On Fri, Aug 27, 2010 at 05:23:05PM -0400, Bob Walton wrote:
> The nmsmax.m function (Nelder-Mead maximization) in the Optim-1.0.12
> package contains three coding errors. Attached is a diff -c file with
> fixes. These fixes reduce the number of function evaluations for my
> test functions by about 40% and also greatly reduces the number of times
> the procedure gets "stuck" at non-optimum points. The fixes conform the
> function to the algorithm published by Nelder and Mead, and also cause
> the routine's results to agree with other implementations of the
> algorithm.
> --
> Bob Walton
> *** nmsmaxorig.m Sun Aug 20 09:37:58 2006
> --- nmsmax.m Thu Aug 26 12:36:46 2010
> ***************
> *** 154,160 ****
> x(:) = ve;
> fe = dirn*feval(fun,x,varargin{:});
> nf = nf + 1;
> ! if fe > f(1)
> vk = ve; fk = fe;
> how = 'expand, ';
> end
> --- 154,160 ----
> x(:) = ve;
> fe = dirn*feval(fun,x,varargin{:});
> nf = nf + 1;
> ! if fe > fr
> vk = ve; fk = fe;
> how = 'expand, ';
> end
> ***************
> *** 168,178 ****
> x(:) = vc;
> fc = dirn*feval(fun,x,varargin{:});
> nf = nf + 1;
> ! if fc > f(n)
> vk = vc; fk = fc;
> how = 'contract,';
> else
> ! for j = 2:n
> V(:,j) = (V(:,1) + V(:,j))/2;
> x(:) = V(:,j);
> f(j) = dirn*feval(fun,x,varargin{:});
> --- 168,178 ----
> x(:) = vc;
> fc = dirn*feval(fun,x,varargin{:});
> nf = nf + 1;
> ! if fc > f(n+1)
> vk = vc; fk = fc;
> how = 'contract,';
> else
> ! for j = 2:n+1
> V(:,j) = (V(:,1) + V(:,j))/2;
> x(:) = V(:,j);
> f(j) = dirn*feval(fun,x,varargin{:});
> ------------------------------------------------------------------------------
nmsmax.m seems to be "foreign" code, only slightly modified for usage
in Octave. If you know the nms algorithm well, I think we should just
accept your correction. But I think for verification you should post
one or more of your test-problems which work better with the
correction.
Olaf
|
