Very nice! Some states have multiple ancestors, some have none, so these are "dead ends" when going backwards. But when you are just looking for an approximation of the final state, as in TFA, you can probably accept quite some perturbation, which could make it significantly easier to go backwards in a naive way and sometimes miss a few cells in the process.
> Some states have multiple ancestors, some have none
Yeah, this is what I didn't understand about that particular Kaggle competition. It doesn't seem possible that you could learn some general set of rules that would allow you to predict previous states with much accuracy.
Yeah, I think it's very interesting because there's a lot of freedom in measuring the similarity and also other aspects of the process, for example optimizing for a "simple" starting configuration as mentioned in other comments.
