function [xny,X1,Y1,A1,X2,Y2,A2,X3,Y3,A3,stplen,eval, Fsq,fnew]=...stprm3(P,T,W1,B1,f1,W2,B2,f2,W3,B3,f3,x,p,mu,f0,Fsq0,Pf,alfin, xc)%Call:% [xny,X1,Y1,A1,X2,Y2,A2,X3,Y3,A3,stplen,eval,Fsq,fnew]=...%stprm3(P,T,W1,B1,f1,W2,B2,f2,W3,B3,f3,x,p,mu,f0,Fsq0,Pf,alfin,xc)%This function computes the steplength (stplen) along the direction%p such that the objective is sufficiently reduced.%ON ENTRY:% P(r*q) holds the input data. (r input nodes and q data points)% T(s2*q) holds the target values.% W1(s1*r) holds the weights corresponding to the first hidden layer% B1(s1*1) holds the offsets for the first hidden layer% f1 is the name of the transfer function used in the first hidden layer% W2(s2*s1) holds the weights corresponding to the second hidden layer% B2(s2*1) holds the offsets for the second hidden layer% f2 is the name of the transfer function used in the second hidden layer% W3(s3*s2) holds the weights corresponding to the output layer% B3(s3*1) holds the offsets for the output layer% f3 is the name of the transfer function used in the output layer% x holds the current point packed as [W1(:); B1(:); W2(:); B2(:); W3(:); B3(:)]% p holds the search direction packed as x is packed% mu holds the current regularization parameter% f0 holds the error vector corresponding to the current point% Fsq0 holds the current value of the augmented objective% Pf holds the projection of faug onto the range of the Jacobian Jaug% alfin holds the suggested steplength% xc holds the centre of the unknowns in packed form%ON RETURN:%xny = x + stplen*p%X1 = the new value of the weights in the first hidden layer%Y1 = the new value of the biases in the first hidden layer%A1 = the new value comming out from the first hidden layer%X2 = the new value of the weights in the second hidden layer%Y2 = the new value of the biases in the second hidden layer%A2 = the new value comming out from the second hidden layer%X3 = the new value of the weights in the output layer%Y3 = the new value of the biases in the output layer%A3 = the new value comming out from the output layer%stplen = the computed steplength%eval = the number of evaluations of the actual NN inside this function%Fsq = the new value of the augmented objective%fnew = the new error value (A3-T)faug=[f0;mu*(x-xc)];v0 = faug;v1 = Pf;F0 = faug;alfa=alfin;tau=0.1;xny=x+alfa*p;nrmaug=(xny-xc)'*(xny-xc)*mu^2/2;%Don't use the suggested alfa if p is unreasonable largeindc=0;while nrmaug > Fsq0  alfa=alfa/2;  indc=indc+1;  xny=x+alfa*p;  nrmaug=(xny-xc)'*(xny-xc)*mu^2/2;end % of whileif indc >0  indc  pauseend[X1,Y1,X2,Y2,X3,Y3]=unpack3(xny,W1,B1,W2,B2,W3,B3);[A1,A2,A3]=simuff(P,X1,Y1,f1,X2,Y2,f2,X3,Y3,f3);ny=A3-T;fnew=ny;fny=ny(:);Fsq=0.5*sumsqr(fny)+nrmaug;i=1;F2 = [fny;mu*(xny-xc)];alf2= alfa;sec = 0;while (Fsq0-Fsq)<tau*(Fsq0-0.5*norm(faug+alfa*Pf)^2)  if sec == 0	FDIFF = F2-F0;	v2 = FDIFF-v1;	[alfa] = parrm(v0, v1, v2,alfa);  	xny=x+alfa*p;	[X1,Y1,X2,Y2,X3,Y3]=unpack3(xny,W1,B1,W2,B2,W3,B3);	[A1,A2,A3]=simuff(P,X1,Y1,f1,X2,Y2,f2,X3,Y3,f3);	ny=A3-T;	fnew=ny;	fny=ny(:);  	Fsq=0.5*(sumsqr(fny)+mu^2*(xny-xc)'*(xny-xc));  	i=i+1;	alf1 = alfa;	F1 = [fny;mu*(xny-xc)];  end  if sec == 1	U1 = (F1-F0) / alf1;	U2 = ((F2-F1)/(alf2-alf1) - (F1-F0)/alf1) / alf2;	v1 = U1-alf1*U2;	v2 = U2;	[alfa] = parrm(v0, v1, v2,alfa);  	xny=x+alfa*p;	[X1,Y1,X2,Y2,X3,Y3]=unpack3(xny,W1,B1,W2,B2,W3,B3);	[A1,A2,A3]=simuff(P,X1,Y1,f1,X2,Y2,f2,X3,Y3,f3);	ny=A3-T;	fnew=ny;	fny=ny(:);  	Fsq=0.5*(sumsqr(fny)+mu^2*(xny-xc)'*(xny-xc));	alf2 = alf1;	F2 = F1;	alf1 = alfa;	F1 = [fny;mu*(xny-xc)];	i = i + 1;  end  sec = 1;endeval=i;stplen=alfa;