mpen_update

 Penalty for matrix constraints update

0001 % Penalty for matrix constraints update
0002 function  [] = mpen_update(obj)
0003 
0004   % add barrier penalty update as well...
0005   idx=obj.Yboxindbar;
0006   pnew=obj.PYbox;   % to get the right shape
0007   pnew(idx)=p_update(obj.PYbox(idx), obj.allopts.mpenalty_update, obj.allopts.mpenalty_border, obj.allopts.mpenalty_min);
0008 
0009   if ~isempty(obj.Yboxindbar)
0010   for k=obj.Yboxindbar
0011     Ykx = obj.Y{obj.Yboxmap(k)};
0012     Akx=obj.Yboxshift(k)*speye(size(Ykx)) + obj.Yboxmlt(k)*Ykx;
0013 
0014     %pkx=pnew(k); %obj.PA(k);
0015     [pnew(k), nfactor] = p_check(-Akx, pnew(k), obj.PYbox(k));
0016   end
0017   end
0018   obj.PYbox(idx) = max(pnew(idx));
0019 
0020   idx=obj.Yboxindphi;
0021   pnew=obj.PYbox;   % to get the right shape
0022   pnew(idx)=p_update(obj.PYbox(idx), obj.allopts.mpenalty_update, obj.allopts.mpenalty_border, obj.allopts.mpenalty_min);
0023 
0024   if ~isempty(obj.Yboxindphi)
0025   for k=obj.Yboxindphi
0026     Ykx = obj.Y{obj.Yboxmap(k)};
0027     Akx=obj.Yboxshift(k)*speye(size(Ykx)) + obj.Yboxmlt(k)*Ykx;
0028 
0029     %pkx=pnew(k); %obj.PA(k);
0030     [pnew(k), nfactor] = p_check(-Akx, pnew(k), obj.PYbox(k));
0031   end
0032   end
0033   obj.PYbox(idx) = max(pnew(idx));
0034 
0035 
0036   idx=obj.Aindphi;
0037   %pold=obj.PA;
0038   pnew=obj.PA;   % to get the right shape
0039   pnew(idx)=p_update(obj.PA(idx), obj.allopts.mpenalty_update, obj.allopts.mpenalty_border, obj.allopts.mpenalty_min);
0040      % p, pfactor, pborder, pmin
0041 
0042   % check that each matrix constraints is still feasible with the new penalty parameter
0043   if ~isempty(obj.Aindphi)
0044   for k=obj.Aindphi
0045     kuser=obj.Amap(k);
0046     [Akuserx, obj.userdata] = obj.mconfun(obj.x, obj.Y, kuser, obj.userdata);
0047     Akx = obj.Ashift(k)*speye(size(Akuserx)) + obj.Amlt(k) .* Akuserx;
0048 
0049     %pkx=pnew(k); %obj.PA(k);
0050     [pnew(k), nfactor] = p_check(-Akx, pnew(k), obj.PA(k));
0051   end
0052   end
0053 
0054   % use the minima for all
0055   obj.PA(idx) = max(pnew(idx));
0056 
0057 end
0058 
0059 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
0060 % penalty vector update according to the rules
0061 %   p - vector of original penalty parameters
0062 %   pfactor - update factor for "normal" operation
0063 %   pborder - where to switch to slower update
0064 %   pmin - minimal penalty which will be used
0065 % OR do it up there??
0066 function [p] = p_update(p, pfactor, pborder, pmin)
0067   idxnormal = p>pborder;
0068 
0069   p(idxnormal) = p(idxnormal)*pfactor;
0070   p(~idxnormal) = p(~idxnormal)*0.9;
0071 
0072   % don't go below pmin
0073   p = max(p,pmin);
0074 
0075 end
0076 
0077 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
0078 % check if the new penalty parameter is usable for the matrix
0079 % constraints, i.e., check that M+p*I >=0 (positive semidefinite)
0080 %   M - matrix of the constraint (M>=0) at the current point
0081 %   p - new penalty parameter
0082 %   pold - previous penalty parameter (to backtrack)
0083 % returns
0084 %   p - valid penalty argument
0085 %   nfactor - how many times it was necessary to factorize
0086 %     (nfactor>1 ==> p had to be shortened)
0087 function [p, nfactor] = p_check(M, p, pold)
0088 
0089   rFactor=0.75;   % if need to refactorize, prefer new penalty parameter
0090   nfactor=0;
0091 
0092   [n m] = size(M);
0093   Missparse = n>10 && issparse(M) && nnz(M)<0.15*n*n;  
0094 
0095   if (Missparse)
0096     perm=amd(M);
0097     M=M(perm,perm);
0098     I=speye(n,n);
0099   else
0100     % it is usually faster to compute with dense matrices in dense format
0101     M=full(M);
0102     I=eye(n,n);
0103   end
0104 
0105   [R,k] = chol(M+p*I);
0106   nfactor=nfactor+1;
0107   if (k==0)
0108     % first match -> go
0109     return;
0110   end
0111 
0112   while (k~=0)
0113     p=rFactor*p + (1-rFactor)*pold;
0114     [R,k] = chol(M+p*I);
0115     %disp(sprintf('up   pert=%e (%i)',p,k));
0116     nfactor=nfactor+1;
0117     % add save bounds??
0118   end
0119 
0120 end
0121 
0122