Difference between revisions of "Linesearch.m"
m (→See also)
|Line 52:||Line 52:|
'', authors: [[David Goodwin]]''
Revision as of 13:58, 3 January 2017
This function finds an appropriate step length from a given search direction at a current iterate. The bracket-section line search based on fminlbfgs.m code from D. Kroon, University of Twente (Nov 2010).
[alpha,fx_1,gfx_1,exitflag,data]=linesearch(cost_function, d_0, x_0, fx_0, gfx_0, data, optim)
Performs a line search to find an appropriate step length called from within a given optimisation algorithm, i.e. fminnewton.m. This function generally assumes cheap gradients. There are currently three line-search methods from which to chose: a Newton-step (step length = 1, equivalent to no line-search); backtracking line-search; bracketing and sectioning with a cubic interpolation.
cost_function - function handle to the objective function. d_0 - current search direction x_0 - current objective (waveform) fx_0 - current objective function value at x_0 gfx_0 - current gradient at x_0 data - data structure inherited from fminnewton.m optim - optim structure inherited from fminnewton.m, contains tolerances and line-search options.
alpha - step length fx_1 - new objective function value (if alpha is not the Newton-step) gfx_1 - new gradient (if alpha is not the Newton-step) exitflag - exit flag that is empty if no errors were encountered. If an error was encountered, the codes produced are: -2 - Line search cannot find an acceptable point along the current search 2 - Change in x was smaller than the specified tolerance, optim.tol_x. 3 - Change in gradient norm less than specified tolerance, optim.tol_gfx. data - data structure with diagnostics from the line-search.
The bracket-section line-search is the default. It is also the most accurate and the most expensive - this generally assumes cheap gradient calculation (e.g. from GRAPE. The newton-step is effectively no line-search, and should only be used with a stable Newton-Raphson method. The backtracking line-search is useful and quite effective for a Newton-Raphson optimisation.
Version 1.9, authors: David Goodwin