These functions continue optimization for a MIP-based array from a stored state.
mosek_MIPcontinue(qco, improve = TRUE, maxtime = Inf, nthread = 2,
mosek.opts = list(verbose = 10, soldetail = 1),
mosek.params = list(dparam = list(LOWER_OBJ_CUT = 0.5,
MIO_TOL_ABS_GAP = 0.2, INTPNT_CO_TOL_PFEAS = 1e-05,
INTPNT_CO_TOL_INFEAS = 1e-07),
iparam = list(PRESOLVE_LINDEP_USE="OFF", LOG_MIO_FREQ=100)))
gurobi_MIPcontinue(qco, improve = TRUE, maxtime = 60, nthread = 2,
heurist = 0.05, MIQCPMethod = 0, MIPFocus = 0,
gurobi.params =list(BestObjStop = 0.5, LogFile=""))an array of class link[DoE.base]{oa}, if not optimized to GMA with info for further continuation (see documentation of mosek_MIParray or gurobi_MIParray)
object of class qco, created by a function in the package; or object of class oa that has a qco object as its MIPinfo attribute
logical; if TRUE (default), try to improve the already obtained solution for word length qco$info$last.k, otherwise optimize the next word length
time in seconds for the optimization call; defaults differ for Mosek (Inf) and Gurobi (60), because a Mosek run can be easily escaped (<ESC>-key), contrary to a Gurobi run
number of cores to use (0=all cores)
CAUTION: nthread should not exceed the available number of cores. Gurobi warns that performance might deteriorate. For Mosek, performance WILL strongly deteriorate, and for extreme choices the R session might even crash (even for small problems)!
for gurobi_MIPcontinue only: the percentage of time (number between 0 and 1) spent on heuristics
for gurobi_MIPcontinue only: the choice of optimization method; the default "0" has been observed to be better than Gurobi's default for most cases (Gurobi version 7.5.1); "-1" leaves the choice to Gurobi, "1" chooses the other method
for gurobi_MIPcontinue only: the choice of strategy; the default "0" leaves this choice to Gurobi; for finding better feasible solutions, "1" is recommended; for improving the speed of increasing the lower bound for eventually proving optimality, "3" can be tried
mosek options
Mosek parameters
Gurobi parameters
Ulrike Groemping
Note that it is possible to continue optimization with Gurobi, if it was started with Mosek, and vice versa. The tool will transform the problem into the respective other format.
Usage of options is analogous to functions mosek_MIParray and gurobi_MIParray, respectively, where these are described in more detail.
For some applications, usability of mosek_MIPcontinue is hampered in Mosek versions up to 8 by the fact that Mosek's presolve routines identify additional integer variables and fail to recognise user-specified starting values for these that are not exactly integer-valued. According to Mosek ApS, this is scheduled to be remedied with Mosek Version 9 (version 9 is now available; I have not checked whether this was indeed fixed).
See also DoE.MIParray for examples of the role of the MIPcontinue functions,
mosek_MIParray and gurobi_MIParray for more detail on the optimization arguments,
mosek_MIPsearch and gurobi_MIPsearch for searching over nlevels orderings
(which may be a very successful alternative to trying to improve an initial optimization based on a fixed nlevels vector).