The solver "xpress" uses FICO XPRESS (by FICO,
http://www.fico.com/xpress) to solve integer, mixed-integer, linear
programming, and quadratic programming problems. Normally xpress is
invoked by AMPL's solve command, which gives the invocation
xpress stub -AMPL
in which stub.nl is an AMPL generic output file (possibly written
by "ampl -obstub" or "ampl -ogstub"). After solving the problem,
xpress writes a stub.sol file for use by ampl's solve and solution
commands. When you run ampl, this all happens automatically if you
give the AMPL commands
option solver xpress;
solve;
You can control xpress by setting the environment variable xpress_options
appropriately (either by using ampl's option command, or by using the
shell's set and export commands before you invoke ampl). You can put
one or more (white-space separated) phrases in $xpress_options. For
details, invoke
xpress -=
or perhaps
xpress '-='
(depending on your shell). The output this gives is shown at the end
of this file.
When XPRESS finds the problem unbounded and a basis is available,
a ray is returned in suffix unbdd.
For problems with several integer (or binary) variable declarations,
it sometimes helps to specify branching priorities for the integer
variables. When XPRESS has a choice of which integer variable to
bound (or fix) during its branch-and-bound algorithm, it chooses
a variable with the highest priority. You can declare suffix priority
and assign individual priorities (between 0 and 2147483647) to each
integer variable.
You can also declare suffix direction and use it to convey individual
branching directions (-1 for down, 0 for no specification, or 1 for
up) for each integer variable.
----------
INSTALLING
==========
On Linux systems, libxprl.so.* and libxprs.so.* (where the values of
"*" depend on the current version of XPRESS) and the lib* files to
which they point need to appear in the current directory when xpress
itself appears there, or in one of the standard places (specified by
/etc/ld.so.conf on some systems), or in a directory named in
$LD_LIBRARY_PATH. An alternative is to add a short shell script,
such as
#!/bin/sh
LD_LIBRARY_PATH=/usr/local/lib
export LD_LIBRARY_PATH
exec /usr/local/bin/xpressx "$@"
to a directory in your usual $PATH (and mark the script executable
with, e.g., "chmod +x xpress"). The above script assumes that the
true "xpress" binary has been moved to /usr/local/bin/xpressx and that
the libxpr* files have been moved to /usr/local/lib.
On MacOSX systems, it suffices for the libxprl.so.* and libxprs.so.*
files to appear in the same directory as xpress; otherwise, MacOSX
systems are similar to Linux systems, but with DYLD_LIBRARY_PATH in
place of LD_LIBRARY_PATH.
On MS Windows systems, xpress.exe and the relevant xprl.dll and
xprs.dll must appear somewhere in your usual search $PATH (or in the
current directory).
-----------------------
solve_result_num values
=======================
Here is a table of solve_result_num values that "xpress" can return
to an AMPL session, along with the text that appears in the associated
solve_message.
Value Message
000 Optimal solution found
001 LP has been optimized
002 Global search complete
100 Objective is worse than cutoff
101 Cutoff in dual
102 Global search incomplete
200 Infeasible problem
201 Global search complete - no integer solution found
300 Unbounded problem
301 Unbounded problem with some integer variables
400 Unfinished optimization
401 Global search incomplete - no integer solution found
500 Problem has not been loaded (should not happen)
501 LP has not been optimized (probably LP Infeasible)
502 Problem unsolved (should not happen)
510 Problem is not convex (indefinite QP)
520 Multiple SIGINTs (Control-C's) received
Additionally, for values < 200, solve_result_num is increased by 10 if
integer variables must be rounded to integer values.
---------------------------------------
Keyword listing (output of "xpress -=")
=======================================
advance whether to use an initial basis, if available:
0 = no, overriding mipstartstatus;
1 = yes (default), subject to mipstartstatus.
In an AMPL session, "option send_statuses 0;" is preferable
to "option xpress_options '... advance=0 ...';".
algaftercrossover algorithm for final cleanup after running the barrier
algorithm:
1 = automatic choice (default)
2 = dual simplex
3 = primal simplex
4 = concurrent
algafternetwork algorithm for final cleanup after the network simplex
algorithm:
1 = automatic choice (default)
2 = dual simplex
3 = primal simplex
autoperturb whether to introduce perturbations when the simplex method
encounters too many degenerate pivots:
1 = yes (default); 0 = no
backtrack choice of next node when solving MIP problems:
-1 = automatic choice (default)
1 = withdrawn; formerly choice 2 until a feasible
integer solution has been found, then
Forrest-Hirst-Tomlin choice
2 = node with best estimated solution
3 = node with best bound on the solution (default)
4 = deepest node (depth-first search)
5 = highest node (breadth-first search)
6 = earliest-created node
7 = most recently created node
8 = random choice
9 = node with fewest LP relaxation infeasibilities
10 = combination of 2 and 9
11 = combination of 2 and 4
backtracktie how to break ties for the next MIP node: same choices as
for "backtrack"
baralg which barrier algorithm to use with "barrier":
-1 = automatic choice (default with just "barrier")
1 = infeasible-start barrier algorithm
2 = homogeneous self-dual barrier algorithm
3 = start with 2 and maybe switch to 1 while solving
barcores if positive, number of CPU cores to assume present when
using the barrier algorithm. Default = -1, which means
automatic choice.
barcrash choice of crash procedure for crossover:
0 = no crash
1-6 = available strategies (default 4):
1 = most conservative, 6 = most agreessive
bardualstop barrier method convergence tolerance on
dual infeasibilities; default = 0 (automatic choice)
bargapstop barrier method convergence tolerance on the relative
duality gap; default = 0
barindeflimit maximum indefinite factorizations to allow in the barrier
algorithm for solving a QP: stop when the limit is hit;
default = 15
bariterlimit maximum number of Newton Barrier iterations; default = 500
barobjscale how the barrier algorthm scales the objective:
-1 = automatic chocie (default)
0 = scale by the geometric mean of the objective
coefficients
> 0 = scale so the argest objective coefficient in
absolute value is <= barobjscale.
When the objective is quadratic, the quadratic diagonal
is used in determining the scale.
barorder Cholesky factorization pivot order for barrier algorithm:
0 = automatic choice (default)
1 = minimum degree
2 = minimum local fill
3 = nested dissection
barorderthreads number of threads to use when choosing a pivot order for
Cholesky factorization; default 0 ==> automatic choice.
baroutput amount of output for the barrier method:
0 = no output
1 = each iteration (default)
barpresolve level of barrier-specific presolve effort:
0 = use standard presolve (default)
1 = use more effort
2 = do full matrix eliminations for size reduction
barprimalstop barrier method convergence tolerance on
primal infeasibilities; default = 0 (automatic choice)
barreg regularization to use with "barrier":
-1 = automatic choice (default with just "barrier")
Values >= 0 are the sum of:
1 = use "standard" regularization
2 = use "reduced" regularization: less perturbation
than "standard" regularization
4 = keep dependent rows in the KKT system
8 = keep degenerate rows in the KKT system
barrier [no assignment] use the Newton Barrier algorithm
barstart choice of starting point for barrier method:
0 = automatic choice (default)
1 = heuristics based on magnitudes of matrix entries
2 = use pseudoinverse of constraint matrix
barstepstop barrier method convergence tolerance: stop when
step size <= barstepstop; default = 1e-10
barthreads number of threads used in the Newton Barrier algorithm;
default = -1 (determined by "threads")
basisin load initial basis from specified file
basisout save final basis to specified file
bestbound [no assignment] return suffix .bestbound for the best known
bound on the objective value. The suffix is on the problem
and objective and is +Infinity for minimization problems and
-Infinity for maximization problems if there are no integer
variables or if an integer feasible solution has not yet
been found.
bigm infeasibility penalty; default = 1024
bigmmethod 0 = phase I/II, 1 = BigM method (default)
branchchoice whether to explore branch with min. or max. estimate first:
0 = explore branch with min. estimate first (default)
1 = explore branch with max. estimate first
2 = if an incumbent solution exists, first explore
the branch satisfied by the incumbent;
otherwise use choice 0 (min. est. first)
default = 3
3 = explore the first branch that moves the branching
variable away from its value at the root node;
if the branching entity is not a simple
variable, assume branchchoice=0
branchdisj whether to branch on general split disjunctions while
solving MIPs:
-1 = automatic choice (default)
0 = disabled
1 = cautious strategy: create branches only for
general integers with a wide range
2 = moderate strategy
3 = aggressive strategy: create disjunctive branches
for both binary and integer variables
branchstruct whether to search for special structure during branch and
bound:
-1 = automatic choice (default)
0 = no
1 = yes
breadthfirst number of MIP nodes included in best-first search
(default 11) before switching to local-first search
cachesize cache size in Kbytes -- relevant to Newton Barrier:
-1 = determined automatically
default = system-dependent (-1 for Intel)
choleskyalg type of Cholesky factorization used for barrier: sum of
1 ==> manual matrix blocking
2 ==> single pass with manual blocking
4 ==> nonseparable QP relaxation
8 ==> manual corrector weight (honor "16" bit)
16 ==> manual corrector weight "on"
32 ==> manual refinement
64 ==> use preconditioned conjugate gradients
128 ==> refine with QMR (quasi-minimal residual)
default = -1 (automatic choice)
choleskytol zero tolerance for Cholesky pivots in the
Newton Barrier algorithm; default = 1e-15
concurrentthreads synonym for lpthreads
conedecomp whether to decompose regular and rotated cone constraints
having more than two elements and to use the result in an
outer approximation:
-1 = automatic choice (default)
0 = no
1 = yes, unless the cone variable is fixed by XPRESS's
presolve
2 = yes, even if the cone variable is fixed
3 = yes, but only for outer approximations
convexitychk whether to check convexity before solving:
0 = no
1 = yes (default)
corespercpu number of cores to assume per cpu; default = -1 ==> number
detected; barrier cache = cachesize / corespercpu
covercuts for MIPS, the number of rounds of lifted-cover inequalities
at the top node; default = -1 ==> automatic choice
cpuplatform whether the Newton Barrier method should use AVX or SSE2
instructions on platforms that offer both:
-1 = automatic choice (default)
0 = use generic code: neither AVX nor SSE2
1 = use SSE2
2 = use AVX
3 = use AVX2
cputime which times to report when logfile is specified:
0 = elapsed time (default)
1 = CPU time
2 = process time
You may need to experiment to see how cputime=1 and
cputime=2 differ (if they do) on your system.
crash type of simplex crash:
0 = none
1 = one-pass search for singletons
2 = multi-pass search for singletons (default)
3 = multi-pass search including slacks
4 = at most 10 passes, only considering slacks
at the end
n = (for n > 10) like 4, but at most n-10 passes
crossover whether to find a simplex basis after the barrier alg.:
-1 = automatic choice (default)
0 = no crossover
1 = primal crossover first
2 = dual crossover first
crossovertol tolerance (default 1e-6) for deciding whether to adjust the
relative pivot tolerance during crossover when a new basis
factorization is necessary. Errors in the recalculated
basic solution above this tolerance cause the pivot
tolerance to be adjusted.
cutdepth maximum MIP tree depth at which to generate cuts:
0 = no cuts
-1 = automatic choice (default)
cutfactor limit on number of cuts and cut coefficients added
while solving MIPs:
-1 = automatic choice (default)
0 = do not add cuts
> 0 ==> multiple of number of original constraints
cutfreq MIP cuts are only generated at tree depths that are integer
multiples of cutfreq; -1 = automatic choice (default)
cutselect detailed control of cuts at MIP root node: sum of
16 = clique cuts
32 = mixed-integer founding (MIR) cuts
64 = lifted cover cuts
1024 = flow path cuts
2048 = implication cuts
4096 = automatic lift-and-project strategy
8192 = disable cutting from cut rows
16384 = lifted GUB cover cuts
-1 = all available cuts (default)
cutstrategy how aggressively to generate MIP cuts; more ==> fewer nodes
but more time per node:
-1 = automatic choice (default)
0 = no cuts
1 = conservative strategy
2 = moderate strategy
3 = aggressive strategy
defaultalg algorithm to use when none of "barrier", "dual", or "primal"
is specified:
1 = automatic choice (default)
2 = dual simplex
3 = primal simplex
4 = Newton Barrier
densecollimit number of nonzeros above which a column is treated as dense
in the barrier algorithm's Cholesky factorization:
0 = automatic choice (default)
deterministic whether a MIP search should be deterministic:
0 = no
1 = yes (default)
dual [no assignment] use the dual simplex algorithm
dualgradient dual simplex pricing strategy:
-1 = automatic choice
0 = Devex
1 = steepest edge
dualize whether to convert the primal problem to its dual and solve
the converted problem:
-1 = automatic choice (default)
0 = no: solve primal problem
1 = yes: solve dual problem
dualizeops when solving the dual problem after deriving it from the
primal, whether to use primal simplex if dual simplex was
specified and vice versa:
0 = no
1 = yes (default)
dualstrategy how to remove infeasibilities when re-optimizing
with the dual algorithm during MIP solves:
0 = use primal algorithm
1 = use dual algorithm (default)
dualthreads limit on number of threads used by parallel dual simplex,
overriding "threads"; default -1 ==> use "threads"
eigenvaltol regard the matrix in a quadratic form as indefinite if its
smallest eigvenalue is < -eigevnaltol; default = 1e-6
elimtol Markowitz tolerance for the elimination phase of
XPRESS's presolve; default = 0.001
etatol zero tolerance on eta elements; default varies with XPRESS
version; default = 1e-12 or 1e-13 with some versions.
Use etatol=? to see the current value.
feaspump whether to run the Feasibility Pump heuristic at the top
node during branch-and-bound: one of
0 = no (default)
1 = yes
2 = only if other heurstics found no integer solution
feastol zero tolerance on RHS; default = 1e-6
feastol_target feasibility tolerance on constraints for solution refiner
(see refineops): if feastol_target > 0 is specified, it is
used instead of feastol
gomcuts gomory cuts at root: -1 = automatic choice (default)
hdive_rand value between 0 and 1 inclusive affecting randomization
in the diving heuristic: 0 (default) ==> none;
1 ==> full;
intermediate values ==> intermediate behavior
hdive_rounding whether to use soft rounding in the MIP diving heuristic
(to push variables to their bounds via the objective rather
than fixing them):
-1 = automatic choice (default)
0 = no soft rounding
1 = cautious soft rounding
2 = aggressive soft rounding
hdive_speed controls tradeoff between speed and solution quality
in the diving heuristic: an integer between -2 and 3:
-2 = automatic bias toward quality
-1 = automatic bias toward speed (default)
0 = emphasize quality
4 = emphasize speed
1-3 = intermediate emphasis
hdive_strategy strategy for diving heuristic: integer between -1 and 10:
-1 = automatic choice (default)
0 = do not use the diving heursistic
1-10 = preset strategies for diving
heurdepth deprecated: no longer has any effect:
maximum depth of branch-and-bound tree search at which to
apply heuristics; 0 = no heuristics; default = -1
heureffort factor affecting how much work local search heuristics
should expend. Default = 1; higher values cause more
local searches over larger neighborhoods.
heurfreq during branch and bound, heuristics are applied at nodes
whose depth from the root is zero modulo heurfreq;
default = -1 (automatic choice)
heurmaxsol deprecated: no longer has any effect:
maximum number of heuristic solutions to find during branch-
and-bound tree search; default = -1 (automatic choice)
heurnodes deprecated: no longer has any effect:
maximum nodes at which to use heuristics during
branch-and-bound tree search; default = 1000
heurroot bit vector controlling local search heuristics to apply at
the root node: sum of
1 = large-neighborhood search: may be slow, but may
find solutions far from the incumbent
2 = small-neighborhood search about node LP solution
4 = small-neighborhood search about integer solutions
8 = local search near multiple integer solutions
16 = no effect
32 = local search without an objective; may only be
done when no feasible solution is available
default = 53
heurrootcutfreq how often to run the local search heuristic while
cutting at the root node:
-1 ==> automatic choice (default)
0 ==> never
n > 0 ==> do n cutting rounds between runs of the
local search heuristic
heursearch how often the local search heurstic should be run during
branch-and-bound:
-1 = automatic choice (default)
0 = never
n > 0 ==> every n nodes
heurstrategy heuristic strategy for branch and bound: one of
-1 = automatic choice (default)
0 = no heuristics
1 = basic heuristics
2 = enhanced heuristics
3 = extensive heuristics
heurthreads number of threads for the root node of
branch-and-bound:
-1 = determined from "threads" keyword
0 = no separate threads (default)
n > 0 ==> use n threads
heurtree heuristics to apply during tree search: sum of
the same values as for heurroot; default 17
iis [no assignment] if the problem is infeasible, find an
Irreducible Independent Set of infeasible constraints
and return it in suffix .iis. If changing the bounds
on just one constraint or variable could remove the
infeasibility, return suffix .iso with value 1 for
each such constraint or variable.
indlinbigm largest "big M" value to use in converting indicator
constraints to regular constraints; default = 1e5.
indprelinbigm largest "big M" value to use in converting indicator
constraints to regular constraints during XPRESS
presolve; default = 100.0
invertfreq maximum simplex iterations before refactoring the basis:
-1 = automatic choice (default)
invertmin minimum simplex iterations before refactoring the basis:
default = 3
keepbasis basis choice for the next LP iteration:
0 = ignore previous basis
1 = use previous basis (default)
2 = use previous basis only if the number of basic
variables == number of constraints
keepnrows 1 (default) if unconstrained rows are to be kept, else 0
lazy whether to regard constraints with nonzero .lazy suffix
values as lazy (i.e., delayed) constraints if the problem
is a MIP:
0 = no
1 = yes (default)
lnpbest number of global infeasible entities for which to create
lift-and-project cuts during each round of Gomory cuts
at the top node; default = 50
lnpiterlimit maximum iterations for each lift-and-project cut;
default = -1 (automatic choice)
localchoice when to backtrack between two child nodes
during a "dive":
1 = never backtrack from the first child unless it
is dropped (i.e., is infeasible or cut off)
2 = always solve both nodes first
3 = automatic choice (default)
logfile name of log file; default = no log file
lpiterlimit simplex iteration limit; default = 2147483647 = 2^31 - 1
lplog frequency of printing simplex iteration log; default = 100
lpref_itlim limit on simplex iterations used by the solution refiner
(see refineops); default = -1 ==> automatic choice
lpthreads number of threads in concurrent LP solves:
-1 = determine from "threads" keyword (default)
n > 0 ==> use n threads
markowitztol Markowitz tolerance used when factoring the basis matrix
default = 0.01
matrixtol zero tolerance on matrix elements; default = 1e-9
maxcuttime maximum time (CPU seconds) to spend generating cuts
and reoptimizing; default = 0 ==> no limit
maxiis maximum number of Irreducible Infeasible Sets to find:
-1 = no limit (default)
0 = none
maxim [no assignment] force maximization of the objective
maximise [no assignment] force maximization of the objective
maximize [no assignment] force maximization of the objective
maximpliedbound when preprocessing MIP problems, only use computed bounds
at most maximpliedbound (default 1e8) in absolute value
maxlocalbt max height above current node to look for a local backtrack
candidate node; default = 1
maxlogscale max log2 of factors used in scaling; must be >= 0 and
<= 64; default 64
maxmemory limit (integer number of megabytes) on memory used:
-1 = automatic choice (default)
>0 = target megabytes of memory to use
maxmipsol maximum number of integer solutions to find:
0 = no limit (default)
maxmiptasks maximum tasks to run in parallel during a MIP solve:
-1 ==> use mipthreads
n > 0 ==> at most n tasks running at once
For maxmiptasks > 0, branch-and-bound nodes are solved in a
deterministic way, but the barrier algorithm (if used) may
cause a nondeterministic MIP solve unless barthreads = 1.
maxnode maximum number of MIP nodes to explore; default = 2147483647
maxpagelines maximum output lines between page breaks in logfile;
default = 23
maxtime maximum solution time allowed; default = 0 ==> no limit
minim [no assignment] force minimization of the objective
minimise [no assignment] force minimization of the objective
minimize [no assignment] force minimization of the objective
mipabscutoff initial MIP cutoff: ignore MIP nodes with objective values
worse than mipabscutoff; default = 1e40 for minimization,
-1e40 for maximization
mipabsstop stop MIP search if abs(MIPOBJVAL - BESTBOUND) <= mipabsstop
default = 0
mipaddcutoff amount to add to the objective function of the best integer
solution found to give the new MIP cutoff; default -1e-5
miplog MIP printing level to logfile (default -100):
-n = print summary line every n MIP nodes
0 = no MIP summary lines
1 = only print a summary at the end
2 = log each solution found
3 = log each node
mipops MIP solver options: one of
0 = traditional primal first phase (default)
1 = Big M primal first phase
2 = traditional dual first
3 = Big M dual first
4 = always use artificial bounds in dual
5 = use original basis only when warmstarting
6 = skip primal bound flips for ranged primals
7 = also do single-pivot crash
8 = suppress aggressive dual perturbations
mippresolve MIP presolve done at each node: sum of
1 = reduced-cost fixing
2 = logical preprocessing of binary variables
4 = ignored; replaced by "preprobing"
8 = allow changing continuous-variable bounds
16 = allow dual reductions
32 = use objective function
default = -1 (automatic choice)
miprefiterlim max. simplex iterations per reoptimization in MIP refiner
when refineops is 2 or 3; default -1 ==> automatic choice
miprelcutoff fraction of best integer solution found to add to MIP
cutoff; default 1e-4
miprelstop stop MIP search if
abs(MIPOBJVAL - BESTBOUND) < miprelstop * abs(BESTBOUND);
default = 0.0001
mipstart synonym for mipstartvalue
mipstartstatus whether to use incoming statuses on MIP problems;
default 1 ==> yes
mipstartvalue whether to use the specified initial guess (if supplied)
when solving a MIP problem:
0 = no
1 = yes (default)
mipstop how to stop a MIP solve when a time or node limit is
reached:
0 = stop tasks as soon as possible (default)
1 = let currently running tasks finish, but do not
start new ones
mipthreads number of threads to use solving mixed-integer
programming problems:
-1 = use "threads" keyword (default)
n > 0 ==> use n threads
miptol integer feasibility tolerance; default = 5e-6
miptoltarget value of miptol used for refining equalities on MIP
problems when refineops is 2 or 3; default = 0
miqcpalg algorithm for solving mixed-integer problems with quadratic
or second-order cone constraints:
-1 = automatic choice (default)
0 = barrier algorithm during branch and bound
1 = outer approximations during branch and bound
network [no assignment] try to find and exploit an embedded network
nodefilebias a value between 0 and 1 (inclusive) that influences
operations when "treememlimit" (on how much of the
branch-and-bound tree should be kept in memory) has
been exceeded:
0 ==> compress every node before writing anything to
the "nodefile";
1 ==> write nodes to the "nodefile" immediately;
values between 0 and 1 give intermediate behavior;
default = 0.5
nodeselection next MIP node control:
1 = local first: choose among descendant and sibling
nodes if available, else from all outstanding nodes
2 = best first of all outstanding nodes
3 = local depth first: choose among descendant and
sibling nodes if available, else from deepest nodes
4 = best first for breadthfirst nodes, then local first
5 = pure depth first: choose among deepest nodes.
The default is determined from matrix characteristics.
objno objective number (0=none, 1=first...)
objrep Whether to replace
minimize obj: v;
with
minimize obj: f(x)
when variable v appears linearly in exactly one
constraint of the form
s.t. c: v >= f(x);
or
s.t. c: v == f(x);
Possible objrep values:
0 = no
1 = yes for v >= f(x)
2 = yes for v == f(x) (default)
3 = yes in both cases
For a maximization problem, "<=" replaces ">=".
optimalitytol tolerance on reduced cost; default = 1e-6
opttol_target feasibility tolerance on reduced costs for solution refiner
(see refineops): default = 0; if opttol_target > 0 is
specified, it is used instead of optimalitytol.
outlev message level:
1 = all
2 = information
3 = warnings & errors only (default)
4 = errors
5 = none
outputtol zero tolerance on print values; default 1e-5
param Used with syntax "param=name=value" (no spaces), where
"name" is the name of an XPRESS control parameter and
"value" is to be assigned to that parameter. If value is
?, report the current value of the parameter. If name is
a string control, value can be a quoted string or a
sequence of nonblank characters other than comma. This
facility provides a way to modify control parameters,
identified by name or number, that have not (yet) been
assigned a keyword. As a special case, "param=?" requests
a list of all control parameters and their current values.
penalty minimum absolute penalty variable coefficient;
default = automatic choice
permuteseed seed for the random-number generator used by prepermute;
default = 1
perturb perturb factor if autoperturb is set to 1;
0 = default = automatic choice
pivottol zero tolerance for pivots; default = 1e-9
pooldualred Whether to suppress removal of dominated solutions (via
"dual reductions") when poolstub is specified:
0 = yes (default, which can be expensive)
1 = no
2 = honor presolveops bit 3 (2^3 = 8)
pooldupcol Whether to suppress duplicate variable removal when
poolstub is specified:
0 = yes (default, which can be expensive)
1 = no
2 = honor presolveops bit 5 (2^5 = 32)
pooldups How poolstub should handle duplicate solutions:
0 = retain all duplicates
1 = discard exact matches
2 = discard exact matches of continuous variables
and matches of rounded values of discrete
varibles
3 = default: discard matches of rounded values of
discrete variables
Rounding of discrete variables is affected by poolmiptol
and poolfeastol.
poolfeastol Zero tolerance for discrete variables in the solution
pool (see poolstub); default = 1e-6.
poolmiptol Error (nonintegrality) allowed in discrete variables
in the solution pool (see poolstub); default = 5e-6.
poolnbest Whether the solution pool (see poolstub) should contain
inferior solutions. When poolstub = n > 1, the
solution pool is allowed to keep the n best solutions.
poolstub Stub for solution files in the MIP solution pool.
Ignored unless some variables are integer or binary.
A pool of alternate MIP solutions is computed if
poolstub is specified, and the solutions in this pool
are written to files
(poolstub & '1') ... (poolstub & |solution pool|),
where |solution pool| is the number of solutions in the
solution pool. That is, file names are obtained by
appending 1, 2, ... |solution pool| to poolstub. The
value of |solution pool| is returned in suffix npool
on the objective and problem.
ppfactor partial-pricing candidate-list size factor; default = 1.0
prebndredcone for MIP problems, whether to use cone constraints to
reduce bounds on variables:
0 = no
1 = yes
-1 = default (undocumented)
prebndredquad for MIP problems, whether to use convex quadratic
constraints to reduce bounds on variables:
0 = no
1 = yes
-1 = default (undocumented)
precoefelim whether XPRESSMP's presolve should recombine constraints:
0 = no,
1 = yes, as many as possible
2 = yes, cautiously (default)
precomponents whether XPRESS's presolve should detect and separately
solve independent MIP subproblems:
-1 = automatic choice (default)
0 = no
1 = yes
predomcol whether XPRESSMP's presolve should remove variables
when solving MIP problems:
-1 = automatic choice (default)
0 = no
1 = yes, cautiously
2 = yes, check all candidates
predomrow whether XPRESSMP's presolve should remove constraints
when solving MIP problems:
-1 = automatic choice (default)
0 = no
1 = yes, cautiously
2 = yes, medium strategy
3 = yes, check all candidates
preduprow how XPRESS's presolve should deal with duplicate rows
in MIP problems:
-1 = automatic choice (default),
0 = do not remove duplicate rows (constraints)
1 = remove duplicate rows identical in all variables
2 = like 1 but allowing simple penalty variables
3 = like 1 but allowing more complex penalty variables
prelindep whether to check for and remove linearly dependent
equality constraints:
-1 = automatic choice (default)
0 = no
1 = yes
preobjcutdetect on MIP problems, whether to check for constraints
that are (nearly) parallel to a linear objective function
and can be removed safely:
0 = no
1 = yes (default)
prepermute whether to randomly permute variables or constraints before
applying XPRESS's presolve: sum of
1 ==> permute constraints
2 ==> permute variables
4 ==> permute global MIP information
default = 0; see permuteseed
preprobing how much probing on binary variables to do during XPRESSMP's
presolve:
-1 = automatic choice (default)
0 = none
1 = light probing
2 = full probing
3 = repeated full probing
presolve whether to use XPRESS's presolver:
0 = no
1 = yes, removing redundant bounds (default)
2 = yes, retaining redundant bounds
presolvemaxgrow factor by which the number of nonzero coefficients
may grow during XPRESS's presolve; default = 0.1
presolveops reductions to use in XPRESSMP's presolve: sum of
1 = 2^0 = remove singleton columns
2 = 2^1 = remove singleton constraints (rows)
4 = 2^2 = forcing row removal (whatever that is)
8 = 2^3 = dual reductions
16 = 2^4 = redundant constraint (row) removal
32 = 2^5 = duplicate variable removal
64 = 2^6 = duplicate constraint removal
128 = 2^7 = strong dual reductions
256 = 2^8 = variable eliminations
512 = 2^9 = no IP reductions
1024 = 2^10 = no semicontinuous variable detection
2048 = 2^11 = no advanced IP reductions
16384 = 2^14 = remove linearly dependent constraints
32768 = 2^15 = no integer variable and SOS detection
default = 511 (bits 0-8 set)
pricingalg primal simplex pricing method:
-1 = partial pricing
0 = automatic choice (default)
1 = Devex pricing
primal [no assignment] use the primal simplex algorithm
primalunshift whether the primal alg. calls the dual to unshift:
0 = yes (default)
1 = no
pseudocost default pseudo-cost assumed for forcing an integer variable
to an integer value; default = 0.01
pseudocost_ud how to update pseudocosts during branch-and-bound:
-1 = automatic choice (default)
0 = no updates
1 = use only regular branches
2 = use regular and strong branch results
3 = use results from all nodes
qccuts when using miqcpalg=1 to solve a mixed-integer problem that
has quadratic constraints or second-order cone constraints,
the number of rounds of outer approximation cuts at the top
node: default = -1 means automatic choice.
qcrootalg when using miqcpalg=1 to solve a mixed-integer problem that
has quadratic constraints or second-order cone constraints,
the algorithm for solving the root node:
-1 = automatic choice (default)
0 = barrier algorithm
1 = dual simplex on outer approximations
quadunshift whether quadratic simplex should do an extra
purification after finding a solution:
-1 = automatic choice (default)
0 = no
1 = yes
ray whether to return a ray of unboundedness in suffix .unbdd:
0 ==> no (default)
1 ==> yes, after suppressing XPRESS's presolve
2 ==> yes, without suppressing XPRESS's presolve
The last setting (ray=2) may give wrong results when
XPRESS's presolve detects infeasibility. Both ray=1 and
ray=2 cause reoptimization with primal simplex if some other
algorithm was used. No ray is returned for MIP problems.
refineops whether refine equalities -- to reduce infeasibilities
in constraints that should hold as equalities: sum of
1 ==> refine LP solutions
2 ==> refine MIP solutions;
default = 3 (do both)
relax [no assignment] ignore integrality
relaxtreemem fraction of memory limit by which to relax "treememlimit"
when too much structural data appears; default 0.1
relpivottol relative pivot tolerance default = 1e-6
repairindefq whether to repair indefinite quadratic forms:
0 = yes
1 = no (default)
rootpresolve whether to presolve after root cutting and heuristics:
-1 = automatic choice (default)
0 = no
1 = yes
round whether to round integer variables to integral values before
returning the solution, and whether to report that XPRESS
returned noninteger values for integer values: sum of
1 ==> round nonintegral integer variables
2 ==> do not modify solve_result
4 ==> do not modify solve_message
8 ==> report modifications even if maxerr < 1e-9
Modifications take place only if XPRESS assigned nonintegral
values to one or more integer variables, and (for round < 8)
are reported if the maximum deviation from integrality
exceeded 1e-9. Default = 1.
sbbest For MIP problems, the number of infeasible
global entities on which to perform strong branching;
default -1 ==> automatic choice.
sbeffort multiplier on strong-branching controls that
are set to "automatic"; default = 1.0
sbestimate how to compute pseudo costs from the local node
when selecting an infeasible entity to branch on:
-1 = automatic choice (default)
1-6 = particular strategies (not described)
sbiterlimit Number of dual iterations to perform the strong branching;
0 ==> none; default = -1 (automatic choice)
sbselect size of candidate list for strong branching:
-2 = low-effort automatic choice (default)
-1 = high-effort automatic choice
n >= 0 ==> include max(n, sbbest) candidates
scaling how to scale the constraint matrix before optimizing: sum of
1 = 2^0 = row scaling
2 = 2^1 = column scaling
4 = 2^2 = row scaling again
8 = 2^3 = maximum scaling
16 = 2^4 = Curtis-Reid
32 = 2^5 = scale by maximum element (rather
than by geometric mean)
128 = 2^7 = objective-function scaling
256 = 2^8 = excluding quadratic part of constraint
when calculating scaling factors
512 = 2^9 = scale before presolve
1024 = 2^10 = do not scale constraints (rows) up
2048 = 2^11 = do not scale variables up
4096 = 2^12 = do global objective function scaling
8192 = 2^13 = do right-hand side scaling
default = 163
sleeponthreadwait whether threads should sleep while awaiting work:
0 = no (busy-wait)
1 = yes (sleep; may add overhead)
default = -1 (automatic choice)
sos whether to use explicit SOS information; default 1 ==> yes
sos2 whether to tell XPRESS about SOS2 constraints for
nonconvex piecewise-linear terms; default 1 ==> yes
sosreftol minimum relative gap between reference row entries;
default = 1e-6
symmetry amount of effort to detect symmetry in MIP problems:
0 = none: do not attempt symmetry detection
1 = modest effort (default)
2 = aggressive effort
tempbounds whether dual simplex should put temporary bounds on
unbounded variables:
-1 = automatic choice (default)
0 = no
1 = yes
threads default number of threads to use:
-1 = automatic choice (based on hardware)
n > 0 ==> use n threads
timing [no assignment] give timing statistics
trace whether to explain infeasibility:
0 = no (default)
1 = yes
treecompress level of effort at data compression when branch-and-bound
memory exceeds "treememlimit": higher ==> greater effort
(taking more time); default = 2
treecovercuts number of rounds of lifted-cover inequalities at MIP nodes
other than the top node (cf covercuts);
default = -1 (automatic choice)
treecuts cuts to generate at nodes during tree search: sum of
32 = 2^5 = clique cuts
64 = 2^6 = mixed-integer rounding (MIR) cuts
64 = 2^7 = lifted-cover cuts
2048 = 2^11 = flow-path cuts
4096 = 2^12 = implication cuts
8192 = 2^13 = lift-and-project cuts
16384 = 2^14 = disable cutting from row cuts
32768 = 2^15 = lifted GUB cover cuts
65536 = 2^16 = zero-half cuts
131072 = 2^17 = indicator cuts
default = 259839 (same effect as -2305)
treegomcuts number of rounds of Gomory cuts to generate at MIP nodes
other than the top node (cf covercuts);
default = -1 (automatic choice)
treememlimit an integer: soft limit in megabytes on memory to use for
branch-and-bound trees. Default = 0 ==> automatic choice.
treememtarget fraction of "treememlimit" to try to recover by compression
or writing to nodefile when "treememlimit" is exceeded.
Default = 0.2
treeoutlev how much to report about branch-and-bound trees
(if allowed by outlev): sum of
1 = regular summaries
2 = report tree compression and output to nodefile
default = 3
treepresolve how much presolving to apply to nodes of the MIP
branch-and-bound tree:
-1 = automatic choice (default)
0 = none
1 = cautious
2 = moderate
3 = aggressive
varselection how to score the integer variables at a MIP node, for
branching on a variable with minimum score:
-1 = automatic choice (default)
1 = minimum of the 'up' and 'down' pseudo-costs
2 = 'up' pseudo-cost + 'down' pseudo-cost
3 = maximum of the 'up' and 'down' pseudo-costs plus
twice their minimum
4 = maximum of the 'up' and 'down' pseudo-costs
5 = the 'down' pseudo-cost
6 = the 'up' pseudo-cost
version Report version details before solving the problem. This is
a single-word "phrase" that does not accept a value
assignment.
wantsol solution report without -AMPL: sum of
1 = write .sol file
2 = print primal variable values
4 = print dual variable values
8 = do not print solution message
----------------------------------------------
Mappings of keywords to Xpress parameter names
==============================================
For the possible convenience of readers of the Xpress-Optimizer
Reference Manual, the following table shows the Xpress parameter names
associated with the above keywords.
Keyword Xpress parameter
algaftercrossover XPRS_ALGAFTERCROSSOVER
algafternetwork XPRS_ALGAFTERNETWORK
autoperturb XPRS_AUTOPERTURB
backtrack XPRS_BACKTRACK
backtracktie XPRS_BACKTRACKTIE
baralg XPRS_BARALG
barcores XPRS_BARCORES
barcrash XPRS_BARCRASH
bardualstop XPRS_BARDUALSTOP
bargapstop XPRS_BARGAPSTOP
barindeflimit XPRS_BARINDEFLIMIT
bariterlimit XPRS_BARITERLIMIT
barobjscale XPRS_BAROBJSCALE
barorder XPRS_BARORDER
barorderthreads XPRS_BARORDERTHREADS
baroutput XPRS_BAROUTPUT
barpresolve XPRS_BARPRESOLVEOPS
barprimalstop XPRS_BARPRIMALSTOP
barreg XPRS_BARREGULARIZE
barstart XPRS_BARSTART
barstepstop XPRS_BARSTEPSTOP
barthreads XPRS_BARTHREADS
bigm XPRS_BIGM
bigmmethod XPRS_BIGMMETHOD
branchchoice XPRS_BRANCHCHOICE
branchdisj XPRS_BRANCHDISJ
branchstruct XPRS_BRANCHSTRUCTURAL
breadthfirst XPRS_BREADTHFIRST
cachesize XPRS_CACHESIZE
choleskyalg XPRS_CHOLESKYALG
choleskytol XPRS_CHOLESKYTOL
concurrentthreads XPRS_LPTHREADS
conedecomp XPRS_PRECONEDECOMP
convexitychk XPRS_IFCHECKCONVEXITY
corespercpu XPRS_CORESPERCPU
covercuts XPRS_COVERCUTS
cpuplatform XPRS_CPUPLATFORM
cputime XPRS_CPUTIME
crash XPRS_CRASH
crossover XPRS_CROSSOVER
crossovertol XPRS_CROSSOVERACCURACYTOL
cutdepth XPRS_CUTDEPTH
cutfactor XPRS_CUTFACTOR
cutfreq XPRS_CUTFREQ
cutselect XPRS_CUTSELECT
cutstrategy XPRS_CUTSTRATEGY
defaultalg XPRS_DEFAULTALG
degradefactor XPRS_DEGRADEFACTOR
densecollimit XPRS_DENSECOLLIMIT
deterministic XPRS_DETERMINISTIC
dualgradient XPRS_DUALGRADIENT
dualize XPRS_DUALIZE
dualizeops XPRS_DUALIZEOPS
dualstrategy XPRS_DUALSTRATEGY
dualthreads XPRS_DUALTHREADS
eigenvaltol XPRS_EIGENVALUETOL
elimtol XPRS_ELIMTOL
etatol XPRS_ETATOL
feaspump XPRS_FEASIBILITYPUMP
feastol XPRS_FEASTOL
feastol_target XPRS_FEASTOLTARGET
gomcuts XPRS_GOMCUTS
hdive_rand XPRS_HEURDIVERANDOMIZE
hdive_rounding XPRS_HEURDIVESOFTROUNDING
hdive_speed XPRS_HEURDIVESPEEDUP
hdive_strategy XPRS_HEURDIVESTRATEGY
heurdepth XPRS_HEURDEPTH
heureffort XPRS_HEURSEARCHEFFORT
heurfreq XPRS_HEURFREQ
heurmaxsol XPRS_HEURMAXSOL
heurnodes XPRS_HEURNODES
heurroot XPRS_HEURSEARCHROOTSELECT
heurrootcutfreq XPRS_HEURSEARCHROOTCUTFREQ
heursearch XPRS_HEURSEARCHFREQ
heurstrategy XPRS_HEURSTRATEGY
heurthreads XPRS_HEURTHREADS
heurtree XPRS_HEURSEARCHTREESELECT
indlinbigm XPRS_INDLINBIGM
indprelinbigm XPRS_INDPRELINBIGM
invertfreq XPRS_INVERTFREQ
invertmin XPRS_INVERTMIN
keepbasis XPRS_KEEPBASIS
keepnrows XPRS_KEEPNROWS
lnpbest XPRS_LNPBEST
lnpiterlimit XPRS_LNPITERLIMIT
localchoice XPRS_LOCALCHOICE
lpiterlimit XPRS_LPITERLIMIT
lplog XPRS_LPLOG
lpref_itlim XPRS_LPREFINEITERLIMIT
lpthreads XPRS_LPTHREADS
markowitztol XPRS_MARKOWITZTOL
matrixtol XPRS_MATRIXTOL
maxcuttime XPRS_MAXCUTTIME
maxiis XPRS_MAXIIS
maxlocalbt XPRS_MAXLOCALBACKTRACK
maxlogscale XPRS_MAXSCALEFACTOR
maximpliedbound XPRS_MAXIMPLIEDBOUND
maxmemory XPRS_MAXMEMORY
maxmipsol XPRS_MAXMIPSOL
maxmiptasks XPRS_MAXMIPTASKS
maxnode XPRS_MAXNODE
maxpagelines XPRS_MAXPAGELINES
maxslaves XPRS_MAXSLAVE
maxtime XPRS_MAXTIME
mipabscutoff XPRS_MIPABSCUTOFF
mipabsstop XPRS_MIPABSSTOP
mipaddcutoff XPRS_MIPADDCUTOFF
miplog XPRS_MIPLOG
mipops XPRS_QSIMPLEXOPS
mippresolve XPRS_MIPPRESOLVE
miprefiterlim XPRS_MIPREFINEITERLIMIT
miprelcutoff XPRS_MIPRELCUTOFF
miprelstop XPRS_MIPRELSTOP
mipstop XPRS_MIPTERMINATIONMETHOD
miptarget XPRS_MIPTARGET
mipthreads XPRS_MIPTHREADS
miptol XPRS_MIPTOL
miptoltarget XPRS_MIPTOLTARGET
miqcpalg XPRS_MIQCPALG
nodefilebias XPRS_GLOBALFILEBIAS
nodeselection XPRS_NODESELECTION
optimalitytol XPRS_OPTIMALITYTOL
opttol_target XPRS_OPTIMALITYTOLTARGET
outputtol XPRS_OUTPUTTOL
penalty XPRS_PENALTY
permuteseed XPRS_PREPERMUTESEED
perturb XPRS_PERTURB
pivottol XPRS_PIVOTTOL
pooldups XPRS_MSP_DUPLICATESOLUTIONSPOLICY
poolfeastol XPRS_MSP_SOL_FEASTOL
poolmiptol XPRS_MSP_SOL_MIPTOL
ppfactor XPRS_PPFACTOR
prebndredcone XPRS_PREBNDREDCONE
prebndredquad XPRS_PREBNDREDQUAD
precoefelim XPRS_PRECOEFELIM
precomponents XPRS_PRECOMPONENTS
predomcol XPRS_PREDOMCOL
predomrow XPRS_PREDOMROW
preduprow XPRS_PREDUPROW
prelindep XPRS_PRELINDEP
preobjcutdetect XPRS_PREOBJCUTDETECT
prepermute XPRS_PREPERMUTE
preprobing XPRS_PREPROBING
presolve XPRS_PRESOLVE
presolvemaxgrow XPRS_PRESOLVEMAXGROW
presolveops XPRS_PRESOLVEOPS
pricingalg XPRS_PRICINGALG
primalops XPRS_PRIMALOPS
primalunshift XPRS_PRIMALUNSHIFT
pseudocost XPRS_PSEUDOCOST
pseudocost_ud XPRS_HISTORYCOSTS
qccuts XPRS_QCCUTS
qcrootalg XPRS_QCROOTALG
quadunshift XPRS_QUADRATICUNSHIFT
refineops XPRS_REFINEOPS
relaxtreemem XPRS_RELAXTREEMEMORYLIMIT
relpivottol XPRS_RELPIVOTTOL
repairindefq XPRS_REPAIRINDEFINITEQ
rootpresolve XPRS_ROOTPRESOLVE
sbbest XPRS_SBBEST
sbeffort XPRS_SBEFFORT
sbestimate XPRS_SBESTIMATE
sbiterlimit XPRS_SBITERLIMIT
sbselect XPRS_SBSELECT
scaling XPRS_SCALING
sleeponthreadwait XPRS_SLEEPONTHREADWAIT
sosreftol XPRS_SOSREFTOL
symmetry XPRS_SYMMETRY
tempbounds XPRS_TEMPBOUNDS
threads XPRS_THREADS
trace XPRS_TRACE
treecompress XPRS_TREECOMPRESSION
treecovercuts XPRS_TREECOVERCUTS
treecuts XPRS_TREECUTSELECT
treegomcuts XPRS_TREEGOMCUTS
treememlimit XPRS_TREEMEMORYLIMIT
treememtarget XPRS_TREEMEMORYSAVINGTARGET
treeoutlev XPRS_TREEDIAGNOSTICS
treepresolve XPRS_TREEPRESOLVE
varselection XPRS_VARSELECTION
*************************
If you have questions about or find bugs with this stuff,
please contact:
David M. Gay
dmg@ampl.com