function [par,Par,Error,Y,iterations,x] = \
ppp_optimise(system_name,x_0,par_0,simpar,u,y_0,free,Q,extras);
## Levenberg-Marquardt optimisation for PPP/MTT
## Usage: [par,Par,Error,Y,iterations,x] = ppp_optimise(system_name,x_0,par_0,simpar,u,y_0,free[,Q,extras]);
## system_name String containing system name
## x_0 Initial state
## par_0 Initial parameter vector estimate
## simpar Simulation parameters:
## .first first time
## .dt time increment
## .stepfactor Euler integration step factor
## u System input (column vector, each row is u')
## y_0 Desired model output
## free one row for each adjustable parameter
## first column parameter indices
## second column corresponding sensitivity indices
## Q vector of positive output weights.
## extras (opt) optimisation parameters
## .criterion convergence criterion
## .max_iterations limit to number of iterations
## .v Initial Levenberg-Marquardt parameter
######################################
##### Model Transformation Tools #####
######################################
###############################################################
## Version control history
###############################################################
## $Id$
## $Log$
## Revision 1.14 2002/08/20 16:14:35 gawthrop
## Include Q in documentation
##
## Revision 1.13 2002/08/20 15:43:45 gawthrop
## Works with ident DIY rep
##
## Revision 1.12 2002/06/11 11:25:25 gawthrop
## No longer delay the simulated data.
##
## Revision 1.11 2002/05/20 13:32:36 gawthrop
## Sanity check on y_0
##
## Revision 1.10 2002/05/13 16:01:09 gawthrop
## Addes Q weighting matrix
##
## Revision 1.9 2002/05/08 10:14:21 gawthrop
## Idetification now OK (Moved data range in ppp_optimise by one sample interval)
##
## Revision 1.8 2002/04/23 17:50:39 gawthrop
## error --> err to avoid name clash with built in function
##
## Revision 1.7 2001/08/10 16:19:06 gawthrop
## Tidied up the optimisation stuff
##
## Revision 1.6 2001/07/03 22:59:10 gawthrop
## Fixed problems with argument passing for CRs
##
## Revision 1.5 2001/06/06 07:54:38 gawthrop
## Further fixes to make nonlinear PPP work ...
##
## Revision 1.4 2001/05/26 15:46:38 gawthrop
## Updated to account for new nonlinear ppp
##
## Revision 1.3 2001/04/05 11:50:12 gawthrop
## Tidied up documentation + verbose mode
##
## Revision 1.2 2001/04/04 08:36:25 gawthrop
## Restuctured to be more logical.
## Data is now in columns to be compatible with MTT.
##
## Revision 1.1 2000/12/28 11:58:07 peterg
## Put under CVS
##
###############################################################
## Copyright (C) 1999,2000,2001,2002 by Peter J. Gawthrop
## Simulation command
sim_command = sprintf("%s_ssim(x_0,par,simpar,u,i_s);", system_name);
## Extract indices
i_t = free(:,1); # Parameters
i_s = free(:,2)'; # Sensitivities
if nargin<9
extras.criterion = 1e-5;
extras.max_iterations = 10;
extras.v = 1e-5;
extras.verbose = 0;
endif
[n_data,n_y] = size(y_0);
if n_data<n_y
error("ppp_optimise: y_0 should be in columns, not rows")
endif
if nargin<8
Q = ones(n_y,1);
endif
n_th = length(i_s);
err_old = inf;
err_old_old = inf;
err = 1e50;
reduction = inf;
predicted_reduction = 0;
par = par_0;
Par = par_0;
step = ones(n_th,1);
Error = [];
Y = [];
iterations = 0;
v = extras.v; # Levenverg-Marquardt parameter.
r = 1; # Step ratio
if extras.verbose # Diagnostics
printf("Iteration: %i\n", iterations);
printf(" error: %g\n", err);
printf(" reduction: %g\n", reduction);
printf(" prediction: %g\n", predicted_reduction);
printf(" ratio: %g\n", r);
printf(" L-M param: %g\n", v);
printf(" parameters: ");
for i_th=1:n_th
printf("%g ", par(i_t(i_th)));
endfor
printf("\n");
endif
while (abs(reduction)>extras.criterion)&&\
(abs(err)>extras.criterion)&&\
(iterations<extras.max_iterations)
iterations = iterations + 1; # Increment iteration counter
[y,y_par,x] = eval(sim_command); # Simulate
[N_data,N_y] = size(y);
if (N_y!=n_y)
mess = sprintf("n_y (%i) in data not same as n_y (%i) in model", n_y,N_y);
error(mess);
endif
## Use the last part of the simulation to compare with data
## ### Removed #### And shift back by one data point
# if ( (N_data-n_data)<1 )
# error(sprintf("y_0 (%i) must be shorter than y (%i)", n_data, N_data));
# endif
y = y(N_data-n_data+1:N_data,:);
y_par = y_par(N_data-n_data+1:N_data,:);
if extras.visual==1
## Plot
title("Optimisation data");
plot([y y_0])
endif
##Evaluate error, cost derivative J and cost second derivative JJ
err = 0;
J = zeros(n_th,1);
JJ = zeros(n_th,n_th);
for i = 1:n_y
E = y(:,i) - y_0(:,i); # Error in ith output
err = err + Q(i)*(E'*E); # Sum the squared error over outputs
y_par_i = y_par(:,i:n_y:n_y*n_th); # sensitivity function (ith output)
J = J + Q(i)*y_par_i'*E; # Jacobian
JJ = JJ + Q(i)*y_par_i'*y_par_i; # Newton Euler approx Hessian
endfor
if iterations>1 # Adjust the Levenberg-Marquardt parameter
reduction = err_old-err;
predicted_reduction = 2*J'*step + step'*JJ*step;
r = predicted_reduction/reduction;
if (r<0.25)||(reduction<0)
v = 4*v;
elseif r>0.75
v = v/2;
endif
if reduction<0 # Its getting worse
par(i_t) = par(i_t) + step; # rewind parameter
err = err_old; # rewind error
err_old = err_old_old; # rewind old error
if extras.verbose
printf(" Rewinding ....\n");
endif
endif
endif
## Compute step using pseudo inverse
JJL = JJ + v*eye(n_th); # Levenberg-Marquardt term
step = pinv(JJL)*J; # Step size
par(i_t) = par(i_t) - step; # Increment parameters
err_old_old = err_old; # Save old error
err_old = err; # Save error
##Some diagnostics
Error = [Error err]; # Save error
Par = [Par par]; # Save parameters
Y = [Y y]; # Save output
if extras.verbose # Diagnostics
printf("Iteration: %i\n", iterations);
printf(" error: %g\n", err);
printf(" reduction: %g\n", reduction);
printf(" prediction: %g\n", predicted_reduction);
printf(" ratio: %g\n", r);
printf(" L-M param: %g\n", v);
printf(" parameters: ");
for i_th=1:n_th
printf("%g ", par(i_t(i_th)));
endfor
printf("\n");
endif
endwhile
endfunction