Reinforcement learning is backwards because it implicitly says "when reward, behavior change"
when actually, change is the *default*, and the thing being called "reward" says "stop changing, you found it"
This figure-ground inversion is key.
The whole concept here of PCT (Perceptual Control Theory) is that brains are in some state of organization that is set up to produce a set of controlled perceptions, like getting money to buy food to not be hungry, or telling jokes to make people laugh to feel connection.
When something is perceived to be at an unwanted level (eg hunger, safety, connection, thirst, sex, meaning) the system goes into a state of reorganization, inhibiting old strategies and trying new ones.
Once the error is resolved, the music stops and the organization stays.
This is independent from whether the new strategy is actually the thing that solved the problem (which is where superstitions come from).
There's also a question of what happens when two strategies for two different needs conflict.
Often this produces oscillatory tug-of-war dynamics, but if you can gently hold all of the desiderata at once, there's usually some sort of win-win to be found.
twitter.com/nickcammarata/status/1403558910984966150One important implication of this whole thing is that you can only motivate/train someone with reward *to the extent that they are otherwise deprived of the thing* and