They are all known to be stable, because they have a ground-truth simulator to test with. Stable doesn't necessarily mean useful, but that wasn't the point.
The benefit here is that training a neural network on simulator data allows you to generate instead of search. The simulator is very computationally expensive (even compared to a deep neural network) and the search space is large and high-dimensional.
I have done something similar using gradient descent and this differentiable vector graphics library. It converges much faster than genetic algorithms.
A good initialization, like the other commenter's voronoi idea, would speed it up considerably.
Aren't these "natural attacks" (rotating or blurring the maze, etc) just distribution shifts?
They are in-distribution for us, since we've seen objects under all sorts of optical conditions. But they're out-of-distribution for the RL model, which has only ever seen this exact maze format.
Word is that AMD support is getting better, but by far everyone is still using NVidia.
It really doesn't, because no one has any clue what Q* is or if it's even real.
TBH this sub would be a lot better if they banned OpenAI news/rumors.
I don't even mind the "I'm a noob, why won't my model train" posts - at least those people have genuine interest in ML and are trying to learn. OpenAI news attracts people who have a more science-fiction idea of AI and are just interested in the hype.
no ML technique has been shown to do anything more than just mimic statistical aspects of the training set.
Reinforcement learning does far more than mimic.
IMO interpretability and debugging are inherently related. The more you know about how the network works, the easier it will be to debug it.
you won't get published without doing proper evaluation
Idk man, I've seen some pretty sketchy papers this year.
Then you lose the 2D grid structure of the image, which is why you want to use a CNN in the first place.
I think it's possible to apply many of these optimizations to 2D convs as well though. This group is just more interested in language modeling than images.
Just built this to try on my CNN and then realized it was only for 1D convolutions. Whoops.
Model based RL is looking a little more stable in the last year. Dreamerv3 and TD-MPC2 claim to be able to train on hundreds of tasks with no per-task hyperparameter tuning, and report smooth loss curves that scale predictably.
Have to wait and see if it pans out though.