#
# Couenne options
#
# Couenne is an open-source solver for nonconvex Mixed-Integer 
# Nonlinear Programming (MINLP) problems. See more at 
# 
# https://projects.coin-or.org/Couenne
#
# The following is a list of option to tweak the performance of
# Couenne.  Each option has a brief description and is set to a
# default value that we believe works well in most cases.
#
#
# Some of the notation used here is close to that of the paper that
# describes Couenne:
#
# P. Belotti, J. Lee, L. Liberti, F. Margot, A. Waechter, "Branching
#   and bounds tightening techniques for non-convex MINLP," 2008.
# Available on Optimization Online at
#
# http://www.optimization-online.org/DB_HTML/2008/08/2059.html
#
# We refer the curious user of Couenne to this paper for more insight
# on how the options are used.

time_limit 15

#
# Verbosity/output level options
#

# display statistics at the end of the run [yes/no]. Set to yes for a
# line of output at the end with some brief data on the problem

display_stats no


# The following are verbosity levels for the main components of Couenne. 
# Values: 0 for quiet, 11 for excessive output [0..11]

branching_print_level       0 # Output level for branching code in Couenne.
boundtightening_print_level 0 # Output level for bound tightening code in Couenne
convexifying_print_level    0 # Output level for convexifying code in Couenne
problem_print_level         0 # Output level for problem manipulation code in Couenne 
                              # (4 prints out original problem)
nlpheur_print_level         0 # Output level for NLP heuristic in Couenne
#disjcuts_print_level       0 # Output level for disjunctive cuts in Couenne - disabled for now

#
# Option for branching rules
#

# Multipliers of pseudocosts for estimating and update estimation of bound
#
# When using pseudocosts, the lower bound of a node is estimated by multiplying 
# the pseudocost by a measure of the "infeasibility" of that variable.
#
# Valid Settings:
#   infeasibility (infeasibility returned by object)
#   projectDist (distance between current LP point and resulting branches' LP points)
#   interval_lp (width of the interval between bound and current lp point)
#   interval_lp_rev (similar to interval_lp, reversed)
#   interval_br (width of the interval between bound and branching point)
#   interval_br_rev (similar to interval_br, reversed)

pseudocost_mult interval_br_rev


# Use distance between LP points to update multipliers of
# pseudocosts. Can give a better estimate of the change in the node as
# a result of the branching rule.

pseudocost_mult_lp no


# Use Special Ordered Sets (SOS) as indicated in the MINLP
# model. Couenne recognizes constraints of the form 
# 
# f_1(x) + f_2(x) ... + f_n (x) = 1,
#
# where f_i (x) are binary expressions, as SOS constraints, and adds
# them to the Branch&Bound solver (disabled now -- still testing)

enable_sos no


# Apply bound tightening before branching
#
# Upon branching, it may be useful to apply a bound reduction
# technique as a preprocessing step for the node, even to check if the
# node is feasible

branch_fbbt yes


# Apply convexification cuts before branching (for now only within
# strong branching)
#
# This is useful to get a more precise lower bound within strong
# branching (note: does not work when performing the real branching
# rule)

branch_conv_cuts yes


# Chooses branching point selection strategy
#
#
# When branching on a continuous variable x that has a bound interval
# [l,u], the branching point is also important. Couenne implements
# several ways of computing the branching point, that may depend on
# the current solution of the LP relaxation or on the characteristics
# of the linearization that would result from the branching.
#
# The default value of this option is a convex combination 
#
#     alpha xp + (1-alpha) xm
#
#
# where xm is the middle point (l+u)/2, xp is the value of x in the
# current LP relaxation, and 0 <= alpha <= 1. Alpha is defined in the
# next option.
#
# Valid Settings:
#   lp-clamped (LP point clamped in [k,1-k] of the bound intervals (k defined by lp_clamp))
#   lp-central (LP point if within [k,1-k] of the bound intervals, 
#               middle point otherwise(k defined by branch_lp_clamp))
#   balanced (minimizes max distance from curve to convexification)
#   min-area (minimizes total area of the two convexifications)
#   mid-point (convex combination of current point and mid point)
#   no-branch (do not branch, return null infeasibility; for testing purposes only)

branch_pt_select mid-point


# Defines convex combination of mid point and current LP point. See
# comments on option "branch_pt_select" above.

#branch_midpoint_alpha 0.25


# Priority of continuous variable branching
#
# In Cbc, the Branch&Bound solver on which Couenne is based, integer
# variables have a priority of 1000. This parameter is the branching
# priority of continuous variables. Setting it to more than 1000 gives
# precedence to integer variables, i.e., as long as one integer
# variable is currently "infeasible" (i.e. fractional) it will be
# branched on.  A value below 1000 will give precedence to continuous
# variables. 

#cont_var_priority 2000


# Apply Reduced Cost Branching (instead of the Violation Transfer) --
# MUST have vt_obj enabled
#
#
# Violation Transfer and reduced cost branching are similar techniques
# for selecting a branching variable. Couenne implements both and lets
# you choose which one to use. Set this to yes to use reduced cost
# branching. Experimentally, Violation Transfer appears slightly
# better, hence it is preferred by default.

red_cost_branching no


# Type of branching object for variable selection
#
#
# This parameter determines the branching variable selection
# technique. With "vt_obj", the Violation Transfer branching technique
# is used. "var_obj" chooses a variable based on the set of nonlinear
# expressions that depends on it, while "expr_obj" selects the most
# violated nonlinear expression and branches on one of the variables
# on which the expression depends. The default is var_obj
#
#
# Valid Settings:
#   vt_obj    use Violation Transfer from Tawarmalani and Sahinidis
#   var_obj   use one object for each variable
#   expr_obj  use one object for each nonlinear expression

branching_object var_obj




#
# Options for bound tightening
#


# Feasibility-based (cheap) bound tightening (FBBT)
#
# A fast bound reduction technique. Not very efficient in eliminating
# vast portions of the solution set, but recommended.

feasibility_bt yes


# Optimality-based (expensive) bound tightening (OBBT)
#
# A slower bound reduction technique that relies on solving 2n LP
# problems (n is the number of variables). Probably more efficient
# that FBBT, but much more computationally intensive. Recommended 
# for small problems. See also the next option.

optimality_bt yes


# Specify the frequency (in terms of nodes) for optimality-based bound
# tightening
#
# As OBBT is expensive, the user may choose to run it only until the
# depth k of the branch&bound tree, and with probability exponentially
# decreasing with the depth of the branch&bound node at any other node
# below depth k.

log_num_obbt_per_level 1


# Aggressive feasibility-based bound tightening (to use with NLP points)
#
#
# See the paper for a detailed explanation. This is also an expensive
# but efficient way to reduce the solution set

aggressive_fbbt yes


# Specify the frequency (in terms of nodes) for aggressive bound tightening.
#
# A parameter analogous to what log_num_obbt_per_level is for OBBT.

log_num_abt_per_level 2


#
# Options for reformulation and linearization
#


# Specify the frequency (in terms of nodes) at which couenne ecp cuts
# are generated.
#
# A default value of 1 has linearization inequalities generated at
# every node.

convexification_cuts 1


# Specify the number of points at which to convexify when
# convexification type.

convexification_points 4


# Yes if only violated convexification cuts should be added 

violated_cuts_only yes



#
# Options for debugging
#
# Some of these options usually slow down Couenne, and are hence only suggested for debugging purposes.
#

# Check all LPs through an independent call to
#  OsiClpSolverInterface::initialSolve() (very expensive)
#check_lp no

# Artificial cutoff. Used when you know a feasible solution and want
# to use it to restrict the solution space

#art_cutoff 1e100

# Window around known optimum
#opt_window 

# Artificial lower bound
#art_lower -1e100


#
# Other options
#


# Do we search for local solutions of NLP's?
local_optimization_heuristic yes


# Specify the logarithm of the number of local optimizations to
# perform.
#
#
# Analogous to log_num_abt_per_level and log_num_obbt_per_level, this
# option determines until which depth of the branch&bound tree the
# call to a nonlinear solver is done at every node. Below this depth,
# calls to the nonlinear solver happen with probability inversely
# proportional to the depth of the node.

log_num_local_optimization_per_level 2


# Tolerance for constraints/auxiliary variables
#
#
# A solution is feasible for the original problem if the maximum
# violation of a constraint is below this number

feas_tolerance 1e-6


# Use quadratic expressions and related exprQuad class
#
# Allows to use a single operator for quadratic forms (not yet enabled)

use_quadratic no