN and evolution working together, and i didnt know it was all the way back from 2001! really old comparing to other techniques. This video is a very clear summary.

[darkerbit]

This is the best explanation of NEAT I've ever seen!

[xeozim]

Really good intro to NEAT, I'm about to give it a go with Python and this is excellent for a short background to it

[MTEXX]

Great presentation. Like the NN itself, you presented many new terms for me to explore.

[karstenhannes]

What I don't get in the cross-over figure at 9:00 is that the excess and disjoint genes of the fittest parent are inherited. As you mentioned, in case Parent 2 is the fittest genes 3, 7 and 8 should be kept but 5 should be discarded. But in the offspring gene 5 exists! So it looks like all disjoint and excess genes are kept in the figure...

[SiOnCaa]

Thanks for this excellent explanation!


I am experimenting with NEAT for a while now, but I stumbled across this detail. You talked about the innovation numbers for connections and its importance for crossover.
Do you handle innovation numbers for the nodes too? And if so, how do you keep track of them?
My initial consideration was to remember for each node-mutation between which nodes it was created to be able to reuse the specific innovation number if another genome was about to mutate the same way.


This seemed to be a good solution but if other node-mutations are about to occur, the node won’t be located between the initial nodes it was created for.


My current setup is to handle innovation numbers of nodes locally within each genome increasing it whenever this genome created a new node.


I would be very happy to hear your thoughts on this! And again, thanks for your content!

[sudeepkuchara]

ngl i have a talk tomorrow and this is an excellent revision material

[mlwithaarvy]

What kind of explanation is this....????
What is the crossover?
9:13 No mention of how parent 2 has better performance without showing any reference. also you said 5 is discarded but its there in the offspring.

If there is a global table and you have to use it one at a time, how it could be asynchronous?
at 8:25 how do you say the crossover is given value..? which operation is being performed here? also which information is missing?

[ryandoesstuffapparently]

I am fascinated by NEAT. I’m still a total beginner, but I’m aiming to use NEAT in Unity to build a rag doll that learns how to stand up and walk and keep itself balanced.

[rigbyb]

This is a great explanation, thanks a lot :)

[JuliusFreiny]

Am I missing something? At 9:15 you're saying Parent 2 has performed better, so we keep 3, 7 and 8 and discard 5. But why is 5 shown in the offspring then? Also: Why is 2 blue and not white?

[laserbean]

The sharing function isn't 0 or 1. It's 0 or 1 - (δ/σ_share)^α

[Sharon-ySharon___e]

Scientific insights: the significance of refund information

[Destructo]

Hello, Great video man! What kind of fitness functions were they using for giving the fitness score to each genome?

[VG-hfsg]

Great Video. But please let me know: How do you solve the problem that basically with each generation, you get children with more and more neurons? If recombination is done that way, amount of neurons will increase parabolic. Given a limit <MAX> as the maximum amount of neurons a network is allowed to have, how to adapt the recombination? Thank you!

[palee]

What does sh mean at the fitness computation?

[NikopolAU]

9:15
> "and discard 5".
Proceeds to not discard 5.