Hello,
I am a hobby AI researcher (so feel free to question the validity of all this), and I am designing a system which can adapt to be any type of neural network. It's a blank slate, the dynamics of the system are all encoded into genes (as the weights of a standard feedforward neural network, which updates some memory variables as well as the neuron output). It is then evolved to produce the most reward-seeking network. I plan to start on simple tests such as XOR, pole balancing, and mountain car.
The standard feed-foward neural networks are universal function approximators, so theoretically they can produce any neural network within the limitations of the data the neural networks operate on (their memory variables, their allowed connectivity).
Right now I have planned to evolve the synaptic update, the activation function (it might end up being spiking, it might not), the connector (decides when neurons to connect/disconnect), an input encoder (takes a single float as input and feeds it to the net) and a decoder (reverse of encoder).
Has anybody ever thought of this before? Surely someone has, but I can't find anything on the web. Just wanted to share this. If someone makes something out of this, at least I feel like I contributed somehow. I will let you all know how it goes. Also, ideas on how to improve the system are welcome.
[link][2 comments]