Hand in your code and results, and answers to the questions, via Canvas. Give your les the names EBLS.m, CG FR.m, CG PRplus.m, SteepDescent.m, and comments.txt. (The last le should contain the outputs from your codes, and your written responses to the questions about the code.)
Question 5.11 from the text.
Question 6.3 from the text.
Question 6.4 from the text.
Write e cient codes for the Fletcher-Reeves nonlinear conjugate gradient algorithm, the Polak-Ribiere method with the modi cation (5.45), and the Steepest Descent method. For all methods, use the EBLS.m routine for extrapolation-bracketing line search, that you coded in an earlier assignment, to nd an approximate k that satis es the strong Wolfe conditions. Modify EBLS slightly so that its calling sequence is as follows:
function [x,alpha,nf,ng] = EBLS(fun, x, d, alpha start)
where the new output parameters nf and nf return the number of function and gradient evaluations, respectively, that were performed inside EBLS.
Terminate your nonlinear CG and steepest descent codes when the following condition is satifsied:
krf(xk)k1 nonCGparams:toler (1 + jf(xk)j);
or else when the number of iterations exceeds nonCGparams.maxit, whichever comes rst.
In your submitted codes, use the following parameter settings for StepSize:
c1 = 10 2; c2 = 0:3; maxit = 100;
and set alpha start to 1 at each call to EBLS.
Test your codes using nonlinearcg.m. Note that the function to be minimized is Powell’s singular function, coded as xpowsing.m. (These les are supplied on Canvas.)
Try modifying the line search parameters (c1, c2, and alpha start) to see if you can improve the performance of the algorithms, and comment on your experiences.
Further Details: You Matlab program SteepDescent.m should have the following rst line:
function [inform,x] = SteepDescent(fun,x,sdparams)
The inputs fun and x are as described in Homework 3, while sdparams is the following structure:
1
sdparams = struct(’maxit’,100000,’toler’,1.0e-5);
The output inform is a structure containing two elds: inform.status is 1 if the gradient tolerance is achieved and 0 if not, while inform.iter is the number of steps taken. The output x is the solution structure, with point, function, and gradient values at the nal value of xk.
Note that the global variables numf and numg are reported by the program nonlinearcg.m and incremented by the function evaluation routines. Be sure to set these to zero at the start of each of your routines for nonlinear CG and steepest descent, and to update them after each call to EBLS.
Do not print out the value of x at each iteration!
Your codes for CG PRplus.m and CG FR.m should have similar input arguments and similar outputs to SteepDescent.m.
2
