Change Genome.dna from List[int] to numpy.ndarray

[jakobj]

The dna list has homogeneous types (all entries are integers), so we might as well use a numpy array to store its contents. For example this would make operations on the whole dna, e.g., as suggested in #157, faster.

[jakobj]

i think as we make this change we can also change from a one-dimensional representation with our custom access functions to a three-dimensional structure with (rows, cols, genes) dimensions. i think this would simplify the implementation significantly. however, we'd need to think a bit about how to treat input/output nodes as their number might exceed the row count.

[HenrikMettler]

You can assign me, if you want. Imo there are 3 options:

• Do a 2 dim array: n_unit x n_genes_per_unit this has the disadvantage that levels-back is more complex to apply
• split the Dna into 3 arrays: 1 for the inputs (n_inputs x 1 x n_genes), 1 for the hidden (n_rows x n_cols x n_genes) 1 for the output (1,1,n_genes) - apart from the obvious disadvantages, this would have the advantage that we could also change the arity of input and output since they are seperate
• fill inputs and outputs with Nan values to match the dimensionality

[jakobj]

good suggestions! this is not as easy as i thought as the number of inputs, rows, and outputs can differ, so a 2d array would always have regions that are not used, as you also observe. maybe we should rethink the benefits of changing the representation of the dna. what would we gain over the current implementation by either of the three options?

[HenrikMettler]

I guess I am not telling any news that the theoretical benefits are less memory and faster compute (see eg: https://www.geeksforgeeks.org/python-lists-vs-numpy-arrays/).
But in practical if I am not mistaking only few things happen with the dna per individual per generation: Evaluation to a function (learning-rule) and replacement of some values in mutation. I don't think that in any practical use case this is computationally expensive compared to the fitness evaluation. So in essence i don't think the benefits would be large

[mschmidt87]

I think that one advantage of a representation as a multi-dimensional array would be that a lot of our code becomes easier. Especially the iter_input_regions, and other iterators become easier because we don't have to do this complicated indexing of the 1-d data structure.

[mschmidt87]

Any opinions here?

[jakobj]

in my opinion there's no need to change it from it's current form, but i would be convinced by an implementation that leads to a decrease in code complexity ;)

[jakobj]

closing for now. feel free to reopen in combination with a concrete implementation.