知识卡片
梯度下降的凸优化误差界
若f凸且可微、存在全局最优点x*,初始距离R=∥x⁽⁰⁾−x*∥>0,沿途梯度范数≤L且L>0;
T≥1时取γ=R/(L√T),前T点均值的目标差f(x̄)−f(x*)≤RL/√T。
此界针对平均输出。
这张卡片看不懂?
这张卡片的反馈已满 10 条,暂时无法继续提交。
已有反馈 · 0 / 10 条
还没有人提交反馈。
在整书笔记中的位置
原书出处 · 1 处
- 第33章 机器学习算法 · 查看原书依据
The analysis of GRADIENT-DESCENT is summarized in the following theorem. Theorem 33.8 Let x* ∈ ℝn be the minimizer of a convex function f, and suppose that an execution of GRADIENT-DESCENT(f, x(0), γ, T) returns x-avg, where and R and L are defined in equations (33.20) and (33.21). Let . Then we have f(x-avg) – f(x*) ≤ ϵ. ▪ We now prove this theorem. We do not give an absolute bound on how much progress each iteration makes. Instead, we use a potential function, as in Section 16.3. Here, we define a potential Φ(t) after computing x(t), such that Φ(t) ≥ 0 for t = 0,
参考来源
- 位置:《算法导论(第4版)》第33章《机器学习算法》(OEBPS/chapter033.xhtml,文本字符[50000,55000))。
- 结论依据:第33.3节式(33.20) Art_P1385 定义 R=∥x⁽⁰⁾−x*∥,式(33.21) Art_P1386 假设沿途∥∇f(x)∥≤L;定理33.8 的 Art_P1387 给出 γ=R/(L√T),Art_P1388 给出 ε=RL/√T,且 f(x̄)−f(x*)≤ε。R,L>0 保证步长式有定义,T≥1。
- 原始内容:The analysis of G RADIENT -D ESCENT is summarized in the following theorem. Theorem 33.8 Let x * ∈ ℝ n be the minimizer of a convex function f , and suppose that an execution of G RADIENT -D ESCENT ( f , x (0) , γ , T ) returns x-avg , where and R and L are defined in equations (33.20) and (33.21). Let . Then we have f ( x-avg ) – f ( x *) ≤ ϵ . ▪ We now prove this theorem. We do not give an absolute bound on how much progress each iteration makes. Instead, we use a potential function, as in Section 16.3 . Here, we define a potential Φ( t ) after computing x ( t ) , such that Φ( t ) ≥ 0 for t = 0, …, T . We define the amortized progress in the iteration that computes x (