function result = dfp_min(f,pin,info,varargin) % PURPOSE: DFP minimization routine to minimize func % (Converted from Numerical Recipes book dfpmin routine) %--------------------------------------------------- % USAGE: result = dfp_min(func,b,info,varargin) % or result = dfp_min(func,b,[],varargin) % (for the case of default options) % Where: func = likelihood function to minimize % b = parameter vector fed to func % info = a structure with fields: % info.maxit = maximum # of iterations (default = 500) % info.tol = tolerance for convergence (default = sqrt(eps)) % info.prt = 1 for printing iterations, 0 for no printing % varargin = list of arguments passed to function %--------------------------------------------------- % RETURNS: a result structure % result.b = (kx1) minimizing vector % result.f = value of func at solution values % result.hess = hessian at solution values % result.iter = # number of iterations % result.time = time taken % result.flag = 0 for convergence, 1 for non-convergence %--------------------------------------------------- % NOTE: func must take the form func(b,varargin) % where: b = parameter vector (k x 1) % varargin = arguments passed to the function %--------------------------------------------------- % SEE ALSO: maxlik, qnewton, pow_min, frpr_min functions % NOTE: calls linmin(), gradnt(), hessian() %--------------------------------------------------- % written by: % James P. LeSage, Dept of Economics % University of Toledo % 2801 W. Bancroft St, % Toledo, OH 43606 % jlesage@spatial-econometrics.com % default options tol = sqrt(eps);; maxit = 500; pflag = 0; if length(info) > 0 if ~isstruct(info) error('dfp_min: options should be in a structure variable'); end; % parse options fields = fieldnames(info); nf = length(fields); xcheck = 0; ycheck = 0; for i=1:nf if strcmp(fields{i},'maxit') maxit = info.maxit; elseif strcmp(fields{i},'tol') tol = info.tol; elseif strcmp(fields{i},'prt') pflag = info.prt; end; end; else % rely on defaults end; funfcn = fcnchk(f,length(varargin)); epss = sqrt(eps); [n junk] = size(pin); xarg = pin; deltagrad = tol; fp = feval(funfcn,xarg,varargin{:}); g = gradnt(funfcn,xarg,deltagrad,varargin{:}); xi = -g'; hessn = eye(n); iter = 1; % these are for the case of printing intermediate results during iteration input.cnames = strvcat('iteration','function value'); input.fmt = strvcat('%5d','%16.8f'); Vname = 'Parameter'; for i=1:n tmp = ['Parameter ',num2str(i)]; Vname = strvcat(Vname,tmp); end; pinf.cnames = strvcat('Estimates','Gradient'); pinf.rnames = Vname; t0 = clock; while (iter <= maxit) [pin fret] = linmin(pin,xi,tol,f,varargin{:}); if pflag == 1 % print iteration results matprt = [iter fp]; mprint(matprt,input); matprt = [pin xi]; mprint(matprt,pinf); end; if ((2*abs(fp-fret)) <= tol*(abs(fp) + abs(fret)+epss)); pout = pin; fout = fret; hess = fdhess(funfcn,xarg,varargin{:}); niter = iter; result.b = pin; result.f = -fret; result.hess = fdhess(funfcn,pout,varargin{:}); result.iter = niter; result.time = etime(clock,t0); result.flag = 0; return; end; xarg = pin; fp = feval(funfcn,xarg,varargin{:}); dg = g'; xarg = pin; fret = feval(funfcn,xarg,varargin{:}); g = gradnt(funfcn,xarg,deltagrad,varargin{:}); dg = g' - dg; hdg = hessn*dg; fac = dg'*xi; fae = dg'*hdg; if (fac == 0) pout = pin; fout = fret; hess = fdhess(funfcn,xarg,varargin{:}); niter = iter; result.b = pin; result.f = -fret; result.hess = fdhess(funfcn,pout,varargin{:}); result.iter = niter; result.time = etime(clock,t0); result.flag = 0; return; else fac = 1/fac; end; if (fae == 0) pout = pin; fout = fret; hess = hessn; niter = iter; result.b = pin; result.f = -fret; result.hess = fdhess(funfcn,pout,varargin{:}); result.iter = niter; result.time = etime(clock,t0); result.flag = 0; return; else fad = 1/fae; end; dg = fac*xi - fad*hdg; hessn = hessn + fac*xi*xi' - fad*hdg*hdg' + fae*dg*dg'; xi = -hessn*g'; iter = iter + 1; end; result.b = pin; result.f = -fret; result.hess = fdhess(funfcn,pout,varargin{:}); result.iter = niter; result.time = etime(clock,t0); result.flag = 1;