The first one may be solely evolutionary, and it must encode neural networks with specific 'genes'. The 'genetic coding' may start arbitrarily as it will progress towards a specific goal. Evolution will be random.

The inner will be goal-specific, and there will be possibly many networks (ensemble) like this within the higher one. Also, the inner will have a environment and fine-tuning, but it should be made generalizable as simulations or goals change. That said, use MNIST for starters.

C++ or python?

Develop FFN and simple back propagation. Keep scalability/evolution in mind

Currently: Array[int] - Each int is the number of nodes in each layer.

I dislike a neural gene that consists of the architecture. It should allow the architecture to exist, but let the architecture grow much further.

Therefore there is no need for a "neural gene". We could just a specific seed that generates a 'starter neural net', and let the neural net grow from that.

That said, neural net cannot get too big. Exploding/vanishing gradients will pretty much 'kill it'.

We do need concepts of termination to happen though.

1. Show neural net parameters and visualization
2. Show the EC parameters at each step and the neural net growth
3. As neural net grows, it may be divided into two agents that can fine tune for itself
4. It must include many of these scenarios at the each step, so figure out how parallelization is working and improve it if needed (efficiency is priority)

Akin to the concept between the "me" who wants to change and talk with the "I". Grows with EC.