jeffreyksmithjr / galapagos_nao Goto Github PK

To DockerHub

This gets away from the diversity/illumination goals of MAP-Elites, so this maybe needs to be a different evolutionary strategy somehow.

Use Task.Supervisor

The current approach for Gluon relies upon the Python Gluon API, but one of the large goals of the MXnet project is to be embedded in a wide range of host languages. This would probably entail lower-level code, but code remove some incidental complexity that stems from dealing with Python in an Elixir app.

The current python script muddles what is generally usable API code for things like layers and the MLP example.

Way too many parameters are embedded in the modules as a development shortcut. It's time to factor these out.

The first step is done in 2212755 and ca9c5fe which added Confex for configuration.

The next step is to allow for all learning parameters that should be user controllable to be accessible during run-time.

There's really two things going on in GN right now: neuroevolution and wrapping Gluon. All of the latter stuff could be in a smaller-scoped, more general library.

This probably starts with centralizing all parameters and structuring them in some way

Introduced in #23 . They should only be published on master builds.