Partially Informed Stochastic Optimization via Zeroth-Order Residual Tracking
This research introduces Partially Informed Stochastic Optimization (PISO) to solve problems where full gradients are unavailable, such as in black-box environments. The method works by decomposing the gradient into known and unknown parts and tracking a residual vector. By centering function evaluations around an estimate of this residual, PISO transforms gradient estimation into an error-feedback contraction, achieving state-of-the-art convergence rates with minimal function calls.