at this point i feel like using generic algorithms to train model weights is basically just getting a bucket of linear algebra and stirring it, hoping it will do something
i've been experimenting with genetic algorithms, neural nets, and neuroevolution. i can tell you that a lot of the academic research on crossover/recombination doesn't really work, and there is way more to be done there. i also need to put a lot more effort into reward shaping.
i've been experimenting with genetic algorithms, neural nets, and neuroevolution. i can tell you that a lot of the academic research on crossover/recombination doesn't really work, and there is way more to be done there. i also need to put a lot more effort into reward shaping.