!************************************************************************* !* * !* Initialization of parameters for A-PWLSVR * !* (last modified 01.02.2022) * !* * !************************************************************************* !* !* Modules included: !* !* initspr ! Initialization of parameters for A-PWLSVR (main program svr). !* initdbdc ! Initialization of DBDC -solver. !* MODULE initspr ! Initialization of parameters. USE constants, ONLY : dp ! Precision for reals. IMPLICIT NONE ! Names of input and output files: CHARACTER(LEN=80), SAVE :: & infile, & ! Name of the dataset file. Given as commandline argument. outfile1 = 'reg_results.txt', & ! Regression results with hyperplanes. outfile2 = 'predictions.txt', & ! Regression results with predictions. outfile3 = 'pwl_points.txt' ! Points of original data and pwl-model obtained ! for drawing purposis. Not working with nft > 2. ! Input real parameters. REAL(KIND=dp), SAVE :: & minv = -2_dp, & ! Minimum value for pwl_points (need to be given by the user), ! used with outfile3, othewise ignored. maxv = 2_dp, & ! Maximum value for pwl_points (need to be given by the user), ! used with outfile3, othewise ignored. tol = 0.05_dp ! Stopping tolerance. ! Input integer parameters. INTEGER, SAVE :: & nft, & ! Number of features in data, from user. nrecord, & ! Number of records in data, from user. ntrain, & ! Size of the training set, ntrain <= nrecord. noutcom, & ! Output column. maxmin = 2, & ! Maximum number of min functions in the model, from user. maxlin = 3, & ! Number of linear functions in the model, optional, from user. iscale = 1, & ! Scaling is ! 0 - not used, ! 1 - used (recommended). inc = 2, & ! Incremental algorithm is ! 0 - not used (the starting point is obtained by regression), ! 1 - used. ! 2 - used with double starting points. ipredi = 1 ! Prediction: ! 0 - only at the termination with complete model. ! 1 - every step. ! Allocatable real arrays. REAL(KIND=dp), SAVE, DIMENSION(:,:), allocatable :: & a, & ! Data set, from input file. a4 ! Training part of the data (ntrain first objects of a). REAL(KIND=dp), SAVE, DIMENSION(:), allocatable :: & ! x0, & ! the starting point. x0reg, & ! SVR solution. x_solution, & ! the solution. x_solution2, & ! the solution. x_unscaled, & ! the unscaled solution. pred, & ! the values of the predicted outputs clin, & dlin ! Other real parameters. REAL(KIND=dp), SAVE :: & f_solution, & ! the objective function value at the solution 'x_solution' f_solution2, & ! the objective function value at the solution 'x_solution2' eps = 0.050_dp, & ! the epsilon used in SVM, change in init_par. CC = 0.50_dp, & ! the coefficient in front of maximum terms (i.e. maximum from minimums) Cnull = 0.0_dp ! the coefficient in front of the quadratic terms, Should be zero! ! Other integer parameters. INTEGER, SAVE :: & ! maxmin2, & ! maxmin2 = 1,...,maxmin maxdim, & ! Maximum dimension in optimization, nft*maxmin*maxlin q ! Total number of linear pieces, maxmin*maxlin ! Allocatable integer tables INTEGER, SAVE, DIMENSION(:), allocatable :: & mk ! mk(maxmin) = number of lines in each minimum function CONTAINS SUBROUTINE init_par() ! Initialization of parameters. IMPLICIT NONE q = maxmin*maxlin ! this if maxlin is constant for all k=1,...,maxmin maxdim = nft*q IF (iscale == 1) eps=0.05_dp END SUBROUTINE init_par !=================================================================== ! Scaling !=================================================================== SUBROUTINE scaling IMPLICIT NONE ! Local variables REAL(KIND=dp) :: & dm,var INTEGER :: & i,j ! Intrinsic Functions INTRINSIC ABS,SQRT DO i=1,nft dm=0_dp DO j=1,ntrain dm=dm+a4(j,i) END DO dm=dm/REAL(ntrain,dp) var=0_dp DO j=1,ntrain var=var+(a4(j,i)-dm)**2 END DO var=var/REAL(ntrain-1,dp) var=SQRT(var) IF (var >= 1.0E-05_dp) THEN clin(i)=1_dp/var dlin(i)=-dm/var DO j=1,ntrain a4(j,i)=clin(i)*a4(j,i)+dlin(i) END DO ELSE IF (ABS(dm) <= 1_dp) THEN clin(i)=1_dp dlin(i)=0_dp ELSE clin(i)=1_dp/dm dlin(i)=0_dp END IF DO j=1,ntrain a4(j,i)=clin(i)*a4(j,i)+dlin(i) END DO END IF END DO RETURN END SUBROUTINE scaling !=================================================================== ! Rescaling !=================================================================== SUBROUTINE rescaling IMPLICIT NONE ! Local variables REAL(KIND=dp) :: & d0,rabs INTEGER :: & i,j,j1,k ! Intrinsic Functions INTRINSIC ABS rabs=ABS(clin(nft)) IF (rabs > 1.0E-08_dp) THEN DO k=1,maxmin2 DO j=1,mk(k) ! This is correct as we always do upto maxlin linear functions i=j+(k-1)*mk(k) !DO j=1,maxlin ! This is correct as we always do upto maxlin linear functions ! i=j+(k-1)*maxlin d0=0_dp DO j1=1,nft-1 x_unscaled(j1+(i-1)*nft)=x_solution(j1+(i-1)*nft)*clin(j1)/rabs d0=d0+x_solution(j1+(i-1)*nft)*dlin(j1) END DO x_unscaled(i*nft)=(x_solution(i*nft)+d0-dlin(nft))/rabs END DO END DO END IF RETURN END SUBROUTINE rescaling END MODULE initspr MODULE initdbdc ! Initialization of parameters for DBDC. There is, usually, no need to change these parameters. USE constants, ONLY : dp ! Precision for reals. IMPLICIT NONE ! Parameters REAL(KIND=dp), PARAMETER :: & user_m = 0.01_dp, & ! The descent parameter. If user_m <= 0.0_dp .OR. user_m >= 1.0_dp ! then DEFAULT value 0.2_dp is used user_c = -0.99_dp, & ! The decrease parameter c in DBDC. If user_c <= 0.0_dp or user_c > 1.0_dp ! then DEFAULT value 0.1_dp is used user_r_dec = -0.99_dp, & ! The decrease parameter r in DBDC user_r_inc = (10.0_dp)**(7), & ! The increase parameter R: If user_r_inc <= 1.0_dp ! then DEFAULT value (10.0_dp)**7 is used user_eps_1 = 5*(10.0_dp)**(-5), & ! The enlargement parameter: If user_eps_1 <= 0.0_dp .OR. user_eps_1 > 1.0_dp ! then DEFAULT value 5*(10.0_dp)**(-5) is used user_crit_tol = (10.0_dp)**(-3), & ! If user_crit_tol <= 0.0_dp then DEFAULT value (10.0_dp)**(-5) is used when n <=200 ! (10.0_dp)**(-4) is used when n > 200 user_m_clarke = 0.01_dp, & ! The descent parameter: If user_m_clarke <= 0.0_dp .OR. user_m_clarke >= 1.0_dp ! then DEFAULT value 0.01_dp is used user_eps = 10.0_dp**(-4) ! The proximity measure: If user_eps <= 0.0_dp ! then DEFAULT value (10.0_dp)**(-6) is used when n <= 50 ! (10.0_dp)**(-5) is used when n > 50 INTEGER, PARAMETER :: & problem1 = 1, & ! Objective function selected for the DC component f1 problem2 = problem1, & ! Objective function selected for the DC component f2 user_size_b2 = 3 ! The biggest possible size of the bundle B_2 ! If user_size_b2 <= 0 then DEFAULT value 3 is used INTEGER, PARAMETER :: & user_n = 2, & user_size_b1 = MIN(user_n + 3,50), & ! The biggest possible size of the bundle B_1 ! If user_size_b1 <= 0 then DEFAULT value MIN(user_n+5,1000) is used user_size = user_n * 2 ! The biggest possible bundle size for the bundle used in ! 'Clarke stationary' algorithm INTEGER, SAVE, DIMENSION(8) :: counter ! Contains the values of different counteres: ! counter(1) = iter_counter: the number of 'main iterations' executed ! counter(2) = subprob_counter: the number of subproblems solved in 'main iteration' ! counter(3) = f_counter: the number of function values evaluated for DC component in 'main iteration' ! counter(4) = subgrad1_counter: the number of subgradients calculated for f_1 in 'main iteration' ! counter(5) = subgrad2_counter: the number of subgradients calculated for f_2 in 'main iteration' ! counter(6) = stop_cond_counter: the number of times 'Clarke stationary algorithm' is executed ! counter(7) = clarke_f_counter: the number of function values evaluated for f in 'Clarke stationary algorithms' ! counter(8) = clarke_sub_counter: the number of subgradients caluculated for f in 'Clarke stationary algorithms' INTEGER :: & iprint = 0, & ! Variable that specifies print option (specified by USER): ! iprint = 0 : print is suppressed ! iprint = 1 : basic print of final result ! iprint = -1: basic print of final result (without the solution vector) ! iprint = 2 : extended print of final result ! iprint = -2: extended print of final result (without the solution vector) ! iprint = 3 : basic print of intermediate results and extended print of final results ! iprint = -3: basic print of intermediate results and extended print of final results (without the solution vector) ! iprint = 4 : extended print of intermediate results and extended print of final results ! iprint = -4: extended print of intermediate results and extended print of final results (without the solution vectors) ! ! If 'iprint' <= -5 .OR. 'iprint' >= 5 then DEFAULT value 'iprint'=1 is used mit = 500, & ! The maximum number of 'main iterations' (specified by USER). ! If 'mit' <=0 then DEFAULT value 'mit'=1000 is used mrounds = 500, & ! The maximum number of rounds during one 'main iteration' (specified by USER). ! If 'mrounds' <=0 then DEFAULT value 'mrounds'=5000 is used mrounds_clarke = 500, & ! The maximum number of rounds during one 'Clarke stationarity' algorithm (specified by USER). ! If 'mrounds_clarke' <=0 then DEFAULT value 'mrounds_clarke'=5000 is used termination ! The reason for termination in DBDC: ! 1 - the stopping condition is satisfied (i.e. Clarke stationarity) ! 2 - the approximate stopping condition is satisfied (i.e. the step-length beta* < user_eps) ! 3 - the maximum number 'mrounds' of rounds is executed in one main iteration ! 4 - the maximum number of 'main iterations' is executed ! 5 - the maximum number 'mrounds_clarke' of rounds is executed in one 'Clarke stationary' algorithm LOGICAL :: & !agg_used = .FALSE., & ! If .TRUE. then aggregation is used in DBDC (specified by USER). agg_used = .TRUE., & ! If .TRUE. then aggregation is used in DBDC (specified by USER). stepsize_used = .TRUE. ! If .TRUE. then simple stepsize determination is done in DBDC after each 'main iteration' (specified by USER). END MODULE initdbdc