penalty.m

matlab算法集 matlab算法集

function [x,ev,j] = penalty (x0,mu,tol,m,f,p,q,dm)
%-----------------------------------------------------------------------
% Usage:       [x,ev,j] = penalty (x0,mu,tol,m,f,p,q,dm)
%
% Description: Use the self-scaled DFP method method with penalty
%              functions to solve the following n-dimensional
%              constrained optimization problem:
%
%                 minimize:   f(x)
%                 subject to: p(x) = 0
%                             q(x) >= 0
%
% Inputs:       x0  = n by 1 vector containing initial guess
%               mu  = penalty parameter (mu >= 0)
%               tol = error tolerance for terminating the search
%               m   = maximum number of iterations (m >= 1)
%               f   = string containing name of objective function: f(x)
%               p   = string containing name of equality constraint 
%                     function: p(x) = 0
%               q   = string containing name of inequality constraint 
%                     function: q(x) >= 0.  The forms of f, p, and q
%                     are:
%
%                        function y = f(x)
%                        function u = p(x)
%                        function v = q(q)
%
%                     When f is called, it must return the value of the
%                     objective function f(x). When p is called, it must
%                     take the n by 1 vector x and compute the r by 1
%                     equality constraint vector u = p(x).  When q is,
%                     called, it must take the n by 1 vector x and
%                     compute the s by 1 inequality constraint vector
%                     v = q(x).  If p = '', then no equality containts
%                     are applied in which case a user-supplied function
%                     is not required.  A similar remark holds for the
%                     inequality containts when q = ''.  Consequently, 
%                     penalty can be used to solve the following 
%                     special cases:
%
%                        ------------------------------------
%                        p = '' =>  no equality constraints
%                        q = '' =>  no inequality constraints
%                        ------------------------------------
%
%               dm  = optional display mode.  If present, 
%                     intermediate results are displayed.  
%              
% Outputs:      x  = n by 1 solution vector
%               ev = number of scalar function evaluations 
%               j  = number of iterations. If it is less than the 
%                    user-specified maximum, m, then the following
%                    convergence criterion was satisfied where em denotes
%                    the machine epsilon, and h is the step length:
%
%                       (||dF(x)/dx|| < eps) or (h*||x|| < em) 
% 
%                    Here F(x) = f(x) + mu*(P(x) + Q(x)) is the generalized
%                    objective function.   P(x) is the penalty function
%                    associated with the equality constraint, p(x) = 0, and
%                    Q(x) is the penalty function associated with the
%                    inequality constraint, q(x) >= 0.
%-----------------------------------------------------------------------
   
% Initialize

   n = length (x0);
   chkvec (x0,1,'penalty');
   mu  = args (mu,0,mu,5,'penalty');
   tol = args (tol,0,tol,6,'penalty');
   m   = args (m,1,m,7,'penalty');
   if getfun(p) 
      chkfun(feval(p,x0),6,'penalty');
   end
   if getfun(q)
      chkfun(feval(q,x0),7,'penalty');
   end
   display = nargin > 7;
   gamma = tol + 1;
   j = 0;
   k = 0;
   h = 1;
   err = 0;
   ev = 0;    
   e = 2*sqrt(eps);
   x = x0;
   dx = zeros (n,1);
   dg = zeros (n,1);
   g1 = zeros (n,1);
   g2 = zeros (n,1);
   d  = zeros (n,1);
   Q  = eye (n);

% Find optimal x

   g1 = gradfmu (f,p,q,x,mu);
   ev = ev + 2*n;
   hwbar = waitbar(0,'Computing Optimum: penalty');
   while (j < m) & (gamma > tol) & (~err)

% Perform line search along d = -Qg 

      waitbar (max(j/m,tol/gamma))
      d = -Q*g1;
      delta = norm (d,inf);

% Perform a line search along direction d 

      [a,b,c,err,eval] = bracket (delta,x,d,mu,f,p,q);
      ev = ev + eval;
      if ~err
         [h,eval] = golden (x,d,a,c,e,mu,f,p,q); 
         ev= ev + eval;  
         if display
            y = fmu (f,p,q,x,mu);
         end
         x = x + h*d;  
         dx = h*d;
      end  
         
% Compute new gradient 

     dg = g1;
     g1 = gradfmu (f,p,q,x,mu);
     ev = ev + 2*n;
     gamma = norm (g1,inf);
     dg = g1 - dg;
     if display
        fprintf ('\n(k h gamma df) = (%5i %10.3g %10.3g %10.3g)',...
                 k,h,gamma,fmu(f,p,q,x,mu)-y);
     end

% Update estimate Q of inverse of Hessian matrix 

      g2 = Q*dg;
      b1 = dot (g2,dg);
      b2 = dot (dx,dg);
      if abs(b1) > eps
         beta = b2/b1;
         Q = beta*(Q - g2*g2'/b1) + dx*dx'/b2;
      end

% Update indices, check for restart 
   
      j = j + 1;
      k = mod(k+1,n);
      if k == 0 
         Q = eye(n);
      end
   end
   close (hwbar)