Share your thoughts, 1 month free Claude Pro on usSee more
WorkDL logo mark

Global Optimization with A Power-Transformed Objective and Gaussian Smoothing

About

We propose a novel method that solves global optimization problems in two steps: (1) perform a (exponential) power-$N$ transformation to the not-necessarily differentiable objective function $f$ and get $f_N$, and (2) optimize the Gaussian-smoothed $f_N$ with stochastic approximations. Under mild conditions on $f$, for any $\delta>0$, we prove that with a sufficiently large power $N_\delta$, this method converges to a solution in the $\delta$-neighborhood of $f$'s global optimum point. The convergence rate is $O(d^2\sigma^4\varepsilon^{-2})$, which is faster than both the standard and single-loop homotopy methods if $\sigma$ is pre-selected to be in $(0,1)$. In most of the experiments performed, our method produces better solutions than other algorithms that also apply smoothing techniques.

Chen Xu• 2024

Related benchmarks

TaskDatasetResultRank
Black-box Targeted Adversarial AttackVitalDB
Success Rate (SR)100
9
Black-Box Targeted Adversarial AttacksCIFAR-10
Success Rate (SR)100
9
Global OptimizationRosenbrock $d=500$
MSE0.37
9
Global OptimizationGriewank
MSE3.32
9
Global OptimizationAckley d = 500
MSE5.98
9
Zeroth-order optimizationTheoretical Objective Functions
Iteration Complexity4
8
Showing 6 of 6 rows

Other info

Follow for update