*DECK DSPLP
SUBROUTINE DSPLP (DUSRMT, MRELAS, NVARS, COSTS, PRGOPT, DATTRV,
+ BL, BU, IND, INFO, PRIMAL, DUALS, IBASIS, WORK, LW, IWORK, LIW)
C***BEGIN PROLOGUE DSPLP
C***PURPOSE Solve linear programming problems involving at
C most a few thousand constraints and variables.
C Takes advantage of sparsity in the constraint matrix.
C***LIBRARY SLATEC
C***CATEGORY G2A2
C***TYPE DOUBLE PRECISION (SPLP-S, DSPLP-D)
C***KEYWORDS LINEAR CONSTRAINTS, LINEAR OPTIMIZATION,
C LINEAR PROGRAMMING, LP, SPARSE CONSTRAINTS
C***AUTHOR Hanson, R. J., (SNLA)
C Hiebert, K. L., (SNLA)
C***DESCRIPTION
C
C These are the short usage instructions; for details about
C other features, options and methods for defining the matrix
C A, see the extended usage instructions which are contained in
C the Long Description section below.
C
C |------------|
C |Introduction|
C |------------|
C The subprogram DSPLP( ) solves a linear optimization problem.
C The problem statement is as follows
C
C minimize (transpose of costs)*x
C subject to A*x=w.
C
C The entries of the unknowns x and w may have simple lower or
C upper bounds (or both), or be free to take on any value. By
C setting the bounds for x and w, the user is imposing the con-
C straints of the problem. The matrix A has MRELAS rows and
C NVARS columns. The vectors costs, x, and w respectively
C have NVARS, NVARS, and MRELAS number of entries.
C
C The input for the problem includes the problem dimensions,
C MRELAS and NVARS, the array COSTS(*), data for the matrix
C A, and the bound information for the unknowns x and w, BL(*),
C BU(*), and IND(*). Only the nonzero entries of the matrix A
C are passed to DSPLP( ).
C
C The output from the problem (when output flag INFO=1) includes
C optimal values for x and w in PRIMAL(*), optimal values for
C dual variables of the equations A*x=w and the simple bounds
C on x in DUALS(*), and the indices of the basic columns,
C IBASIS(*).
C
C |------------------------------|
C |Fortran Declarations Required:|
C |------------------------------|
C
C DIMENSION COSTS(NVARS),PRGOPT(*),DATTRV(*),
C *BL(NVARS+MRELAS),BU(NVARS+MRELAS),IND(NVARS+MRELAS),
C *PRIMAL(NVARS+MRELAS),DUALS(MRELAS+NVARS),IBASIS(NVARS+MRELAS),
C *WORK(LW),IWORK(LIW)
C
C EXTERNAL DUSRMT
C
C The dimensions of PRGOPT(*) and DATTRV(*) must be at least 1.
C The exact lengths will be determined by user-required options and
C data transferred to the subprogram DUSRMT( ).
C
C The values of LW and LIW, the lengths of the arrays WORK(*)
C and IWORK(*), must satisfy the inequalities
C
C LW .GE. 4*NVARS+ 8*MRELAS+LAMAT+ LBM
C LIW.GE. NVARS+11*MRELAS+LAMAT+2*LBM
C
C It is an error if they do not both satisfy these inequalities.
C (The subprogram will inform the user of the required lengths
C if either LW or LIW is wrong.) The values of LAMAT and LBM
C nominally are
C
C LAMAT=4*NVARS+7
C and LBM =8*MRELAS
C
C LAMAT determines the length of the sparse matrix storage area.
C The value of LBM determines the amount of storage available
C to decompose and update the active basis matrix.
C
C |------|
C |Input:|
C |------|
C
C MRELAS,NVARS
C ------------
C These parameters are respectively the number of constraints (the
C linear relations A*x=w that the unknowns x and w are to satisfy)
C and the number of entries in the vector x. Both must be .GE. 1.
C Other values are errors.
C
C COSTS(*)
C --------
C The NVARS entries of this array are the coefficients of the
C linear objective function. The value COSTS(J) is the
C multiplier for variable J of the unknown vector x. Each
C entry of this array must be defined.
C
C DUSRMT
C ------
C This is the name of a specific subprogram in the DSPLP( ) package
C used to define the matrix A. In this usage mode of DSPLP( )
C the user places the nonzero entries of A in the
C array DATTRV(*) as given in the description of that parameter.
C The name DUSRMT must appear in a Fortran EXTERNAL statement.
C
C DATTRV(*)
C ---------
C The array DATTRV(*) contains data for the matrix A as follows:
C Each column (numbered J) requires (floating point) data con-
C sisting of the value (-J) followed by pairs of values. Each pair
C consists of the row index immediately followed by the value
C of the matrix at that entry. A value of J=0 signals that there
C are no more columns. The required length of
C DATTRV(*) is 2*no. of nonzeros + NVARS + 1.
C
C BL(*),BU(*),IND(*)
C ------------------
C The values of IND(*) are input parameters that define
C the form of the bounds for the unknowns x and w. The values for
C the bounds are found in the arrays BL(*) and BU(*) as follows.
C
C For values of J between 1 and NVARS,
C if IND(J)=1, then X(J) .GE. BL(J); BU(J) is not used.
C if IND(J)=2, then X(J) .LE. BU(J); BL(J) is not used.
C if IND(J)=3, then BL(J) .LE. X(J) .LE. BU(J),(BL(J)=BU(J) ok)
C if IND(J)=4, then X(J) is free to have any value,
C and BL(J), BU(J) are not used.
C
C For values of I between NVARS+1 and NVARS+MRELAS,
C if IND(I)=1, then W(I-NVARS) .GE. BL(I); BU(I) is not used.
C if IND(I)=2, then W(I-NVARS) .LE. BU(I); BL(I) is not used.
C if IND(I)=3, then BL(I) .LE. W(I-NVARS) .LE. BU(I),
C (BL(I)=BU(I) is ok).
C if IND(I)=4, then W(I-NVARS) is free to have any value,
C and BL(I), BU(I) are not used.
C
C A value of IND(*) not equal to 1,2,3 or 4 is an error. When
C IND(I)=3, BL(I) must be .LE. BU(I). The condition BL(I).GT.
C BU(I) indicates infeasibility and is an error.
C
C PRGOPT(*)
C ---------
C This array is used to redefine various parameters within DSPLP( ).
C Frequently, perhaps most of the time, a user will be satisfied
C and obtain the solutions with no changes to any of these
C parameters. To try this, simply set PRGOPT(1)=1.D0.
C
C For users with more sophisticated needs, DSPLP( ) provides several
C options that may be used to take advantage of more detailed
C knowledge of the problem or satisfy other utilitarian needs.
C The complete description of how to use this option array to
C utilize additional subprogram features is found under the
C heading of DSPLP( ) Subprogram Options in the Extended
C Usage Instructions.
C
C Briefly, the user should note the following value of the parameter
C KEY and the corresponding task or feature desired before turning
C to that document.
C
C Value Brief Statement of Purpose for Option
C of KEY
C ------ -------------------------------------
C 50 Change from a minimization problem to a
C maximization problem.
C 51 Change the amount of printed output.
C Normally, no printed output is obtained.
C 52 Redefine the line length and precision used
C for the printed output.
C 53 Redefine the values of LAMAT and LBM that
C were discussed above under the heading
C Fortran Declarations Required.
C 54 Redefine the unit number where pages of the sparse
C data matrix A are stored. Normally, the unit
C number is 1.
C 55 A computation, partially completed, is
C being continued. Read the up-to-date
C partial results from unit number 2.
C 56 Redefine the unit number where the partial results
C are stored. Normally, the unit number is 2.
C 57 Save partial results on unit 2 either after
C maximum iterations or at the optimum.
C 58 Redefine the value for the maximum number of
C iterations. Normally, the maximum number of
C iterations is 3*(NVARS+MRELAS).
C 59 Provide DSPLP( ) with a starting (feasible)
C nonsingular basis. Normally, DSPLP( ) starts
C with the identity matrix columns corresponding
C to the vector w.
C 60 The user has provided scale factors for the
C columns of A. Normally, DSPLP( ) computes scale
C factors that are the reciprocals of the max. norm
C of each column.
C 61 The user has provided a scale factor
C for the vector costs. Normally, DSPLP( ) computes
C a scale factor equal to the reciprocal of the
C max. norm of the vector costs after the column
C scaling for the data matrix has been applied.
C 62 Size parameters, namely the smallest and
C largest magnitudes of nonzero entries in
C the matrix A, are provided. Values noted
C outside this range are to be considered errors.
C 63 Redefine the tolerance required in
C evaluating residuals for feasibility.
C Normally, this value is set to RELPR,
C where RELPR = relative precision of the arithmetic.
C 64 Change the criterion for bringing new variables
C into the basis from the steepest edge (best
C local move) to the minimum reduced cost.
C 65 Redefine the value for the number of iterations
C between recalculating the error in the primal
C solution. Normally, this value is equal to ten.
C 66 Perform "partial pricing" on variable selection.
C Redefine the value for the number of negative
C reduced costs to compute (at most) when finding
C a variable to enter the basis. Normally this
C value is set to NVARS. This implies that no
C "partial pricing" is used.
C 67 Adjust the tuning factor (normally one) to apply
C to the primal and dual error estimates.
C 68 Pass information to the subprogram DFULMT(),
C provided with the DSPLP() package, so that a Fortran
C two-dimensional array can be used as the argument
C DATTRV(*).
C 69 Pass an absolute tolerance to use for the feasibility
C test when the usual relative error test indicates
C infeasibility. The nominal value of this tolerance,
C TOLABS, is zero.
C
C
C |---------------|
C |Working Arrays:|
C |---------------|
C
C WORK(*),LW,
C IWORK(*),LIW
C ------------
C The arrays WORK(*) and IWORK(*) are respectively floating point
C and type INTEGER working arrays for DSPLP( ) and its
C subprograms. The lengths of these arrays are respectively
C LW and LIW. These parameters must satisfy the inequalities
C noted above under the heading "Fortran Declarations Required:"
C It is an error if either value is too small.
C
C |----------------------------|
C |Input/Output files required:|
C |----------------------------|
C
C Fortran unit 1 is used by DSPLP( ) to store the sparse matrix A
C out of high-speed memory. A crude
C upper bound for the amount of information written on unit 1
C is 6*nz, where nz is the number of nonzero entries in A.
C
C |-------|
C |Output:|
C |-------|
C
C INFO,PRIMAL(*),DUALS(*)
C -----------------------
C The integer flag INFO indicates why DSPLP( ) has returned to the
C user. If INFO=1 the solution has been computed. In this case
C X(J)=PRIMAL(J) and W(I)=PRIMAL(I+NVARS). The dual variables
C for the equations A*x=w are in the array DUALS(I)=dual for
C equation number I. The dual value for the component X(J) that
C has an upper or lower bound (or both) is returned in
C DUALS(J+MRELAS). The only other values for INFO are .LT. 0.
C The meaning of these values can be found by reading
C the diagnostic message in the output file, or by looking for
C error number = (-INFO) in the Extended Usage Instructions
C under the heading:
C
C List of DSPLP( ) Error and Diagnostic Messages.
C
C BL(*),BU(*),IND(*)
C ------------------
C These arrays are output parameters only under the (unusual)
C circumstances where the stated problem is infeasible, has an
C unbounded optimum value, or both. These respective conditions
C correspond to INFO=-1,-2 or -3. See the Extended
C Usage Instructions for further details.
C
C IBASIS(I),I=1,...,MRELAS
C ------------------------
C This array contains the indices of the variables that are
C in the active basis set at the solution (INFO=1). A value
C of IBASIS(I) between 1 and NVARS corresponds to the variable
C X(IBASIS(I)). A value of IBASIS(I) between NVARS+1 and NVARS+
C MRELAS corresponds to the variable W(IBASIS(I)-NVARS).
C
C *Long Description:
C
C SUBROUTINE DSPLP(DUSRMT,MRELAS,NVARS,COSTS,PRGOPT,DATTRV,
C * BL,BU,IND,INFO,PRIMAL,DUALS,IBASIS,WORK,LW,IWORK,LIW)
C
C |------------|
C |Introduction|
C |------------|
C The subprogram DSPLP( ) solves a linear optimization problem.
C The problem statement is as follows
C
C minimize (transpose of costs)*x
C subject to A*x=w.
C
C The entries of the unknowns x and w may have simple lower or
C upper bounds (or both), or be free to take on any value. By
C setting the bounds for x and w, the user is imposing the con-
C straints of the problem.
C
C (The problem may also be stated as a maximization
C problem. This is done by means of input in the option array
C PRGOPT(*).) The matrix A has MRELAS rows and NVARS columns. The
C vectors costs, x, and w respectively have NVARS, NVARS, and
C MRELAS number of entries.
C
C The input for the problem includes the problem dimensions,
C MRELAS and NVARS, the array COSTS(*), data for the matrix
C A, and the bound information for the unknowns x and w, BL(*),
C BU(*), and IND(*).
C
C The output from the problem (when output flag INFO=1) includes
C optimal values for x and w in PRIMAL(*), optimal values for
C dual variables of the equations A*x=w and the simple bounds
C on x in DUALS(*), and the indices of the basic columns in
C IBASIS(*).
C
C |------------------------------|
C |Fortran Declarations Required:|
C |------------------------------|
C
C DIMENSION COSTS(NVARS),PRGOPT(*),DATTRV(*),
C *BL(NVARS+MRELAS),BU(NVARS+MRELAS),IND(NVARS+MRELAS),
C *PRIMAL(NVARS+MRELAS),DUALS(MRELAS+NVARS),IBASIS(NVARS+MRELAS),
C *WORK(LW),IWORK(LIW)
C
C EXTERNAL DUSRMT (or 'NAME', if user provides the subprogram)
C
C The dimensions of PRGOPT(*) and DATTRV(*) must be at least 1.
C The exact lengths will be determined by user-required options and
C data transferred to the subprogram DUSRMT( ) ( or 'NAME').
C
C The values of LW and LIW, the lengths of the arrays WORK(*)
C and IWORK(*), must satisfy the inequalities
C
C LW .GE. 4*NVARS+ 8*MRELAS+LAMAT+ LBM
C LIW.GE. NVARS+11*MRELAS+LAMAT+2*LBM
C
C It is an error if they do not both satisfy these inequalities.
C (The subprogram will inform the user of the required lengths
C if either LW or LIW is wrong.) The values of LAMAT and LBM
C nominally are
C
C LAMAT=4*NVARS+7
C and LBM =8*MRELAS
C
C These values will be as shown unless the user changes them by
C means of input in the option array PRGOPT(*). The value of LAMAT
C determines the length of the sparse matrix "staging" area.
C For reasons of efficiency the user may want to increase the value
C of LAMAT. The value of LBM determines the amount of storage
C available to decompose and update the active basis matrix.
C Due to exhausting the working space because of fill-in,
C it may be necessary for the user to increase the value of LBM.
C (If this situation occurs an informative diagnostic is printed
C and a value of INFO=-28 is obtained as an output parameter.)
C
C |------|
C |Input:|
C |------|
C
C MRELAS,NVARS
C ------------
C These parameters are respectively the number of constraints (the
C linear relations A*x=w that the unknowns x and w are to satisfy)
C and the number of entries in the vector x. Both must be .GE. 1.
C Other values are errors.
C
C COSTS(*)
C --------
C The NVARS entries of this array are the coefficients of the
C linear objective function. The value COSTS(J) is the
C multiplier for variable J of the unknown vector x. Each
C entry of this array must be defined. This array can be changed
C by the user between restarts. See options with KEY=55,57 for
C details of checkpointing and restarting.
C
C DUSRMT
C ------
C This is the name of a specific subprogram in the DSPLP( ) package
C that is used to define the matrix entries when this data is passed
C to DSPLP( ) as a linear array. In this usage mode of DSPLP( )
C the user gives information about the nonzero entries of A
C in DATTRV(*) as given under the description of that parameter.
C The name DUSRMT must appear in a Fortran EXTERNAL statement.
C Users who are passing the matrix data with DUSRMT( ) can skip
C directly to the description of the input parameter DATTRV(*).
C Also see option 68 for passing the constraint matrix data using
C a standard Fortran two-dimensional array.
C
C If the user chooses to provide a subprogram 'NAME'( ) to
C define the matrix A, then DATTRV(*) may be used to pass floating
C point data from the user's program unit to the subprogram
C 'NAME'( ). The content of DATTRV(*) is not changed in any way.
C
C The subprogram 'NAME'( ) can be of the user's choice
C but it must meet Fortran standards and it must appear in a
C Fortran EXTERNAL statement. The first statement of the subprogram
C has the form
C
C SUBROUTINE 'NAME'(I,J,AIJ, INDCAT, PRGOPT, DATTRV, IFLAG)
C
C The variables I,J, INDCAT, IFLAG(10) are type INTEGER,
C while AIJ, PRGOPT(*),DATTRV(*) are type REAL.
C
C The user interacts with the contents of IFLAG(*) to
C direct the appropriate action. The algorithmic steps are
C as follows.
C
C Test IFLAG(1).
C
C IF(IFLAG(1).EQ.1) THEN
C
C Initialize the necessary pointers and data
C for defining the matrix A. The contents
C of IFLAG(K), K=2,...,10, may be used for
C storage of the pointers. This array remains intact
C between calls to 'NAME'( ) by DSPLP( ).
C RETURN
C
C END IF
C
C IF(IFLAG(1).EQ.2) THEN
C
C Define one set of values for I,J,AIJ, and INDCAT.
C Each nonzero entry of A must be defined this way.
C These values can be defined in any convenient order.
C (It is most efficient to define the data by
C columns in the order 1,...,NVARS; within each
C column define the entries in the order 1,...,MRELAS.)
C If this is the last matrix value to be
C defined or updated, then set IFLAG(1)=3.
C (When I and J are positive and respectively no larger
C than MRELAS and NVARS, the value of AIJ is used to
C define (or update) row I and column J of A.)
C RETURN
C
C END IF
C
C END
C
C Remarks: The values of I and J are the row and column
C indices for the nonzero entries of the matrix A.
C The value of this entry is AIJ.
C Set INDCAT=0 if this value defines that entry.
C Set INDCAT=1 if this entry is to be updated,
C new entry=old entry+AIJ.
C A value of I not between 1 and MRELAS, a value of J
C not between 1 and NVARS, or a value of INDCAT
C not equal to 0 or 1 are each errors.
C
C The contents of IFLAG(K), K=2,...,10, can be used to
C remember the status (of the process of defining the
C matrix entries) between calls to 'NAME'( ) by DSPLP( ).
C On entry to 'NAME'( ), only the values 1 or 2 will be
C in IFLAG(1). More than 2*NVARS*MRELAS definitions of
C the matrix elements is considered an error because
C it suggests an infinite loop in the user-written
C subprogram 'NAME'( ). Any matrix element not
C provided by 'NAME'( ) is defined to be zero.
C
C The REAL arrays PRGOPT(*) and DATTRV(*) are passed as
C arguments directly from DSPLP( ) to 'NAME'( ).
C The array PRGOPT(*) contains any user-defined program
C options. In this usage mode the array DATTRV(*) may
C now contain any (type REAL) data that the user needs
C to define the matrix A. Both arrays PRGOPT(*) and
C DATTRV(*) remain intact between calls to 'NAME'( )
C by DSPLP( ).
C Here is a subprogram that communicates the matrix values for A,
C as represented in DATTRV(*), to DSPLP( ). This subprogram,
C called DUSRMT( ), is included as part of the DSPLP( ) package.
C This subprogram 'decodes' the array DATTRV(*) and defines the
C nonzero entries of the matrix A for DSPLP( ) to store. This
C listing is presented here as a guide and example
C for the users who find it necessary to write their own subroutine
C for this purpose. The contents of DATTRV(*) are given below in
C the description of that parameter.
C
C SUBROUTINE DUSRMT(I,J,AIJ, INDCAT,PRGOPT,DATTRV,IFLAG)
C DIMENSION PRGOPT(*),DATTRV(*),IFLAG(10)
C
C IF(IFLAG(1).EQ.1) THEN
C
C THIS IS THE INITIALIZATION STEP. THE VALUES OF IFLAG(K),K=2,3,4,
C ARE RESPECTIVELY THE COLUMN INDEX, THE ROW INDEX (OR THE NEXT COL.
C INDEX), AND THE POINTER TO THE MATRIX ENTRY'S VALUE WITHIN
C DATTRV(*). ALSO CHECK (DATTRV(1)=0.) SIGNIFYING NO DATA.
C IF(DATTRV(1).EQ.0.) THEN
C I = 0
C J = 0
C IFLAG(1) = 3
C ELSE
C IFLAG(2)=-DATTRV(1)
C IFLAG(3)= DATTRV(2)
C IFLAG(4)= 3
C END IF
C
C RETURN
C ELSE
C J=IFLAG(2)
C I=IFLAG(3)
C L=IFLAG(4)
C IF(I.EQ.0) THEN
C
C SIGNAL THAT ALL OF THE NONZERO ENTRIES HAVE BEEN DEFINED.
C IFLAG(1)=3
C RETURN
C ELSE IF(I.LT.0) THEN
C
C SIGNAL THAT A SWITCH IS MADE TO A NEW COLUMN.
C J=-I
C I=DATTRV(L)
C L=L+1
C END IF
C
C AIJ=DATTRV(L)
C
C UPDATE THE INDICES AND POINTERS FOR THE NEXT ENTRY.
C IFLAG(2)=J
C IFLAG(3)=DATTRV(L+1)
C IFLAG(4)=L+2
C
C INDCAT=0 DENOTES THAT ENTRIES OF THE MATRIX ARE ASSIGNED THE
C VALUES FROM DATTRV(*). NO ACCUMULATION IS PERFORMED.
C INDCAT=0
C RETURN
C END IF
C END
C
C DATTRV(*)
C ---------
C If the user chooses to use the provided subprogram DUSRMT( ) then
C the array DATTRV(*) contains data for the matrix A as follows:
C Each column (numbered J) requires (floating point) data con-
C sisting of the value (-J) followed by pairs of values. Each pair
C consists of the row index immediately followed by the value
C of the matrix at that entry. A value of J=0 signals that there
C are no more columns. (See "Example of DSPLP( ) Usage," below.)
C The dimension of DATTRV(*) must be 2*no. of nonzeros
C + NVARS + 1 in this usage. No checking of the array
C length is done by the subprogram package.
C
C If the Save/Restore feature is in use (see options with
C KEY=55,57 for details of checkpointing and restarting)
C DUSRMT( ) can be used to redefine entries of the matrix.
C The matrix entries are redefined or overwritten. No accum-
C ulation is performed.
C Any other nonzero entry of A, defined in a previous call to
C DSPLP( ), remain intact.
C
C BL(*),BU(*),IND(*)
C ------------------
C The values of IND(*) are input parameters that define
C the form of the bounds for the unknowns x and w. The values for
C the bounds are found in the arrays BL(*) and BU(*) as follows.
C
C For values of J between 1 and NVARS,
C if IND(J)=1, then X(J) .GE. BL(J); BU(J) is not used.
C if IND(J)=2, then X(J) .LE. BU(J); BL(J) is not used.
C if IND(J)=3, then BL(J) .LE. X(J) .LE. BU(J),(BL(J)=BU(J) ok)
C if IND(J)=4, then X(J) is free to have any value,
C and BL(J), BU(J) are not used.
C
C For values of I between NVARS+1 and NVARS+MRELAS,
C if IND(I)=1, then W(I-NVARS) .GE. BL(I); BU(I) is not used.
C if IND(I)=2, then W(I-NVARS) .LE. BU(I); BL(I) is not used.
C if IND(I)=3, then BL(I) .LE. W(I-NVARS) .LE. BU(I),
C (BL(I)=BU(I) is ok).
C if IND(I)=4, then W(I-NVARS) is free to have any value,
C and BL(I), BU(I) are not used.
C
C A value of IND(*) not equal to 1,2,3 or 4 is an error. When
C IND(I)=3, BL(I) must be .LE. BU(I). The condition BL(I).GT.
C BU(I) indicates infeasibility and is an error. These
C arrays can be changed by the user between restarts. See
C options with KEY=55,57 for details of checkpointing and
C restarting.
C
C PRGOPT(*)
C ---------
C This array is used to redefine various parameters within DSPLP( ).
C Frequently, perhaps most of the time, a user will be satisfied
C and obtain the solutions with no changes to any of these
C parameters. To try this, simply set PRGOPT(1)=1.D0.
C
C For users with more sophisticated needs, DSPLP( ) provides several
C options that may be used to take advantage of more detailed
C knowledge of the problem or satisfy other utilitarian needs.
C The complete description of how to use this option array to
C utilize additional subprogram features is found under the
C heading "Usage of DSPLP( ) Subprogram Options."
C
C Briefly, the user should note the following value of the parameter
C KEY and the corresponding task or feature desired before turning
C to that section.
C
C Value Brief Statement of Purpose for Option
C of KEY
C ------ -------------------------------------
C 50 Change from a minimization problem to a
C maximization problem.
C 51 Change the amount of printed output.
C Normally, no printed output is obtained.
C 52 Redefine the line length and precision used
C for the printed output.
C 53 Redefine the values of LAMAT and LBM that
C were discussed above under the heading
C Fortran Declarations Required.
C 54 Redefine the unit number where pages of the sparse
C data matrix A are stored. Normally, the unit
C number is 1.
C 55 A computation, partially completed, is
C being continued. Read the up-to-date
C partial results from unit number 2.
C 56 Redefine the unit number where the partial results
C are stored. Normally, the unit number is 2.
C 57 Save partial results on unit 2 either after
C maximum iterations or at the optimum.
C 58 Redefine the value for the maximum number of
C iterations. Normally, the maximum number of
C iterations is 3*(NVARS+MRELAS).
C 59 Provide DSPLP( ) with a starting (feasible)
C nonsingular basis. Normally, DSPLP( ) starts
C with the identity matrix columns corresponding
C to the vector w.
C 60 The user has provided scale factors for the
C columns of A. Normally, DSPLP( ) computes scale
C factors that are the reciprocals of the max. norm
C of each column.
C 61 The user has provided a scale factor
C for the vector costs. Normally, DSPLP( ) computes
C a scale factor equal to the reciprocal of the
C max. norm of the vector costs after the column
C scaling for the data matrix has been applied.
C 62 Size parameters, namely the smallest and
C largest magnitudes of nonzero entries in
C the matrix A, are provided. Values noted
C outside this range are to be considered errors.
C 63 Redefine the tolerance required in
C evaluating residuals for feasibility.
C Normally, this value is set to the value RELPR,
C where RELPR = relative precision of the arithmetic.
C 64 Change the criterion for bringing new variables
C into the basis from the steepest edge (best
C local move) to the minimum reduced cost.
C 65 Redefine the value for the number of iterations
C between recalculating the error in the primal
C solution. Normally, this value is equal to ten.
C 66 Perform "partial pricing" on variable selection.
C Redefine the value for the number of negative
C reduced costs to compute (at most) when finding
C a variable to enter the basis. Normally this
C value is set to NVARS. This implies that no
C "partial pricing" is used.
C 67 Adjust the tuning factor (normally one) to apply
C to the primal and dual error estimates.
C 68 Pass information to the subprogram DFULMT(),
C provided with the DSPLP() package, so that a Fortran
C two-dimensional array can be used as the argument
C DATTRV(*).
C 69 Pass an absolute tolerance to use for the feasibility
C test when the usual relative error test indicates
C infeasibility. The nominal value of this tolerance,
C TOLABS, is zero.
C
C
C |---------------|
C |Working Arrays:|
C |---------------|
C
C WORK(*),LW,
C IWORK(*),LIW
C ------------
C The arrays WORK(*) and IWORK(*) are respectively floating point
C and type INTEGER working arrays for DSPLP( ) and its
C subprograms. The lengths of these arrays are respectively
C LW and LIW. These parameters must satisfy the inequalities
C noted above under the heading "Fortran Declarations Required."
C It is an error if either value is too small.
C
C |----------------------------|
C |Input/Output files required:|
C |----------------------------|
C
C Fortran unit 1 is used by DSPLP( ) to store the sparse matrix A
C out of high-speed memory. This direct access file is opened
C within the package under the following two conditions.
C 1. When the Save/Restore feature is used. 2. When the
C constraint matrix is so large that storage out of high-speed
C memory is required. The user may need to close unit 1
C (with deletion from the job step) in the main program unit
C when several calls are made to DSPLP( ). A crude
C upper bound for the amount of information written on unit 1
C is 6*nz, where nz is the number of nonzero entries in A.
C The unit number may be redefined to any other positive value
C by means of input in the option array PRGOPT(*).
C
C Fortran unit 2 is used by DSPLP( ) only when the Save/Restore
C feature is desired. Normally this feature is not used. It is
C activated by means of input in the option array PRGOPT(*).
C On some computer systems the user may need to open unit
C 2 before executing a call to DSPLP( ). This file is type
C sequential and is unformatted.
C
C Fortran unit=I1MACH(2) (check local setting) is used by DSPLP( )
C when the printed output feature (KEY=51) is used. Normally
C this feature is not used. It is activated by input in the
C options array PRGOPT(*). For many computer systems I1MACH(2)=6.
C
C |-------|
C |Output:|
C |-------|
C
C INFO,PRIMAL(*),DUALS(*)
C -----------------------
C The integer flag INFO indicates why DSPLP( ) has returned to the
C user. If INFO=1 the solution has been computed. In this case
C X(J)=PRIMAL(J) and W(I)=PRIMAL(I+NVARS). The dual variables
C for the equations A*x=w are in the array DUALS(I)=dual for
C equation number I. The dual value for the component X(J) that
C has an upper or lower bound (or both) is returned in
C DUALS(J+MRELAS). The only other values for INFO are .LT. 0.
C The meaning of these values can be found by reading
C the diagnostic message in the output file, or by looking for
C error number = (-INFO) under the heading "List of DSPLP( ) Error
C and Diagnostic Messages."
C The diagnostic messages are printed using the error processing
C subprogram XERMSG( ) with error category LEVEL=1.
C See the document "Brief Instr. for Using the Sandia Math.
C Subroutine Library," SAND79-2382, Nov., 1980, for further inform-
C ation about resetting the usual response to a diagnostic message.
C
C BL(*),BU(*),IND(*)
C ------------------
C These arrays are output parameters only under the (unusual)
C circumstances where the stated problem is infeasible, has an
C unbounded optimum value, or both. These respective conditions
C correspond to INFO=-1,-2 or -3. For INFO=-1 or -3 certain comp-
C onents of the vectors x or w will not satisfy the input bounds.
C If component J of X or component I of W does not satisfy its input
C bound because of infeasibility, then IND(J)=-4 or IND(I+NVARS)=-4,
C respectively. For INFO=-2 or -3 certain
C components of the vector x could not be used as basic variables
C because the objective function would have become unbounded.
C In particular if component J of x corresponds to such a variable,
C then IND(J)=-3. Further, if the input value of IND(J)
C =1, then BU(J)=BL(J);
C =2, then BL(J)=BU(J);
C =4, then BL(J)=0.,BU(J)=0.
C
C (The J-th variable in x has been restricted to an appropriate
C feasible value.)
C The negative output value for IND(*) allows the user to identify
C those constraints that are not satisfied or those variables that
C would cause unbounded values of the objective function. Note
C that the absolute value of IND(*), together with BL(*) and BU(*),
C are valid input to DSPLP( ). In the case of infeasibility the
C sum of magnitudes of the infeasible values is minimized. Thus
C one could reenter DSPLP( ) with these components of x or w now
C fixed at their present values. This involves setting
C the appropriate components of IND(*) = 3, and BL(*) = BU(*).
C
C IBASIS(I),I=1,...,MRELAS
C ------------------------
C This array contains the indices of the variables that are
C in the active basis set at the solution (INFO=1). A value
C of IBASIS(I) between 1 and NVARS corresponds to the variable
C X(IBASIS(I)). A value of IBASIS(I) between NVARS+1 and NVARS+
C MRELAS corresponds to the variable W(IBASIS(I)-NVARS).
C
C Computing with the Matrix A after Calling DSPLP( )
C --------------------------------------------------
C Following the return from DSPLP( ), nonzero entries of the MRELAS
C by NVARS matrix A are available for usage by the user. The method
C for obtaining the next nonzero in column J with a row index
C strictly greater than I in value, is completed by executing
C
C CALL DPNNZR(I,AIJ,IPLACE,WORK,IWORK,J)
C
C The value of I is also an output parameter. If I.LE.0 on output,
C then there are no more nonzeroes in column J. If I.GT.0, the
C output value for component number I of column J is in AIJ. The
C parameters WORK(*) and IWORK(*) are the same arguments as in the
C call to DSPLP( ). The parameter IPLACE is a single INTEGER
C working variable.
C
C The data structure used for storage of the matrix A within DSPLP()
C corresponds to sequential storage by columns as defined in
C SAND78-0785. Note that the names of the subprograms LNNZRS(),
C LCHNGS(),LINITM(),LLOC(),LRWPGE(), and LRWVIR() have been
C changed to DPNNZR(),DPCHNG(),PINITM(),IPLOC(),DPRWPG(), and
C DPRWVR() respectively. The error processing subprogram LERROR()
C is no longer used; XERMSG() is used instead.
C
C |--------------------------------|
C |Subprograms Required by DSPLP( )|
C |--------------------------------|
C Called by DSPLP() are DPLPMN(),DPLPUP(),DPINIT(),DPOPT(),
C DPLPDM(),DPLPCE(),DPINCW(),DPLPFL(),
C DPLPFE(),DPLPMU().
C
C Error Processing Subprograms XERMSG(),I1MACH(),D1MACH()
C
C Sparse Matrix Subprograms DPNNZR(),DPCHNG(),DPRWPG(),DPRWVR(),
C PINITM(),IPLOC()
C
C Mass Storage File Subprograms SOPENM(),SCLOSM(),DREADP(),DWRITP()
C
C Basic Linear Algebra Subprograms DCOPY(),DASUM(),DDOT()
C
C Sparse Matrix Basis Handling Subprograms LA05AD(),LA05BD(),
C LA05CD(),LA05ED(),MC20AD()
C
C Vector Output Subprograms DVOUT(),IVOUT()
C
C Machine-sensitive Subprograms I1MACH( ),D1MACH( ),
C SOPENM(),SCLOSM(),DREADP(),DWRITP().
C COMMON Block Used
C -----------------
C /LA05DD/ SMALL,LP,LENL,LENU,NCP,LROW,LCOL
C See the document AERE-R8269 for further details.
C |-------------------------|
C |Example of DSPLP( ) Usage|
C |-------------------------|
C PROGRAM LPEX
C THE OPTIMIZATION PROBLEM IS TO FIND X1, X2, X3 THAT
C MINIMIZE X1 + X2 + X3, X1.GE.0, X2.GE.0, X3 UNCONSTRAINED.
C
C THE UNKNOWNS X1,X2,X3 ARE TO SATISFY CONSTRAINTS
C
C X1 -3*X2 +4*X3 = 5
C X1 -2*X2 .LE.3
C 2*X2 - X3.GE.4
C
C WE FIRST DEFINE THE DEPENDENT VARIABLES
C W1=X1 -3*X2 +4*X3
C W2=X1- 2*X2
C W3= 2*X2 -X3
C
C WE NOW SHOW HOW TO USE DSPLP( ) TO SOLVE THIS LINEAR OPTIMIZATION
C PROBLEM. EACH REQUIRED STEP WILL BE SHOWN IN THIS EXAMPLE.
C DIMENSION COSTS(03),PRGOPT(01),DATTRV(18),BL(06),BU(06),IND(06),
C *PRIMAL(06),DUALS(06),IBASIS(06),WORK(079),IWORK(103)
C
C EXTERNAL DUSRMT
C MRELAS=3
C NVARS=3
C
C DEFINE THE ARRAY COSTS(*) FOR THE OBJECTIVE FUNCTION.
C COSTS(01)=1.
C COSTS(02)=1.
C COSTS(03)=1.
C
C PLACE THE NONZERO INFORMATION ABOUT THE MATRIX IN DATTRV(*).
C DEFINE COL. 1:
C DATTRV(01)=-1
C DATTRV(02)=1
C DATTRV(03)=1.
C DATTRV(04)=2
C DATTRV(05)=1.
C
C DEFINE COL. 2:
C DATTRV(06)=-2
C DATTRV(07)=1
C DATTRV(08)=-3.
C DATTRV(09)=2
C DATTRV(10)=-2.
C DATTRV(11)=3
C DATTRV(12)=2.
C
C DEFINE COL. 3:
C DATTRV(13)=-3
C DATTRV(14)=1
C DATTRV(15)=4.
C DATTRV(16)=3
C DATTRV(17)=-1.
C
C DATTRV(18)=0
C
C CONSTRAIN X1,X2 TO BE NONNEGATIVE. LET X3 HAVE NO BOUNDS.
C BL(1)=0.
C IND(1)=1
C BL(2)=0.
C IND(2)=1
C IND(3)=4
C
C CONSTRAIN W1=5,W2.LE.3, AND W3.GE.4.
C BL(4)=5.
C BU(4)=5.
C IND(4)=3
C BU(5)=3.
C IND(5)=2
C BL(6)=4.
C IND(6)=1
C
C INDICATE THAT NO MODIFICATIONS TO OPTIONS ARE IN USE.
C PRGOPT(01)=1
C
C DEFINE THE WORKING ARRAY LENGTHS.
C LW=079
C LIW=103
C CALL DSPLP(DUSRMT,MRELAS,NVARS,COSTS,PRGOPT,DATTRV,
C *BL,BU,IND,INFO,PRIMAL,DUALS,IBASIS,WORK,LW,IWORK,LIW)
C
C CALCULATE VAL, THE MINIMAL VALUE OF THE OBJECTIVE FUNCTION.
C VAL=DDOT(NVARS,COSTS,1,PRIMAL,1)
C
C STOP
C END
C |------------------------|
C |End of Example of Usage |
C |------------------------|
C
C |-------------------------------------|
C |Usage of DSPLP( ) Subprogram Options.|
C |-------------------------------------|
C
C Users frequently have a large variety of requirements for linear
C optimization software. Allowing for these varied requirements
C is at cross purposes with the desire to keep the usage of DSPLP( )
C as simple as possible. One solution to this dilemma is as follows.
C (1) Provide a version of DSPLP( ) that solves a wide class of
C problems and is easy to use. (2) Identify parameters within
C DSPLP() that certain users may want to change. (3) Provide a
C means of changing any selected number of these parameters that
C does not require changing all of them.
C
C Changing selected parameters is done by requiring
C that the user provide an option array, PRGOPT(*), to DSPLP( ).
C The contents of PRGOPT(*) inform DSPLP( ) of just those options
C that are going to be modified within the total set of possible
C parameters that can be modified. The array PRGOPT(*) is a linked
C list consisting of groups of data of the following form
C
C LINK
C KEY
C SWITCH
C data set
C
C that describe the desired options. The parameters LINK, KEY and
C switch are each one word and are always required. The data set
C can be comprised of several words or can be empty. The number of
C words in the data set for each option depends on the value of
C the parameter KEY.
C
C The value of LINK points to the first entry of the next group
C of data within PRGOPT(*). The exception is when there are no more
C options to change. In that case, LINK=1 and the values for KEY,
C SWITCH and data set are not referenced. The general layout of
C PRGOPT(*) is as follows:
C ...PRGOPT(1)=LINK1 (link to first entry of next group)
C . PRGOPT(2)=KEY1 (KEY to the option change)
C . PRGOPT(3)=SWITCH1 (on/off switch for the option)
C . PRGOPT(4)=data value
C . .
C . .
C . .
C ...PRGOPT(LINK1)=LINK2 (link to first entry of next group)
C . PRGOPT(LINK1+1)=KEY2 (KEY to option change)
C . PRGOPT(LINK1+2)=SWITCH2 (on/off switch for the option)
C . PRGOPT(LINK1+3)=data value
C ... .
C . .
C . .
C ...PRGOPT(LINK)=1 (no more options to change)
C
C A value of LINK that is .LE.0 or .GT. 10000 is an error.
C In this case DSPLP( ) returns with an error message, INFO=-14.
C This helps prevent using invalid but positive values of LINK that
C will probably extend beyond the program limits of PRGOPT(*).
C Unrecognized values of KEY are ignored. If the value of SWITCH is
C zero then the option is turned off. For any other value of SWITCH
C the option is turned on. This is used to allow easy changing of
C options without rewriting PRGOPT(*). The order of the options is
C arbitrary and any number of options can be changed with the
C following restriction. To prevent cycling in processing of the
C option array PRGOPT(*), a count of the number of options changed
C is maintained. Whenever this count exceeds 1000 an error message
C (INFO=-15) is printed and the subprogram returns.
C
C In the following description of the options, the value of
C LATP indicates the amount of additional storage that a particular
C option requires. The sum of all of these values (plus one) is
C the minimum dimension for the array PRGOPT(*).
C
C If a user is satisfied with the nominal form of DSPLP( ),
C set PRGOPT(1)=1 (or PRGOPT(1)=1.D0).
C
C Options:
C
C -----KEY = 50. Change from a minimization problem to a maximization
C problem.
C If SWITCH=0 option is off; solve minimization problem.
C =1 option is on; solve maximization problem.
C data set =empty
C LATP=3
C
C -----KEY = 51. Change the amount of printed output. The nominal form
C of DSPLP( ) has no printed output.
C The first level of output (SWITCH=1) includes
C
C (1) Minimum dimensions for the arrays COSTS(*),BL(*),BU(*),IND(*),
C PRIMAL(*),DUALS(*),IBASIS(*), and PRGOPT(*).
C (2) Problem dimensions MRELAS,NVARS.
C (3) The types of and values for the bounds on x and w,
C and the values of the components of the vector costs.
C (4) Whether optimization problem is minimization or
C maximization.
C (5) Whether steepest edge or smallest reduced cost criteria used
C for exchanging variables in the revised simplex method.
C
C Whenever a solution has been found, (INFO=1),
C
C (6) the value of the objective function,
C (7) the values of the vectors x and w,
C (8) the dual variables for the constraints A*x=w and the
C bounded components of x,
C (9) the indices of the basic variables,
C (10) the number of revised simplex method iterations,
C (11) the number of full decompositions of the basis matrix.
C
C The second level of output (SWITCH=2) includes all for SWITCH=1
C plus
C
C (12) the iteration number,
C (13) the column number to enter the basis,
C (14) the column number to leave the basis,
C (15) the length of the step taken.
C
C The third level of output (SWITCH=3) includes all for SWITCH=2
C plus
C (16) critical quantities required in the revised simplex method.
C This output is rather voluminous. It is intended to be used
C as a diagnostic tool in case of a failure in DSPLP( ).
C
C If SWITCH=0 option is off; no printed output.
C =1 summary output.
C =2 lots of output.
C =3 even more output.
C data set =empty
C LATP=3
C
C -----KEY = 52. Redefine the parameter, IDIGIT, which determines the
C format and precision used for the printed output. In the printed
C output, at least ABS(IDIGIT) decimal digits per number is
C printed. If IDIGIT.LT.0, 72 printing columns are used. If
C IDIGIT.GT.0, 133 printing columns are used.
C If SWITCH=0 option is off; IDIGIT=-4.
C =1 option is on.
C data set =IDIGIT
C LATP=4
C
C -----KEY = 53. Redefine LAMAT and LBM, the lengths of the portions of
C WORK(*) and IWORK(*) that are allocated to the sparse matrix
C storage and the sparse linear equation solver, respectively.
C LAMAT must be .GE. NVARS+7 and LBM must be positive.
C If SWITCH=0 option is off; LAMAT=4*NVARS+7
C LBM =8*MRELAS.
C =1 option is on.
C data set =LAMAT
C LBM
C LATP=5
C
C -----KEY = 54. Redefine IPAGEF, the file number where the pages of the
C sparse data matrix are stored. IPAGEF must be positive and
C different from ISAVE (see option 56).
C If SWITCH=0 option is off; IPAGEF=1.
C =1 option is on.
C data set =IPAGEF
C LATP=4
C
C -----KEY = 55. Partial results have been computed and stored on unit
C number ISAVE (see option 56), during a previous run of
C DSPLP( ). This is a continuation from these partial results.
C The arrays COSTS(*),BL(*),BU(*),IND(*) do not have to have
C the same values as they did when the checkpointing occurred.
C This feature makes it possible for the user to do certain
C types of parameter studies such as changing costs and varying
C the constraints of the problem. This file is rewound both be-
C fore and after reading the partial results.
C If SWITCH=0 option is off; start a new problem.
C =1 option is on; continue from partial results
C that are stored in file ISAVE.
C data set = empty
C LATP=3
C
C -----KEY = 56. Redefine ISAVE, the file number where the partial
C results are stored (see option 57). ISAVE must be positive and
C different from IPAGEF (see option 54).
C If SWITCH=0 option is off; ISAVE=2.
C =1 option is on.
C data set =ISAVE
C LATP=4
C
C -----KEY = 57. Save the partial results after maximum number of
C iterations, MAXITR, or at the optimum. When this option is on,
C data essential to continuing the calculation is saved on a file
C using a Fortran binary write operation. The data saved includes
C all the information about the sparse data matrix A. Also saved
C is information about the current basis. Nominally the partial
C results are saved on Fortran unit 2. This unit number can be
C redefined (see option 56). If the save option is on,
C this file must be opened (or declared) by the user prior to the
C call to DSPLP( ). A crude upper bound for the number of words
C written to this file is 6*nz. Here nz= number of nonzeros in A.
C If SWITCH=0 option is off; do not save partial results.
C =1 option is on; save partial results.
C data set = empty
C LATP=3
C
C -----KEY = 58. Redefine the maximum number of iterations, MAXITR, to
C be taken before returning to the user.
C If SWITCH=0 option is off; MAXITR=3*(NVARS+MRELAS).
C =1 option is on.
C data set =MAXITR
C LATP=4
C
C -----KEY = 59. Provide DSPLP( ) with exactly MRELAS indices which
C comprise a feasible, nonsingular basis. The basis must define a
C feasible point: values for x and w such that A*x=w and all the
C stated bounds on x and w are satisfied. The basis must also be
C nonsingular. The failure of either condition will cause an error
C message (INFO=-23 or =-24, respectively). Normally, DSPLP( ) uses
C identity matrix columns which correspond to the components of w.
C This option would normally not be used when restarting from
C a previously saved run (KEY=57).
C In numbering the unknowns,
C the components of x are numbered (1-NVARS) and the components
C of w are numbered (NVARS+1)-(NVARS+MRELAS). A value for an
C index .LE. 0 or .GT. (NVARS+MRELAS) is an error (INFO=-16).
C If SWITCH=0 option is off; DSPLP( ) chooses the initial basis.
C =1 option is on; user provides the initial basis.
C data set =MRELAS indices of basis; order is arbitrary.
C LATP=MRELAS+3
C
C -----KEY = 60. Provide the scale factors for the columns of the data
C matrix A. Normally, DSPLP( ) computes the scale factors as the
C reciprocals of the max. norm of each column.
C If SWITCH=0 option is off; DSPLP( ) computes the scale factors.
C =1 option is on; user provides the scale factors.
C data set =scaling for column J, J=1,NVARS; order is sequential.
C LATP=NVARS+3
C
C -----KEY = 61. Provide a scale factor, COSTSC, for the vector of
C costs. Normally, DSPLP( ) computes this scale factor to be the
C reciprocal of the max. norm of the vector costs after the column
C scaling has been applied.
C If SWITCH=0 option is off; DSPLP( ) computes COSTSC.
C =1 option is on; user provides COSTSC.
C data set =COSTSC
C LATP=4
C
C -----KEY = 62. Provide size parameters, ASMALL and ABIG, the smallest
C and largest magnitudes of nonzero entries in the data matrix A,
C respectively. When this option is on, DSPLP( ) will check the
C nonzero entries of A to see if they are in the range of ASMALL and
C ABIG. If an entry of A is not within this range, DSPLP( ) returns
C an error message, INFO=-22. Both ASMALL and ABIG must be positive
C with ASMALL .LE. ABIG. Otherwise, an error message is returned,
C INFO=-17.
C If SWITCH=0 option is off; no checking of the data matrix is done
C =1 option is on; checking is done.
C data set =ASMALL
C ABIG
C LATP=5
C
C -----KEY = 63. Redefine the relative tolerance, TOLLS, used in
C checking if the residuals are feasible. Normally,
C TOLLS=RELPR, where RELPR is the machine precision.
C If SWITCH=0 option is off; TOLLS=RELPR.
C =1 option is on.
C data set =TOLLS
C LATP=4
C
C -----KEY = 64. Use the minimum reduced cost pricing strategy to choose
C columns to enter the basis. Normally, DSPLP( ) uses the steepest
C edge pricing strategy which is the best local move. The steepest
C edge pricing strategy generally uses fewer iterations than the
C minimum reduced cost pricing, but each iteration costs more in the
C number of calculations done. The steepest edge pricing is
C considered to be more efficient. However, this is very problem
C dependent. That is why DSPLP( ) provides the option of either
C pricing strategy.
C If SWITCH=0 option is off; steepest option edge pricing is used.
C =1 option is on; minimum reduced cost pricing is used.
C data set =empty
C LATP=3
C
C -----KEY = 65. Redefine MXITBR, the number of iterations between
C recalculating the error in the primal solution. Normally, MXITBR
C is set to 10. The error in the primal solution is used to monitor
C the error in solving the linear system. This is an expensive
C calculation and every tenth iteration is generally often enough.
C If SWITCH=0 option is off; MXITBR=10.
C =1 option is on.
C data set =MXITBR
C LATP=4
C
C -----KEY = 66. Redefine NPP, the number of negative reduced costs
C (at most) to be found at each iteration of choosing
C a variable to enter the basis. Normally NPP is set
C to NVARS which implies that all of the reduced costs
C are computed at each such step. This "partial
C pricing" may very well increase the total number
C of iterations required. However it decreases the
C number of calculations at each iteration.
C therefore the effect on overall efficiency is quite
C problem-dependent.
C
C if SWITCH=0 option is off; NPP=NVARS
C =1 option is on.
C data set =NPP
C LATP=4
C
C -----KEY = 67. Redefine the tuning factor (PHI) used to scale the
C error estimates for the primal and dual linear algebraic systems
C of equations. Normally, PHI = 1.D0, but in some environments it
C may be necessary to reset PHI to the range 0.001-0.01. This is
C particularly important for machines with short word lengths.
C
C if SWITCH = 0 option is off; PHI=1.D0.
C = 1 option is on.
C Data Set = PHI
C LATP=4
C
C -----KEY = 68. Used together with the subprogram DFULMT(), provided
C with the DSPLP() package, for passing a standard Fortran two-
C dimensional array containing the constraint matrix. Thus the sub-
C program DFULMT must be declared in a Fortran EXTERNAL statement.
C The two-dimensional array is passed as the argument DATTRV.
C The information about the array and problem dimensions are passed
C in the option array PRGOPT(*). It is an error if DFULMT() is
C used and this information is not passed in PRGOPT(*).
C
C if SWITCH = 0 option is off; this is an error is DFULMT() is
C used.
C = 1 option is on.
C Data Set = IA = row dimension of two-dimensional array.
C MRELAS = number of constraint equations.
C NVARS = number of dependent variables.
C LATP = 6
C -----KEY = 69. Normally a relative tolerance (TOLLS, see option 63)
C is used to decide if the problem is feasible. If this test fails
C an absolute test will be applied using the value TOLABS.
C Nominally TOLABS = zero.
C If SWITCH = 0 option is off; TOLABS = zero.
C = 1 option is on.
C Data set = TOLABS
C LATP = 4
C
C |-----------------------------|
C |Example of Option array Usage|
C |-----------------------------|
C To illustrate the usage of the option array, let us suppose that
C the user has the following nonstandard requirements:
C
C a) Wants to change from minimization to maximization problem.
C b) Wants to limit the number of simplex steps to 100.
C c) Wants to save the partial results after 100 steps on
C Fortran unit 2.
C
C After these 100 steps are completed the user wants to continue the
C problem (until completed) using the partial results saved on
C Fortran unit 2. Here are the entries of the array PRGOPT(*)
C that accomplish these tasks. (The definitions of the other
C required input parameters are not shown.)
C
C CHANGE TO A MAXIMIZATION PROBLEM; KEY=50.
C PRGOPT(01)=4
C PRGOPT(02)=50
C PRGOPT(03)=1
C
C LIMIT THE NUMBER OF SIMPLEX STEPS TO 100; KEY=58.
C PRGOPT(04)=8
C PRGOPT(05)=58
C PRGOPT(06)=1
C PRGOPT(07)=100
C
C SAVE THE PARTIAL RESULTS, AFTER 100 STEPS, ON FORTRAN
C UNIT 2; KEY=57.
C PRGOPT(08)=11
C PRGOPT(09)=57
C PRGOPT(10)=1
C
C NO MORE OPTIONS TO CHANGE.
C PRGOPT(11)=1
C The user makes the CALL statement for DSPLP( ) at this point.
C Now to restart, using the partial results after 100 steps, define
C new values for the array PRGOPT(*):
C
C AGAIN INFORM DSPLP( ) THAT THIS IS A MAXIMIZATION PROBLEM.
C PRGOPT(01)=4
C PRGOPT(02)=50
C PRGOPT(03)=1
C
C RESTART, USING SAVED PARTIAL RESULTS; KEY=55.
C PRGOPT(04)=7
C PRGOPT(05)=55
C PRGOPT(06)=1
C
C NO MORE OPTIONS TO CHANGE. THE SUBPROGRAM DSPLP( ) IS NO LONGER
C LIMITED TO 100 SIMPLEX STEPS BUT WILL RUN UNTIL COMPLETION OR
C MAX.=3*(MRELAS+NVARS) ITERATIONS.
C PRGOPT(07)=1
C The user now makes a CALL to subprogram DSPLP( ) to compute the
C solution.
C |--------------------------------------------|
C |End of Usage of DSPLP( ) Subprogram Options.|
C |--------------------------------------------|
C
C |-----------------------------------------------|
C |List of DSPLP( ) Error and Diagnostic Messages.|
C |-----------------------------------------------|
C This section may be required to understand the meanings of the
C error flag =-INFO that may be returned from DSPLP( ).
C
C -----1. There is no set of values for x and w that satisfy A*x=w and
C the stated bounds. The problem can be made feasible by ident-
C ifying components of w that are now infeasible and then rede-
C signating them as free variables. Subprogram DSPLP( ) only
C identifies an infeasible problem; it takes no other action to
C change this condition. Message:
C DSPLP( ). THE PROBLEM APPEARS TO BE INFEASIBLE.
C ERROR NUMBER = 1
C
C 2. One of the variables in either the vector x or w was con-
C strained at a bound. Otherwise the objective function value,
C (transpose of costs)*x, would not have a finite optimum.
C Message:
C DSPLP( ). THE PROBLEM APPEARS TO HAVE NO FINITE SOLN.
C ERROR NUMBER = 2
C
C 3. Both of the conditions of 1. and 2. above have occurred.
C Message:
C DSPLP( ). THE PROBLEM APPEARS TO BE INFEASIBLE AND TO
C HAVE NO FINITE SOLN.
C ERROR NUMBER = 3
C
C -----4. The REAL and INTEGER working arrays, WORK(*) and IWORK(*),
C are not long enough. The values (I1) and (I2) in the message
C below will give you the minimum length required. Also redefine
C LW and LIW, the lengths of these arrays. Message:
C DSPLP( ). WORK OR IWORK IS NOT LONG ENOUGH. LW MUST BE (I1)
C AND LIW MUST BE (I2).
C IN ABOVE MESSAGE, I1= 0
C IN ABOVE MESSAGE, I2= 0
C ERROR NUMBER = 4
C
C -----5. and 6. These error messages often mean that one or more
C arguments were left out of the call statement to DSPLP( ) or
C that the values of MRELAS and NVARS have been over-written
C by garbage. Messages:
C DSPLP( ). VALUE OF MRELAS MUST BE .GT.0. NOW=(I1).
C IN ABOVE MESSAGE, I1= 0
C ERROR NUMBER = 5
C
C DSPLP( ). VALUE OF NVARS MUST BE .GT.0. NOW=(I1).
C IN ABOVE MESSAGE, I1= 0
C ERROR NUMBER = 6
C
C -----7.,8., and 9. These error messages can occur as the data matrix
C is being defined by either DUSRMT( ) or the user-supplied sub-
C program, 'NAME'( ). They would indicate a mistake in the contents
C of DATTRV(*), the user-written subprogram or that data has been
C over-written.
C Messages:
C DSPLP( ). MORE THAN 2*NVARS*MRELAS ITERS. DEFINING OR UPDATING
C MATRIX DATA.
C ERROR NUMBER = 7
C
C DSPLP( ). ROW INDEX (I1) OR COLUMN INDEX (I2) IS OUT OF RANGE.
C IN ABOVE MESSAGE, I1= 1
C IN ABOVE MESSAGE, I2= 12
C ERROR NUMBER = 8
C
C DSPLP( ). INDICATION FLAG (I1) FOR MATRIX DATA MUST BE
C EITHER 0 OR 1.
C IN ABOVE MESSAGE, I1= 12
C ERROR NUMBER = 9
C
C -----10. and 11. The type of bound (even no bound) and the bounds
C must be specified for each independent variable. If an independent
C variable has both an upper and lower bound, the bounds must be
C consistent. The lower bound must be .LE. the upper bound.
C Messages:
C DSPLP( ). INDEPENDENT VARIABLE (I1) IS NOT DEFINED.
C IN ABOVE MESSAGE, I1= 1
C ERROR NUMBER = 10
C
C DSPLP( ). LOWER BOUND (R1) AND UPPER BOUND (R2) FOR INDEP.
C VARIABLE (I1) ARE NOT CONSISTENT.
C IN ABOVE MESSAGE, I1= 1
C IN ABOVE MESSAGE, R1= 0.
C IN ABOVE MESSAGE, R2= -.1000000000E+01
C ERROR NUMBER = 11
C
C -----12. and 13. The type of bound (even no bound) and the bounds
C must be specified for each dependent variable. If a dependent
C variable has both an upper and lower bound, the bounds must be
C consistent. The lower bound must be .LE. the upper bound.
C Messages:
C DSPLP( ). DEPENDENT VARIABLE (I1) IS NOT DEFINED.
C IN ABOVE MESSAGE, I1= 1
C ERROR NUMBER = 12
C
C DSPLP( ). LOWER BOUND (R1) AND UPPER BOUND (R2) FOR DEP.
C VARIABLE (I1) ARE NOT CONSISTENT.
C IN ABOVE MESSAGE, I1= 1
C IN ABOVE MESSAGE, R1= 0.
C IN ABOVE MESSAGE, R2= -.1000000000E+01
C ERROR NUMBER = 13
C
C -----14. - 21. These error messages can occur when processing the
C option array, PRGOPT(*), supplied by the user. They would
C indicate a mistake in defining PRGOPT(*) or that data has been
C over-written. See heading Usage of DSPLP( )
C Subprogram Options, for details on how to define PRGOPT(*).
C Messages:
C DSPLP( ). THE USER OPTION ARRAY HAS UNDEFINED DATA.
C ERROR NUMBER = 14
C
C DSPLP( ). OPTION ARRAY PROCESSING IS CYCLING.
C ERROR NUMBER = 15
C
C DSPLP( ). AN INDEX OF USER-SUPPLIED BASIS IS OUT OF RANGE.
C ERROR NUMBER = 16
C
C DSPLP( ). SIZE PARAMETERS FOR MATRIX MUST BE SMALLEST AND LARGEST
C MAGNITUDES OF NONZERO ENTRIES.
C ERROR NUMBER = 17
C
C DSPLP( ). THE NUMBER OF REVISED SIMPLEX STEPS BETWEEN CHECK-POINTS
C MUST BE POSITIVE.
C ERROR NUMBER = 18
C
C DSPLP( ). FILE NUMBERS FOR SAVED DATA AND MATRIX PAGES MUST BE
C POSITIVE AND NOT EQUAL.
C ERROR NUMBER = 19
C
C DSPLP( ). USER-DEFINED VALUE OF LAMAT (I1)
C MUST BE .GE. NVARS+7.
C IN ABOVE MESSAGE, I1= 1
C ERROR NUMBER = 20
C
C DSPLP( ). USER-DEFINED VALUE OF LBM MUST BE .GE. 0.
C ERROR NUMBER = 21
C
C -----22. The user-option, number 62, to check the size of the matrix
C data has been used. An element of the matrix does not lie within
C the range of ASMALL and ABIG, parameters provided by the user.
C (See the heading: Usage of DSPLP( ) Subprogram Options,
C for details about this feature.) Message:
C DSPLP( ). A MATRIX ELEMENT'S SIZE IS OUT OF THE SPECIFIED RANGE.
C ERROR NUMBER = 22
C
C -----23. The user has provided an initial basis that is singular.
C In this case, the user can remedy this problem by letting
C subprogram DSPLP( ) choose its own initial basis. Message:
C DSPLP( ). A SINGULAR INITIAL BASIS WAS ENCOUNTERED.
C ERROR NUMBER = 23
C
C -----24. The user has provided an initial basis which is infeasible.
C The x and w values it defines do not satisfy A*x=w and the stated
C bounds. In this case, the user can let subprogram DSPLP( )
C choose its own initial basis. Message:
C DSPLP( ). AN INFEASIBLE INITIAL BASIS WAS ENCOUNTERED.
C ERROR NUMBER = 24
C
C -----25.Subprogram DSPLP( ) has completed the maximum specified number
C of iterations. (The nominal maximum number is 3*(MRELAS+NVARS).)
C The results, necessary to continue on from
C this point, can be saved on Fortran unit 2 by activating option
C KEY=57. If the user anticipates continuing the calculation, then
C the contents of Fortran unit 2 must be retained intact. This
C is not done by subprogram DSPLP( ), so the user needs to save unit
C 2 by using the appropriate system commands. Message:
C DSPLP( ). MAX. ITERS. (I1) TAKEN. UP-TO-DATE RESULTS
C SAVED ON FILE (I2). IF(I2)=0, NO SAVE.
C IN ABOVE MESSAGE, I1= 500
C IN ABOVE MESSAGE, I2= 2
C ERROR NUMBER = 25
C
C -----26. This error should never happen. Message:
C DSPLP( ). MOVED TO A SINGULAR POINT. THIS SHOULD NOT HAPPEN.
C ERROR NUMBER = 26
C
C -----27. The subprogram LA05A( ), which decomposes the basis matrix,
C has returned with an error flag (R1). (See the document,
C "Fortran subprograms for handling sparse linear programming
C bases", AERE-R8269, J.K. Reid, Jan., 1976, H.M. Stationery Office,
C for an explanation of this error.) Message:
C DSPLP( ). LA05A( ) RETURNED ERROR FLAG (R1) BELOW.
C IN ABOVE MESSAGE, R1= -.5000000000E+01
C ERROR NUMBER = 27
C
C -----28. The sparse linear solver package, LA05*( ), requires more
C space. The value of LBM must be increased. See the companion
C document, Usage of DSPLP( ) Subprogram Options, for details on how
C to increase the value of LBM. Message:
C DSPLP( ). SHORT ON STORAGE FOR LA05*( ) PACKAGE. USE PRGOPT(*)
C TO GIVE MORE.
C ERROR NUMBER = 28
C
C -----29. The row dimension of the two-dimensional Fortran array,
C the number of constraint equations (MRELAS), and the number
C of variables (NVARS), were not passed to the subprogram
C DFULMT(). See KEY = 68 for details. Message:
C DFULMT() OF DSPLP() PACKAGE. ROW DIM., MRELAS, NVARS ARE
C MISSING FROM PRGOPT(*).
C ERROR NUMBER = 29
C
C |-------------------------------------------------------|
C |End of List of DSPLP( ) Error and Diagnostic Messages. |
C |-------------------------------------------------------|
C***REFERENCES R. J. Hanson and K. L. Hiebert, A sparse linear
C programming subprogram, Report SAND81-0297, Sandia
C National Laboratories, 1981.
C***ROUTINES CALLED DPLPMN, XERMSG
C***REVISION HISTORY (YYMMDD)
C 811215 DATE WRITTEN
C 890531 Changed all specific intrinsics to generic. (WRB)
C 890605 Corrected references to XERRWV. (WRB)
C 890605 Removed unreferenced labels. (WRB)
C 891006 Cosmetic changes to prologue. (WRB)
C 891006 REVISION DATE from Version 3.2
C 891214 Prologue converted to Version 4.0 format. (BAB)
C 900315 CALLs to XERROR changed to CALLs to XERMSG. (THJ)
C 900510 Convert XERRWV calls to XERMSG calls. (RWC)
C 920501 Reformatted the REFERENCES section. (WRB)
C***END PROLOGUE DSPLP
DOUBLE PRECISION BL(*),BU(*),COSTS(*),DATTRV(*),DUALS(*),
* PRGOPT(*),PRIMAL(*),WORK(*),ZERO
C
INTEGER IBASIS(*),IND(*),IWORK(*)
CHARACTER*8 XERN1, XERN2
C
EXTERNAL DUSRMT
C
C***FIRST EXECUTABLE STATEMENT DSPLP
ZERO=0.D0
IOPT=1
C
C VERIFY THAT MRELAS, NVARS .GT. 0.
C
IF (MRELAS.LE.0) THEN
WRITE (XERN1, '(I8)') MRELAS
CALL XERMSG ('SLATEC', 'DSPLP', 'VALUE OF MRELAS MUST BE ' //
* '.GT. 0. NOW = ' // XERN1, 5, 1)
INFO = -5
RETURN
ENDIF
C
IF (NVARS.LE.0) THEN
WRITE (XERN1, '(I8)') NVARS
CALL XERMSG ('SLATEC', 'DSPLP', 'VALUE OF NVARS MUST BE ' //
* '.GT. 0. NOW = ' // XERN1, 6, 1)
INFO = -6
RETURN
ENDIF
C
LMX=4*NVARS+7
LBM=8*MRELAS
LAST = 1
IADBIG=10000
ICTMAX=1000
ICTOPT= 0
C
C LOOK IN OPTION ARRAY FOR CHANGES TO WORK ARRAY LENGTHS.
20008 NEXT=PRGOPT(LAST)
IF (.NOT.(NEXT.LE.0 .OR. NEXT.GT.IADBIG)) GO TO 20010
C
C THE CHECKS FOR SMALL OR LARGE VALUES OF NEXT ARE TO PREVENT
C WORKING WITH UNDEFINED DATA.
NERR=14
CALL XERMSG ('SLATEC', 'DSPLP',
+ 'THE USER OPTION ARRAY HAS UNDEFINED DATA.', NERR, IOPT)
INFO=-NERR
RETURN
20010 IF (.NOT.(NEXT.EQ.1)) GO TO 10001
GO TO 20009
10001 IF (.NOT.(ICTOPT.GT.ICTMAX)) GO TO 10002
NERR=15
CALL XERMSG ('SLATEC', 'DSPLP',
+ 'OPTION ARRAY PROCESSING IS CYCLING.', NERR, IOPT)
INFO=-NERR
RETURN
10002 CONTINUE
KEY = PRGOPT(LAST+1)
C
C IF KEY = 53, USER MAY SPECIFY LENGTHS OF PORTIONS
C OF WORK(*) AND IWORK(*) THAT ARE ALLOCATED TO THE
C SPARSE MATRIX STORAGE AND SPARSE LINEAR EQUATION
C SOLVING.
IF (.NOT.(KEY.EQ.53)) GO TO 20013
IF (.NOT.(PRGOPT(LAST+2).NE.ZERO)) GO TO 20016
LMX=PRGOPT(LAST+3)
LBM=PRGOPT(LAST+4)
20016 CONTINUE
20013 ICTOPT = ICTOPT+1
LAST = NEXT
GO TO 20008
C
C CHECK LENGTH VALIDITY OF SPARSE MATRIX STAGING AREA.
C
20009 IF (LMX.LT.NVARS+7) THEN
WRITE (XERN1, '(I8)') LMX
CALL XERMSG ('SLATEC', 'DSPLP', 'USER-DEFINED VALUE OF ' //
* 'LAMAT = ' // XERN1 // ' MUST BE .GE. NVARS+7.', 20, 1)
INFO = -20
RETURN
ENDIF
C
C TRIVIAL CHECK ON LENGTH OF LA05*() MATRIX AREA.
C
IF (.NOT.(LBM.LT.0)) GO TO 20022
NERR=21
CALL XERMSG ('SLATEC', 'DSPLP',
+ 'USER-DEFINED VALUE OF LBM MUST BE .GE. 0.', NERR, IOPT)
INFO=-NERR
RETURN
20022 CONTINUE
C
C DEFINE POINTERS FOR STARTS OF SUBARRAYS USED IN WORK(*)
C AND IWORK(*) IN OTHER SUBPROGRAMS OF THE PACKAGE.
LAMAT=1
LCSC=LAMAT+LMX
LCOLNR=LCSC+NVARS
LERD=LCOLNR+NVARS
LERP=LERD+MRELAS
LBASMA=LERP+MRELAS
LWR=LBASMA+LBM
LRZ=LWR+MRELAS
LRG=LRZ+NVARS+MRELAS
LRPRIM=LRG+NVARS+MRELAS
LRHS=LRPRIM+MRELAS
LWW=LRHS+MRELAS
LWORK=LWW+MRELAS-1
LIMAT=1
LIBB=LIMAT+LMX
LIBRC=LIBB+NVARS+MRELAS
LIPR=LIBRC+2*LBM
LIWR=LIPR+2*MRELAS
LIWORK=LIWR+8*MRELAS-1
C
C CHECK ARRAY LENGTH VALIDITY OF WORK(*), IWORK(*).
C
IF (LW.LT.LWORK .OR. LIW.LT.LIWORK) THEN
WRITE (XERN1, '(I8)') LWORK
WRITE (XERN2, '(I8)') LIWORK
CALL XERMSG ('SLATEC', 'DSPLP', 'WORK OR IWORK IS NOT LONG ' //
* 'ENOUGH. LW MUST BE = ' // XERN1 // ' AND LIW MUST BE = ' //
* XERN2, 4, 1)
INFO = -4
RETURN
ENDIF
C
CALL DPLPMN(DUSRMT,MRELAS,NVARS,COSTS,PRGOPT,DATTRV,
* BL,BU,IND,INFO,PRIMAL,DUALS,WORK(LAMAT),
* WORK(LCSC),WORK(LCOLNR),WORK(LERD),WORK(LERP),WORK(LBASMA),
* WORK(LWR),WORK(LRZ),WORK(LRG),WORK(LRPRIM),WORK(LRHS),
* WORK(LWW),LMX,LBM,IBASIS,IWORK(LIBB),IWORK(LIMAT),
* IWORK(LIBRC),IWORK(LIPR),IWORK(LIWR))
C
C CALL DPLPMN(DUSRMT,MRELAS,NVARS,COSTS,PRGOPT,DATTRV,
C 1 BL,BU,IND,INFO,PRIMAL,DUALS,AMAT,
C 2 CSC,COLNRM,ERD,ERP,BASMAT,
C 3 WR,RZ,RG,RPRIM,RHS,
C 4 WW,LMX,LBM,IBASIS,IBB,IMAT,
C 5 IBRC,IPR,IWR)
C
RETURN
END