基于定向攻击的优化:用神经网络代理求解问题
Optimization by Directional Attacks: Solving Problems with Neural Network Surrogates
AI总结:
本文提出一种结合定向神经网络攻击算子与无导数优化技术的混合算法,用于求解含已训练非黑盒神经网络的优化问题,实验显示其性能优于标准基线。
摘要:
本文研究目标函数和约束条件包含已训练神经网络(NN)的优化问题,目标是最大化 $f(\Phi(x))$,约束为 $c(\Phi(x)) \leq 0$,其中 $f$ 是光滑函数,$c$ 是一般且非严格的函数,$\Phi$ 是已训练好的非黑盒神经网络。我们解决该问题的两个挑战:识别局部搜索的上升方向,以及确保可靠收敛到相关局部解。为此,我们将定向神经网络攻击的概念重新用作高效优化子例程,因为定向神经网络攻击旨在高效计算能将 $\Phi(x)$ 导向指定方向的 $x$ 的扰动。具体而言,我们开发了一个攻击算子,用于在任意 $x$ 处沿方向 $\nabla f(\Phi(x))$ 计算 $\Phi$ 的攻击。接着,我们提出一种混合算法,将该攻击算子与无导数优化(DFO)技术相结合,该算法因无需关注问题结构而具备数值可靠性。我们采用覆盖直接搜索方法(CDSM),在问题的温和假设下,该方法提供渐近收敛到局部解的保证。所得方法交替执行基于攻击的步骤(对当前候选解进行启发式但快速的局部改进)和 CDSM 步骤(用于保证收敛和数值可靠性)。对三个问题的实验表明,这种混合方法优于标准基线。
英文摘要:
This paper tackles optimization problems whose objective and constraints involve a trained Neural Network (NN), where the goal is to maximize $f(Φ(x))$ subject to $c(Φ(x)) \leq 0$, with $f$ smooth, $c$ general and non-stringent, and $Φ$ a NN already trained and not given as a glass-box. We address two challenges regarding this problem: identifying ascent directions for local search, and ensuring reliable convergence towards relevant local solutions. To this end, we re-purpose the notion of directional NN attacks as efficient optimization subroutines, since directional NN attacks are designed to efficiently compute perturbations of $x$ that steer $Φ(x)$ in prescribed directions. Precisely, we develop an attack operator that computes attacks of $Φ$ at any $x$ along the direction $\nabla f(Φ(x))$. Then, we propose a hybrid algorithm combining the attack operator with derivative-free optimization (DFO) techniques, designed for numerical reliability by remaining oblivious to the structure of the problem. We consider the Covering Direct Search Method (CDSM), which offers asymptotic guarantees to converge to a local solution under mild assumptions on the problem. The resulting method alternates between attack-based steps for heuristic yet fast local improvements of the current incumbent solution and CDSM steps for certified convergence and numerical reliability. Experiments on three problems show that this hybrid approach outperforms standard baselines.
机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。