Marking recurrent connections gets very slow once the creature phenotypes get more complicated (see image below). Need to find a way to optimise this part if possible.

Navigating to the end point /users/foo leads to a nice error screen and gives the following:

sqlalchemy.exc.DataError

sqlalchemy.exc.DataError: (psycopg2.DataError) invalid input syntax for integer: "foo"
LINE 3: WHERE "user".id = 'foo'
^
[SQL: 'SELECT "user".id AS user_id, "user".username AS user_username, "user".email AS user_email \nFROM "user" \nWHERE "user".id = %(id_1)s \n LIMIT %(param_1)s'] [parameters: {'id_1': 'foo', 'param_1': 1}] (Background on this error at: http://sqlalche.me/e/9h9h)

It should handle this a bit more gracefully, e.g. returning a 404 or something.

To start, refer to page 55 of the pdf for the paper.

For example, on a quad core system one could run four genetic algorithms in parallel and then choose the best of the four.

Given an input column vector x, a weight column vector w, we can express the output vector y in a graph structure.
Working with just the vectors, and using padding in the convolution we get conv(x, w) = [w₃*x₁ + w₂*0 + w₁*0,
w₃*x₂ + w₂*x₁ + w₁*0,
w₃*x₃ + w₂*x₂ + w₁*x₁,
w₃*x₄ + w₂*x₃ + w₁*x₂,
w₃*x₅ + w₂*x₄ + w₁*x₃,
w₃*0 + w₂*x₅ + w₁*x₄,
w₃*0 + w₂*0 + w₁*x₅]

Figure: the above convolution in a graph structure.

So, essentially the main two constraints are:

1. The repeating structure of similar connections
2. the same set of weights are used over all the connections.

There may also be two main approaches when trying to introduce convolution into a creature's neural structure:

1. Try to engineer a way to insert the convolutional structure into a creature's genome.
2. Allow it to happen naturally, and try to detect the convolutional structures to see if any creatures developed it.

Some problems with approach (1) are:

• How do you mutate the weights?
• Do you make the connections explicitly target multiple inputs and outputs?
• What if a random mutation happens and the structure is modified and no longer a pure convolutional structure?

Some problems with approach (2) are:

• How do you detect a convolutional structure?
• How long will it take for such a structure to appear?
• Will the structure be stable (i.e. not get changed such that it is no longer a convolutional structure)?

How is the forward pass computed for the neural networks represented in the genes?