A comment from the zoom chat (I forget who from) said something like: If the environment gives you a hash of an observation, and then the observation itself, then you have compression but not prediction.
(I.e., you can take the output and remove the hashes to compress it, but you can’t predict what’s coming next.)
A comment from the zoom chat (I forget who from) said something like: If the environment gives you a hash of an observation, and then the observation itself, then you have compression but not prediction.
(I.e., you can take the output and remove the hashes to compress it, but you can’t predict what’s coming next.)
What if I take my model’s predictions of future observations, and hash them in order of posterior probability until I run out of time or get a hit?