Learning rates and online testing

An idea has been floating of combining the learning rate sequence with an online hypothesis test; the test would check whether the iterates \theta_n have entered a critical area around \theta_star. This could be effective with implicit methods because there is more flexibility in the learning rates.