EMSOL Task Stack

Urgent

• Complete PWL
  Fix bug when LP from SLP has BP variables.
  • ems_reset_1_non_bc_pwl_vr_st needs to be completed
  • ems_g_lp_ph uses its own code to handle BP variables---should call generic code.
• Sort out assembler version
• Add option to move nonbasic variables onto their bounds at optimality
• Prevent reset loops
• Look at Jan's CHUZC bug
• Change tl_pr_ifs rather than use small pivots.

Developments

• Maintain n_pwl_vr_in_bs n_bp_vr_in_bs and use them to determine which CHUZR to do.
• Add dual facility. Remember ems_un_bd_aux_sol. Use n_struc_vr/n_lg_vr to refer to number of columns/rows in rsmi and n_c/n_r to refer to number of columns/rows in model---same in primal but swapped in dual.
• Cheap FTRAN
• See how much faster the code is with shifted/swapped bounds so every non-free variable has a lower bound of zero.
• Dense block in INVERT?
• Check consistency of sparse matrix definition passed to ems_lmdl
• Schur complement updating (and switch from PF when numerical going gets tough).
• Comprehensive test routines (a la HSL).

Not critical

• Use current section numbers 0 and n_sn+1 to indicate PWL variables violating bounds
• Code to test for infinte values in PWL data, non-convexities or non-monotone breakpoints and eliminate duplicate breakpoints/costs is written but not checked.
• L1-CHUZC for DANTZIG
• L1-CHUZC for U_STPST_ED_WT
• Although MPS read is give half of the available space for fixed proportions of rows, cols and elements, it fails if the limit is reached rather than re-allocating space more intelligently: use more or increase proportion for data type whose limit has been reached.
• Crash when problem has BP/PWL variables
• Prevent overflow when MPS file contains discrete variables.
• Is the MPS default for rows to have RHS=0 and RANGES=inf
• May try to read/write a basis for a model without names.
• May try to write a basis for a model without names.
• ems_prts prints status PWL for non-basic PWL variables: ems_wr_st_t_ch4 can't tell if they are at LB/BP/UB.
• ems_reset_1_(non_)bc_bp_vr_st updates rsmi_lb, rsmi_co, rsmi_ub by forming costs by adding delta causing (unlikely) unnecessary rounding error if upper/lower costs differ greatly in magnitude.
• ems_reset_1_bc_bp_vr_st
• Sort out hot start with scaling and when rsmi_bd are there or not.
• Revise Add-row, Add-column for new basis scheme and when additional block of new values is needed.
• Complete NW_MTX when several blocks of new of new values are needed. Can't extend space for matrix if mx_n_a_el is exceeded. Scaling?
• Look at GEMV with scaling (and new blocks of values?)

Can wait

• Is mod guaranteed to be >=0 on IBM, SGI or PENTIUM?
• Is ems_iz_sos needed?
• Update ems_se_com_undn
• Error message when trying to add mx_n_r+1st row is opaque: what does OSL do? Ditto for mx_n_c+1st column.
• In ems_nget, what pointers should be returned if no user copy of the matrix has been made?
• ems_iz_btran_eta_fi_p is coded for multiple eta groups containing multiple eta sets for an INVERT but not tested since INVERT only returns one eta group containing a single set.
• BTRAN/FTRAN coded for eta groups containing multiple eta sets but not tested.
• Bounds on primal infeasibility tolerance changed from [1d-12, 1d-1] to [0,inf) to allow experiments.
• ems_ope_u_eta_grp should be able to see if any data such as other INVERTs can be deleted and the memory manager re-compressed rather than returning a serious error.

Unimportant

• ems_irand for IBM---ems_irand is not used anywhere!
• The tests for error in ems_ck_inv should exploit knowledge of indices of non-zeros.

  ---

  Last modified: