0
$\begingroup$

I have the following optimization problem:

$$\underset{\alpha\in\mathbb{R}}{\text{min}}:\;\;f(\textbf{x}+\alpha\textbf{d})$$ $$\text{subject to}:\;\;0\leq\alpha\leq \alpha_{max},$$

where $f(\textbf{x})=\frac{1}{2}\textbf{x}^T\textbf{Q}\textbf{x}+\textbf{c}^T\textbf{x}, \;\textbf{x}\in\mathbb{R}^n, \;\textbf{d}\in\mathbb{R}^n, \;\textbf{c}\in\mathbb{R}^n,\;$ $\alpha_{max}\in \mathbb{R}$ and $\textbf{Q}$ is $n\times n$ symmetric positive semi-definite real matrix. I'm trying to solve a general support vector machine - classifier parameters by using active set method with gradient projection. I'm almost done with my formulations but my last task is to solve the optimization problem I presented above. How should I solve for optimal $\alpha$?

$\endgroup$
1
$\begingroup$

In reduced gradient - active set methods (RG-ASM): $\alpha^* \in [0, \alpha_{max}]$ and this $\alpha^*$ must satisfy Armijo-Wolfe conditions (sufficient decrease condition and curvature condition).

$f(x^k + \alpha d^k) \le f(x^k) + \alpha c_1\nabla f(x^k)d^k\;$ where $c_1 \in (0, 1)$

$\nabla f(x^k + \alpha d^k)^T d^k \ge c_2\nabla f(x^k)d^k\;$ where $0 < c1 < c2 < 1$

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.