Online Learning with Adaptive Local Step Sizes

Almeida et al. have recently proposed online algorithms for local step size adaptation in nonlinear systems trained by gradient descent. Here we develop an alternative to their approach by extending Sutton’s work on linear systems to the general, nonlinear case. The resulting algorithms are computationally little more expensive than other acceleration techniques, do not assume statistical independence between successive training patterns, and do not require an arbitrary smoothing parameter. In our benchmark experiments, they consistently outperform other acceleration methods as well as stochastic gradient descent with fixed learning rate and momentum.