Poster
Stochastic Second-Order Methods Improve Best-Known Sample Complexity of SGD for Gradient-Dominated Functions
Saeed Masiha · Saber Salehkaleybar · Niao He · Negar Kiyavash · Patrick Thiran
Hall J (level 1) #826
Keywords: [ Gradient-dominated functions ] [ Reinforcement Learning ] [ Stochastic Optimization ] [ second-order methods ]
Abstract:
We study the performance of Stochastic Cubic Regularized Newton (SCRN) on a class of functions satisfying gradient dominance property with 1≤α≤2 which holds in a wide range of applications in machine learning and signal processing. This condition ensures that any first-order stationary point is a global optimum. We prove that the total sample complexity of SCRN in achieving ϵ-global optimum is O(ϵ−7/(2α)+1) for 1≤α<3/2 and ˜O(ϵ−2/(α)) for 3/2≤α≤2. SCRN improves the best-known sample complexity of stochastic gradient descent. Even under a weak version of gradient dominance property, which is applicable to policy-based reinforcement learning (RL), SCRN achieves the same improvement over stochastic policy gradient methods. Additionally, we show that the average sample complexity of SCRN can be reduced to O(ϵ−2) for α=1 using a variance reduction method with time-varying batch sizes. Experimental results in various RL settings showcase the remarkable performance of SCRN compared to first-order methods.
Chat is not available.