def set_mip_parameters(self, param = None):
"""
updates MIP parameters
:param param:
:return:
"""
if param is None:
param = self._cpx_parameters
self._mip = set_cpx_parameters(self._mip, param)
def build_mip(self):
"""
returns an optimization problem that can be solved to determine an item in a flipset for x
:return:
"""
# MIP parameters
cost_type = self.mip_cost_type
min_items = max(self.min_items, 1)
max_items = self.max_items
# cost/action information
build_info, indices = self._get_mip_build_info()
## TODO: note: if actiongrid is empty, build_info, indices == {}. Correct handling?
# initialize mip
mip = Cplex()
mip.set_problem_type(mip.problem_type.MILP)
vars = mip.variables
cons = mip.linear_constraints
n_actionable = len(build_info)
n_indicators = len(indices['action_ind_names'])
# define a[j]
vars.add(names = indices['action_var_names'],
types = ['C'] * n_actionable,
lb = indices['action_lb'],
ub = indices['action_ub'])
# sum_j w[j] a[j] > -score
cons.add(names = ['score'],
lin_expr = [SparsePair(ind = indices['action_var_names'], val = indices['coefficients'])],
senses = ['G'],
rhs = [-self.score()])
# define indicators u[j][k] = 1 if a[j] = actions[j][k]
vars.add(names = indices['action_ind_names'], types = ['B'] * n_indicators)
# restrict a[j] to feasible values using a 1 of K constraint setup
for info in build_info.values():
# restrict a[j] to actions in feasible set and make sure exactly 1 indicator u[j][k] is on
# 1. a[j] = sum_k u[j][k] * actions[j][k] - > 0.0 = sum u[j][k] * actions[j][k] - a[j]
# 2.sum_k u[j][k] = 1.0
cons.add(names = ['set_a[%d]' % info['idx'], 'pick_a[%d]' % info['idx']],
lin_expr = [SparsePair(ind = info['action_var_name'] + info['action_ind_names'], val = [-1.0] + info['actions']),
SparsePair(ind = info['action_ind_names'], val = [1.0] * len(info['actions']))],
senses = ["E", "E"],
rhs = [0.0, 1.0])
# declare indicator variables as SOS set
mip.SOS.add(type = "1", name = "sos_u[%d]" % info['idx'], SOS = SparsePair(ind = info['action_ind_names'], val = info['actions']))
# limit number of features per action
#
# size := n_actionable - n_null where n_null := sum_j u[j][0] = sum_j 1[a[j] = 0]
#
# size <= max_size
# n_actionable - sum_j u[j][0] <= max_size
# n_actionable - max_size <= sum_j u[j][0]
#
# min_size <= size:
# min_size <= n_actionable - sum_j u[j][0]
# sum_j u[j][0] <= n_actionable - min_size
size_expr = SparsePair(ind = indices['action_off_names'], val = [1.0] * n_actionable)
cons.add(names = ['max_items', 'min_items'],
lin_expr = [size_expr, size_expr],
senses = ['G', 'L'],
rhs = [float(n_actionable - max_items), float(n_actionable - min_items)])
# add constraints for cost function
if cost_type in ('total', 'local'):
indices.pop('cost_var_names')
objval_pairs = list(chain(*[list(zip(v['action_ind_names'], v['costs'])) for v in build_info.values()]))
mip.objective.set_linear(objval_pairs)
elif cost_type == 'max':
indices['max_cost_var_name'] = ['max_cost']
## handle empty actionsets
indices['epsilon'] = np.min(indices['cost_df'] or np.inf) / np.sum(indices['cost_ub'])
vars.add(names = indices['max_cost_var_name'] + indices['cost_var_names'],
types = ['C'] * (n_actionable + 1),
obj = [1.0] + [indices['epsilon']] * n_actionable)
#lb = [0.0] * (n_actionable + 1)) # default values are 0.0
cost_constraints = {
'names': [],
'lin_expr': [],
'senses': ["E", "G"] * n_actionable,
'rhs': [0.0, 0.0] * n_actionable,
}
for info in build_info.values():
cost_constraints['names'].extend([
'def_cost[%d]' % info['idx'],
'set_max_cost[%d]' % info['idx']
])
cost_constraints['lin_expr'].extend([
SparsePair(ind = info['cost_var_name'] + info['action_ind_names'], val = [-1.0] + info['costs']),
SparsePair(ind = indices['max_cost_var_name'] + info['cost_var_name'], val = [1.0, -1.0])
])
cons.add(**cost_constraints)
# old code (commented out for speed)
#
# vars.add(names = indices['cost_var_names'],
# types = ['C'] * n_actionable,
# obj = [indices['epsilon']] * n_actionable,
# #ub = [CPX_INFINITY] * n_actionable, #indices['cost_ub'], #indices['cost_ub'],
# lb = [0.0] * n_actionable)
#
# vars.add(names = indices['max_cost_var_name'],
# types = ['C'],
# obj = [1.0],
# #ub = [np.max(indices['cost_ub'])],
# lb = [0.0])
#
# for info in build_info.values():
# cost[j] = sum c[j][k] u[j][k]
# cons.add(names = ['def_cost[%d]' % info['idx']],
# lin_expr = [SparsePair(ind = info['cost_var_name'] + info['action_ind_names'], val = [-1.0] + info['costs'])]
# senses = ["E"],
# rhs = [0.0])
#
# max_cost > cost[j]
# cons.add(names = ['set_max_cost[%d]' % info['idx']],
# lin_expr = [SparsePair(ind = indices['max_cost_var_name'] + info['cost_var_name'], val = [1.0, -1.0])],
# senses = ["G"],
# rhs = [0.0])
mip = set_cpx_parameters(mip, self._cpx_parameters)
self._mip = mip
self._mip_indices = indices