On the selection of optimal nonlinearities for stochastic gradient adaptive algorithms