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.
Related benchmarks
| Task | Dataset | Result | Rank | |
|---|---|---|---|---|
| Black-box Targeted Adversarial Attack | VitalDB | Success Rate (SR)100 | 9 | |
| Black-Box Targeted Adversarial Attacks | CIFAR-10 | Success Rate (SR)100 | 9 | |
| Global Optimization | Rosenbrock $d=500$ | MSE0.37 | 9 | |
| Global Optimization | Griewank | MSE3.32 | 9 | |
| Global Optimization | Ackley d = 500 | MSE5.98 | 9 | |
| Zeroth-order optimization | Theoretical Objective Functions | Iteration Complexity4 | 8 |