This is the fastest evolving neural network I have so far:
https://drive.google.com/open?id=0BwsgM ... F9BTGhfa3M
Linux AMD64 only.
Quick evolve neural network AMD64
-
- Posts: 3906
- Joined: Jan 01, 2009 7:03
- Location: Australia
Re: Quick evolve neural network AMD64
And what does it evolve into?sean_vn wrote:This is the fastest evolving neural network I have so far:
I have seen examples of evolving behaviors (eg. swimming)
https://www.youtube.com/watch?v=AVU8dza89Mc
Evolving the ability to value a board game.
What is its environment? Its feedback goal?
Any utube video of it in action?
.
Re: Quick evolve neural network AMD64
The code is just infrastructure. It's not embedded in any environment for it to do anything interesting yet.
So the associative memory will just remember a number of <input,target> patterns in arrays that you show it.
The ENet version is somewhat similar but you have to give it all the examples at once, and it can begin to reason out what features in the input result in what features in the output.
The idea is to put the associative memory together with the "smarter" ENet to make some sort of artificial life (alife.)
The difficulty is an alife needs some sort of environment to live in, like a video game or something. That's too troublesome to do at the moment so just to test if a viable alife is possible I will use text prediction as a sort of environment for it to live in.
I have been sort of delaying by translating the code to JavaScript etc. Maybe next week I'll have the text prediction code.
The reason for the truncate code it to use it as a gating function for the associative memory. If the output of the truncate function is zero nothing is written to the associative memory. Any non-zero value is written. That means the smart network can output say 0.9 and with a truncate function with a magnitude diminishing value of 1, the output will be zero and nothing written to the associative memory. If the smart networks outputs 1.1 then 0.1 will be written to memory, if -1.1 then -0.1 will be written. Obviously 0.0 cannot be written to memory using such a scheme, but say 1e-30 could which is near enough.
So the associative memory will just remember a number of <input,target> patterns in arrays that you show it.
The ENet version is somewhat similar but you have to give it all the examples at once, and it can begin to reason out what features in the input result in what features in the output.
The idea is to put the associative memory together with the "smarter" ENet to make some sort of artificial life (alife.)
The difficulty is an alife needs some sort of environment to live in, like a video game or something. That's too troublesome to do at the moment so just to test if a viable alife is possible I will use text prediction as a sort of environment for it to live in.
I have been sort of delaying by translating the code to JavaScript etc. Maybe next week I'll have the text prediction code.
The reason for the truncate code it to use it as a gating function for the associative memory. If the output of the truncate function is zero nothing is written to the associative memory. Any non-zero value is written. That means the smart network can output say 0.9 and with a truncate function with a magnitude diminishing value of 1, the output will be zero and nothing written to the associative memory. If the smart networks outputs 1.1 then 0.1 will be written to memory, if -1.1 then -0.1 will be written. Obviously 0.0 cannot be written to memory using such a scheme, but say 1e-30 could which is near enough.