Reya now uses neuroevolution to fit tasks

After my last reya-related devlog, i have made the decision to ditch training data entirely and use either:

1. imitation reinforcement training

or

2. neuroevolution

i have decided to use option 2 for now, although it may be limited for the complex tasks this AI will end up having to do. a combination of bothg methods could be useful (e.g. neuroevolution to create modules, imitation training to use those modules together). though note that i would prefer pure evolution, as it is amazingly simple.

after using a basic genetic algorithm (stolen from a previous project) and failing spectacularly at optimizing the task (the task being predicting the sum of sin(t) where t ranges from 0 to x, at 25 values of x, unifromly distributed in the range 0 to pi), i tried to use gradient descent with approximated gradients, but that also didnt work (even with RMSprop), most likely because the cost function uses discrete points with a relatively large step size (~0.12566).

then, i tried using differential evolution. although it is intended for continuous functions, it worked surprisignly very well (relatively, anyway).

i am eyeing the covariance matrix adaptation evolution strategy as a more powerful alternative.