PISO enables efficient optimization when gradients are partially unavailable by using zeroth-order information to estimate missing components, achieving state-of-the-art function-evaluation complexity without requiring bounds on the unknown gradient.
This paper introduces PISO, an optimization algorithm for problems where you can only access part of the gradient directly and must estimate the rest from function evaluations. By tracking the unknown gradient component and centering observations around this estimate, PISO reduces variance and achieves better convergence rates than existing methods.