I would do the following. 1) Train a model to predict what happens next given an input 2) Each frame, predict the next frames given all possible inputs 3) Choo

I would do the following. 1) Train a model to predict what happens next given an input 2) Each frame, predict the next frames given all possible inputs 3) Choose the input that maximizes uncertainty I would expect this to learn to avoid deaths relatively quickly. It doesn’t need to be good at knowing what will happen next, just better at recognizing specific dead ends (e.g. spikes or holes).