home *** CD-ROM | disk | FTP | other *** search
- function th=bj(z,nn,maxiter,tol,lim,maxsize,Tsamp)
- %BJ Computes the prediction error estimate of a Box-Jenkins model.
- %
- % TH=bj(Z,NN)
- %
- % TH: Returned as the estimated parameters of the Box-Jenkins model
- % y(t) = [B(q)/F(q)] u(t-nk) + [C(q)/D(q)] e(t)
- % along with estimated covariances and structure information. For
- % the exact format of TH, see HELP THETA.
- %
- % Z : The output-input data Z=[y u], with y and u being column vectors.
- % This routine only works for single-input systems. Use PEM otherwise.
- %
- % NN: Initial value and structure information, given either as
- % NN=[nb nc nd nf nk], the orders and delay of the above model,
- % or as NN=THI, with THI being a theta matrix of standard format.
- % In the latter case the criterion minimization is initialized at THI.
- %
- % Some parameters associated with the algorithm are accessed by
- % TH = bj(Z,NN,maxiter,tol,lim,maxsize,T)
- % See HELP AUXVAR for an explanation of these and their default values.
-
- % L. Ljung 10-1-86,11-2-91
- % Copyright (c) 1986-92 by the MathWorks, Inc.
- % All Rights Reserved.
-
- % *** Set up default values ***
- [Ncap,nz]=size(z);
- maxsdef=idmsize(Ncap);
-
- if nargin<7, Tsamp=[];end
- if nargin<6, maxsize=[];end
- if nargin<5, lim=[];end
- if nargin<4, tol=[];, end
- if nargin<3, maxiter=[]; end
- Tdef=0;if isempty(Tsamp),Tsamp=1;Tdef=1;end
- if isempty(maxsize),maxsize=maxsdef;end
- if isempty(lim),lim=1.6;end,if isempty(tol),tol=0.01;end
- if isempty(maxiter),maxiter=10;end
- if Tsamp<0,Tsamp=1;end, if maxsize<0,maxsize=maxsdef;end, if lim<0,lim=1.6;end
- if tol<0,tol=0.01;end,if maxiter<0, maxiter=10;end
-
- % *** Do some consistency tests ***
-
- [nr,cc]=size(nn); nu=nz-1;
- if nz>Ncap, error('The data should be organized in column vectors')
- return,end
- if nu>1, error('This routine only works for single-input systems. Use PEM instead!')
- return,end
- if nu==0, error('For time series, use ARMAX instead!'),
- return,end
-
- if nr==1 & cc~=5,
- error('Incorrect number of orders specified: nn should be nn=[nb nc nd nf nk]')
- return,end
-
- % *** if nn has more than 1 row, it is a theta matrix
- % and we jump directly to the minimization ***
- %
- if nr>1, nu=nn(1,3);
- if nu>1, error('This routine only works for single input systems. Use PEM instead')
- return,end
- if nu==0, error('This routine is only meaningful for a SISO system. Use ARMAX instead!'),end
- na=nn(1,4);nb=nn(1,5);nc=nn(1,6);nd=nn(1,7);
- nf=nn(1,8);nk=nn(1,9);
- if na>0,error('na should be zero for this routine!'),
- return,end
- n=nf+nb+nc+nd; t=nn(3,1:n)';ni=max([na+nd nd+nb+nk-1 nc+nf]);
- if nc>0, c=[1 t(nb+1:nb+nc)'];else c=1;,end
- if nf>0,f=[1 t(nb+nc+nd+1:n)'];else f=1;,end
- if nb>0,b=[zeros(1,nk) t(na+1:na+nb)'];,else b=0;,end
- if nd>0,d=[1 t(nb+nc+1:nb+nc+nd)'];else d=1;end
- w=filter(b,f,z(:,2));v=z(:,1)-w;e=pefilt(d,c,v,zeros(1,ni));
- if Tdef,Tsamp=gett(nn);end
- end
- if nr==1,
- %
- nb=nn(1); nc=nn(2); nd=nn(3); nf=nn(4);
- nk=nn(5); n=sum(nn(1:4));ni=max([na+nd nd+nb+nk-1 nc+nf]);
- %
- % *** compute initial estimate via LS and IV ***
- %
- t=arx(z,[nf nb nk],maxsize,Tsamp,0);
- f=fstab([1 t(1:nf)']);
- t=iv(z,[nf nb nk],f,[zeros(1,nk) t(nf+1:nf+nb)'],maxsize,Tsamp,0);
- f=fstab([1 t(1:nf)']);t(1:nf)=f(2:nf+1)';
- %
- % *** Determine the initial C- and D-estimates by first building a
- % high order AR model of the residuals, and then
- % using the LS-method ***
- %
- w=pefilt([zeros(1,nk) t(nf+1:nf+nb)'],f,z(:,2),z(1:ni,1));
- v=z(:,1)-w;
- c=1; t2=[];
- if nc>0
- t1=arx(v,n,maxsize,Tsamp,0);
- e=pefilt([1 t1'],1,v);
- t2=arx([v e],[nd nc 1],maxsize,Tsamp,0);
- %
- % ** test the stability of the C-estimate **
- %
- c=fstab([1 t2(nd+1:nd+nc)']); t2(nd+1:nd+nc)=c(2:nc+1)';
- end
- %
- if nc==0
- t2=arx(v,nd,maxsize,Tsamp,0);
- end
-
- if nf>0, tf=t(1:nf);else tf=[];end
- if nc>0, tc=t2(nd+1:nd+nc);else tc=[];end
- if nd>0, td=t2(1:nd);else td=[];end
- t=[t(nf+1:nf+nb);tc;td;tf];
-
- %
- if nd==0, d=1; else d=[1 t2(1:nd)'];end
- e=pefilt(d,c,v,zeros(1,ni));
- end
- % *** Display initial estimate ***
- %
- Vcap=e'*e/(Ncap-ni);
- clc, disp([' INITIAL ESTIMATE'])
- disp(['Current loss:' num2str(Vcap)])
- disp(['th-vector'])
- disp(t)
- %
- % *** start minimizing ***
- %
- % ** determine limit for robustification **
- if lim~=0, lim=median(abs(e-median(e)))*lim/0.7;end
- if lim==0, lim=1000*Vcap;end
- [me,ne]=size(e);
- ll=ones(me,ne)*lim;el=max(min(e,ll),-ll); clear ll
- g=ones(n,1); l=0; st=0; nmax=max([nf nb+nk-1 nc nd]);
- h=conv(f,c);
- %
- % ** the minimization loop **
- %
- while [norm(g)>tol l<maxiter st~=1]
- l=l+1;
- % * compute gradient *
- wf=filter(-d,h,w); ef=filter(1,c,e);
- vf=filter(-1,c,v); uf=filter(d,h,z(:,2));
- M=floor(maxsize/n);
- R=zeros(n);Fcap=zeros(n,1);
- for k1=nmax:M:Ncap-1
- jj=(k1+1:min(Ncap,k1+M));
- psi=zeros(length(jj),n);
- for k=1:nb, psi(:,k)=uf(jj-k-nk+1);end
- for k=1:nc, psi(:,nb+k)=ef(jj-k);end
- for k=1:nd, psi(:,k+nb+nc)=vf(jj-k);end
- for k=1:nf, psi(:,k+nb+nc+nd)=wf(jj-k);end
- if Ncap>M, R=R+psi'*psi; Fcap=Fcap+psi'*el(jj);end
- end
- if Ncap>M, g=R\Fcap; else g=psi\el(jj);end
- %
- % * search along the g-direction *
- %
-
- [t1,e,el,v,w,V1,c,d,h,st]=searchbj(z,t,g,lim,Vcap,nb,nc,nd,nf,nk,ni);
- home
- disp([' ITERATION # ' int2str(l)])
- disp(['Current loss: ' num2str(V1) ' Previous loss: ' num2str(Vcap)])
- disp(['Current th prev. th gn-dir'])
- disp([t1 t g])
- disp(['Norm of gn-vector: ' num2str(norm(g))])
- if st==1
- disp(['No improvement of the criterion possible along the search direction'])
- disp(['Iterations therefore terminated']),end
- t=t1; Vcap=V1;
- end
- th=zeros(3+n,max([9 n]));
- Vcap=e'*e/(length(e)-ni);
- th(1,1:9)=[Vcap Tsamp 1 0 nb nc nd nf nk];
- th(2,1)=Vcap*(1+n/Ncap)/(1-n/Ncap);
- ti=fix(clock); ti(1)=ti(1)/100;
- th(2,2:6)=ti(1:5);
- th(2,7)=2;
- th(3,1:n)=t';
- if maxiter==0,th(2,7)=2;return,end
- if Ncap>M, PP=inv(R); else PP=inv(psi'*psi);end
- th(4:3+n,1:n)=Vcap*PP;
-
-