Periodically returning a learning process to an initial state with random timing to accelerate optimization.