.Modelica_LinearSystems2.WorkInProgress.Optimizer.constrainedLeastSquares .Modelica_LinearSystems2.WorkInProgress.Optimizer.constrainedLeastSquaresThis function solves a least squares problem with linear equality and linear inequality constraints. The problem to be solved is mathematically formulated as:
min |A*x - b|^2
subject to E*x = f;
G*x >= h;
Note, that the inequality relation is treated element wise. Remarks:
As input argument W the assembled matrices have to be given:
| E f |
W = | A b |
| G h |
With output argument mode the type of solution is characterized:
mode = 0: A solution has been computed. Both equality and
inequality constraints are compatible and have been satisfied.
= 1: A solution has been computed. Equality constraints are
contradictory. A generalized inverse solution of E*x=f was used
to minimize the residual vector length f-E*x.
In this sense, the solution is still meaningful.
> 1: Error, no solution has been computed
= 2: Inequality constraints are contradictory.
= 3: Both equality and inequality constraints are contradictory.
= 4: Usage error (input arguments are wrong).
Copyright © 2004-2012, DLR Institute of Robotics and Mechatronics
The Fortran interface of DLSEI is defined as:
SUBROUTINE DLSEI (W, MDW, ME, MA, MG, N, PRGOPT, X, RNORME,
+ RNORML, MODE, WS, IP)
***BEGIN PROLOGUE DLSEI
***PURPOSE Solve a linearly constrained least squares problem with
equality and inequality constraints, and optionally compute
a covariance matrix.
***LIBRARY SLATEC
***CATEGORY K1A2A, D9
***TYPE DOUBLE PRECISION (LSEI-S, DLSEI-D)
***KEYWORDS CONSTRAINED LEAST SQUARES, CURVE FITTING, DATA FITTING,
EQUALITY CONSTRAINTS, INEQUALITY CONSTRAINTS,
QUADRATIC PROGRAMMING
***AUTHOR Hanson, R. J., (SNLA)
Haskell, K. H., (SNLA)
***DESCRIPTION
Abstract
This subprogram solves a linearly constrained least squares
problem with both equality and inequality constraints, and, if the
user requests, obtains a covariance matrix of the solution
parameters.
Suppose there are given matrices E, A and G of respective
dimensions ME by N, MA by N and MG by N, and vectors F, B and H of
respective lengths ME, MA and MG. This subroutine solves the
linearly constrained least squares problem
EX = F, (E ME by N) (equations to be exactly
satisfied)
AX = B, (A MA by N) (equations to be
approximately satisfied,
least squares sense)
GX .GE. H,(G MG by N) (inequality constraints)
The inequalities GX .GE. H mean that every component of the
product GX must be .GE. the corresponding component of H.
In case the equality constraints cannot be satisfied, a
generalized inverse solution residual vector length is obtained
for F-EX. This is the minimal length possible for F-EX.
Any values ME .GE. 0, MA .GE. 0, or MG .GE. 0 are permitted. The
rank of the matrix E is estimated during the computation. We call
this value KRANKE. It is an output parameter in IP(1) defined
below. Using a generalized inverse solution of EX=F, a reduced
least squares problem with inequality constraints is obtained.
The tolerances used in these tests for determining the rank
of E and the rank of the reduced least squares problem are
given in Sandia Tech. Rept. SAND-78-1290. They can be
modified by the user if new values are provided in
the option list of the array PRGOPT(*).
The user must dimension all arrays appearing in the call list..
W(MDW,N+1),PRGOPT(*),X(N),WS(2*(ME+N)+K+(MG+2)*(N+7)),IP(MG+2*N+2)
where K=MAX(MA+MG,N). This allows for a solution of a range of
problems in the given working space. The dimension of WS(*)
given is a necessary overestimate. Once a particular problem
has been run, the output parameter IP(3) gives the actual
dimension required for that problem.
The parameters for DLSEI( ) are
Input.. All TYPE REAL variables are DOUBLE PRECISION
W(*,*),MDW, The array W(*,*) is doubly subscripted with
ME,MA,MG,N first dimensioning parameter equal to MDW.
For this discussion let us call M = ME+MA+MG. Then
MDW must satisfy MDW .GE. M. The condition
MDW .LT. M is an error.
The array W(*,*) contains the matrices and vectors
(E F)
(A B)
(G H)
in rows and columns 1,...,M and 1,...,N+1
respectively.
The integers ME, MA, and MG are the
respective matrix row dimensions
of E, A and G. Each matrix has N columns.
PRGOPT(*) This real-valued array is the option vector.
If the user is satisfied with the nominal
subprogram features set
PRGOPT(1)=1 (or PRGOPT(1)=1.0)
Otherwise PRGOPT(*) is a linked list consisting of
groups of data of the following form
LINK
KEY
DATA SET
The parameters LINK and KEY are each one word.
The DATA SET can be comprised of several words.
The number of items depends on the value of KEY.
The value of LINK points to the first
entry of the next group of data within
PRGOPT(*). The exception is when there are
no more options to change. In that
case, LINK=1 and the values KEY and DATA SET
are not referenced. The general layout of
PRGOPT(*) is as follows.
...PRGOPT(1) = LINK1 (link to first entry of next group)
. PRGOPT(2) = KEY1 (key to the option change)
. PRGOPT(3) = data value (data value for this change)
. .
. .
. .
...PRGOPT(LINK1) = LINK2 (link to the first entry of
. next group)
. PRGOPT(LINK1+1) = KEY2 (key to the option change)
. PRGOPT(LINK1+2) = data value
... .
. .
. .
...PRGOPT(LINK) = 1 (no more options to change)
Values of LINK that are nonpositive are errors.
A value of LINK .GT. NLINK=100000 is also an error.
This helps prevent using invalid but positive
values of LINK that will probably extend
beyond the program limits of PRGOPT(*).
Unrecognized values of KEY are ignored. The
order of the options is arbitrary and any number
of options can be changed with the following
restriction. To prevent cycling in the
processing of the option array, a count of the
number of options changed is maintained.
Whenever this count exceeds NOPT=1000, an error
message is printed and the subprogram returns.
Options..
KEY=1
Compute in W(*,*) the N by N
covariance matrix of the solution variables
as an output parameter. Nominally the
covariance matrix will not be computed.
(This requires no user input.)
The data set for this option is a single value.
It must be nonzero when the covariance matrix
is desired. If it is zero, the covariance
matrix is not computed. When the covariance matrix
is computed, the first dimensioning parameter
of the array W(*,*) must satisfy MDW .GE. MAX(M,N).
KEY=10
Suppress scaling of the inverse of the
normal matrix by the scale factor RNORM**2/
MAX(1, no. of degrees of freedom). This option
only applies when the option for computing the
covariance matrix (KEY=1) is used. With KEY=1 and
KEY=10 used as options the unscaled inverse of the
normal matrix is returned in W(*,*).
The data set for this option is a single value.
When it is nonzero no scaling is done. When it is
zero scaling is done. The nominal case is to do
scaling so if option (KEY=1) is used alone, the
matrix will be scaled on output.
KEY=2
Scale the nonzero columns of the
entire data matrix.
(E)
(A)
(G)
to have length one. The data set for this
option is a single value. It must be
nonzero if unit length column scaling
is desired.
KEY=3
Scale columns of the entire data matrix
(E)
(A)
(G)
with a user-provided diagonal matrix.
The data set for this option consists
of the N diagonal scaling factors, one for
each matrix column.
KEY=4
Change the rank determination tolerance for
the equality constraint equations from
the nominal value of SQRT(DRELPR). This quantity can
be no smaller than DRELPR, the arithmetic-
storage precision. The quantity DRELPR is the
largest positive number such that T=1.+DRELPR
satisfies T .EQ. 1. The quantity used
here is internally restricted to be at
least DRELPR. The data set for this option
is the new tolerance.
KEY=5
Change the rank determination tolerance for
the reduced least squares equations from
the nominal value of SQRT(DRELPR). This quantity can
be no smaller than DRELPR, the arithmetic-
storage precision. The quantity used
here is internally restricted to be at
least DRELPR. The data set for this option
is the new tolerance.
For example, suppose we want to change
the tolerance for the reduced least squares
problem, compute the covariance matrix of
the solution parameters, and provide
column scaling for the data matrix. For
these options the dimension of PRGOPT(*)
must be at least N+9. The Fortran statements
defining these options would be as follows:
PRGOPT(1)=4 (link to entry 4 in PRGOPT(*))
PRGOPT(2)=1 (covariance matrix key)
PRGOPT(3)=1 (covariance matrix wanted)
PRGOPT(4)=7 (link to entry 7 in PRGOPT(*))
PRGOPT(5)=5 (least squares equas. tolerance key)
PRGOPT(6)=... (new value of the tolerance)
PRGOPT(7)=N+9 (link to entry N+9 in PRGOPT(*))
PRGOPT(8)=3 (user-provided column scaling key)
CALL DCOPY (N, D, 1, PRGOPT(9), 1) (Copy the N
scaling factors from the user array D(*)
to PRGOPT(9)-PRGOPT(N+8))
PRGOPT(N+9)=1 (no more options to change)
The contents of PRGOPT(*) are not modified
by the subprogram.
The options for WNNLS( ) can also be included
in this array. The values of KEY recognized
by WNNLS( ) are 6, 7 and 8. Their functions
are documented in the usage instructions for
subroutine WNNLS( ). Normally these options
do not need to be modified when using DLSEI( ).
IP(1), The amounts of working storage actually
IP(2) allocated for the working arrays WS(*) and
IP(*), respectively. These quantities are
compared with the actual amounts of storage
needed by DLSEI( ). Insufficient storage
allocated for either WS(*) or IP(*) is an
error. This feature was included in DLSEI( )
because miscalculating the storage formulas
for WS(*) and IP(*) might very well lead to
subtle and hard-to-find execution errors.
The length of WS(*) must be at least
LW = 2*(ME+N)+K+(MG+2)*(N+7)
where K = max(MA+MG,N)
This test will not be made if IP(1).LE.0.
The length of IP(*) must be at least
LIP = MG+2*N+2
This test will not be made if IP(2).LE.0.
Output.. All TYPE REAL variables are DOUBLE PRECISION
X(*),RNORME, The array X(*) contains the solution parameters
RNORML if the integer output flag MODE = 0 or 1.
The definition of MODE is given directly below.
When MODE = 0 or 1, RNORME and RNORML
respectively contain the residual vector
Euclidean lengths of F - EX and B - AX. When
MODE=1 the equality constraint equations EX=F
are contradictory, so RNORME .NE. 0. The residual
vector F-EX has minimal Euclidean length. For
MODE .GE. 2, none of these parameters is defined.
MODE Integer flag that indicates the subprogram
status after completion. If MODE .GE. 2, no
solution has been computed.
MODE =
0 Both equality and inequality constraints
are compatible and have been satisfied.
1 Equality constraints are contradictory.
A generalized inverse solution of EX=F was used
to minimize the residual vector length F-EX.
In this sense, the solution is still meaningful.
2 Inequality constraints are contradictory.
3 Both equality and inequality constraints
are contradictory.
The following interpretation of
MODE=1,2 or 3 must be made. The
sets consisting of all solutions
of the equality constraints EX=F
and all vectors satisfying GX .GE. H
have no points in common. (In
particular this does not say that
each individual set has no points
at all, although this could be the
case.)
4 Usage error occurred. The value
of MDW is .LT. ME+MA+MG, MDW is
.LT. N and a covariance matrix is
requested, or the option vector
PRGOPT(*) is not properly defined,
or the lengths of the working arrays
WS(*) and IP(*), when specified in
IP(1) and IP(2) respectively, are not
long enough.
W(*,*) The array W(*,*) contains the N by N symmetric
covariance matrix of the solution parameters,
provided this was requested on input with
the option vector PRGOPT(*) and the output
flag is returned with MODE = 0 or 1.
IP(*) The integer working array has three entries
that provide rank and working array length
information after completion.
IP(1) = rank of equality constraint
matrix. Define this quantity
as KRANKE.
IP(2) = rank of reduced least squares
problem.
IP(3) = the amount of storage in the
working array WS(*) that was
actually used by the subprogram.
The formula given above for the length
of WS(*) is a necessary overestimate.
If exactly the same problem matrices
are used in subsequent executions,
the declared dimension of WS(*) can
be reduced to this output value.
User Designated
Working Arrays..
WS(*),IP(*) These are respectively type real
and type integer working arrays.
Their required minimal lengths are
given above.
***REFERENCES K. H. Haskell and R. J. Hanson, An algorithm for
linear least squares problems with equality and
nonnegativity constraints, Report SAND77-0552, Sandia
Laboratories, June 1978.
K. H. Haskell and R. J. Hanson, Selected algorithms for
the linearly constrained least squares problem - a
users guide, Report SAND78-1290, Sandia Laboratories,
August 1979.
K. H. Haskell and R. J. Hanson, An algorithm for
linear least squares problems with equality and
nonnegativity constraints, Mathematical Programming
21 (1981), pp. 98-118.
R. J. Hanson and K. H. Haskell, Two algorithms for the
linearly constrained least squares problem, ACM
Transactions on Mathematical Software, September 1982.
***ROUTINES CALLED D1MACH, DASUM, DAXPY, DCOPY, DDOT, DH12, DLSI,
DNRM2, DSCAL, DSWAP, XERMSG
***REVISION HISTORY (YYMMDD)
790701 DATE WRITTEN
890531 Changed all specific intrinsics to generic. (WRB)
890618 Completely restructured and extensively revised (WRB & RWC)
890831 REVISION DATE from Version 3.2
891214 Prologue converted to Version 4.0 format. (BAB)
900315 CALLs to XERROR changed to CALLs to XERMSG. (THJ)
900510 Convert XERRWV calls to XERMSG calls. (RWC)
900604 DP version created from SP version. (RWC)
920501 Reformatted the REFERENCES section. (WRB)
function constrainedLeastSquares
extends Modelica.Icons.Function;
input Real W[:, :] "System matrix = [E,f; A,b; G,h]";
input Integer n_E = 0 "Number of equations to be exactly satisfied (number of rows of matrix E)";
input Integer n_G = 0 "Number of inequality equations (number of rows of matrix G)";
input Real options[:] = {1.0} "Option vector (see info)";
output Real x[size(W, 2) - 1] "Solution vector if mode = 0 or 1";
output Integer mode "Status after completion (mode = 0/1 is successful, 2/3/4 is error)";
output Real residueNorm_A = 0 "Length of b-A*x, if mode = 0 or 1";
output Real residueNorm_E = 0 "Length of f-E*x, if mode = 0 or 1";
output Real W_out[size(W, 1), size(W, 2)] = W "Covariance matrix, if requested by options vector";
end constrainedLeastSquares;Release Notes: