next up previous contents
Next: 4.1.4 Newton Direction Up: 4.1 Optimization Methods Previous: 4.1.2 Step direction


4.1.3 Steepest-descent

The modeling of the target function can be done by a linear approximation of $f(\vec{x}_0)$ based on a Taylor-series expansion of first order


\begin{displaymath}
f(\vec{x}_0 + \vec{p}) \approx
f(\vec{x}_0) + ( {\mathop{\nabla }\nolimits f(\vec{x}_0)} ) ^{\cal T} \vec{p}
\end{displaymath} (4.6)

where $\vec{p}$ denotes a small step away from the current point $\vec{x}$. A reduction of the function can be achieved by a step $\vec{p}$ where $({\mathop{\nabla }\nolimits f(\vec{x})})^{\cal T} \vec{p} $ is a negative number with a large absolute value. The direction can be found by solving the following minimization problem

\begin{displaymath}
\min_{\vec{p} \in \mathbb{R}^{n}}
\frac{({\mathop{\nabla }...
...ec{x}_0)})^{\cal T} \vec{p}}
{\vert\vert\vec{p}\vert\vert _2}
\end{displaymath} (4.7)

which gives direction $\vec{p}$


\begin{displaymath}
\vec{p} = - {\mathop{\nabla }\nolimits f(\vec{x}_0)}
\end{displaymath} (4.8)

along the negative gradient of the target function f.




R. Plasun