BARON:

Branch and Reduce Optimization

Navigator
A High Level Overview for CME334
Assembled by Thomas Lipp
 What is BARON?
• BARON is a branch and bound non-linear, mixed
integer global optimization solver dating back to 1991,
although last updated in 2010
• Developed at University of Illionis at Urbana-
Champaign by Nikolaos Sahinidis.
• Problems can be stated in either the AIMMS or GAMS
modeling languages, although computing time is
available on the NEOS server for GAMS
• Objective and constraint functions must be factorable
• All non-linear variables and expressions must be
bounded from above and below
Branch and Bound Algorithms

A quick reminder of Branch and

Bound algorithms.

From:
Sahinidis, N. V. and M. Tawarmalani, BARON 9.0.4: Global
Optimization of Mixed-Integer Nonlinear Programs, User's
manual, 2010
What does the Baron Algorithm Look
like?

From:
Sahinidis, N. V. and M. Tawarmalani, BARON 9.0.4: Global
Optimization of Mixed-Integer Nonlinear Programs, User's
manual, 2010
 What does each step do?
• Unfortunately, the Baron documentation does
not address each step fully. BARON is written to
be modular so that user subroutines can be used
to replace certain steps as needed.
• The creators are also trying to market BARON:
http://www.theoptimizationfirm.com/

• The subproblems are generally fed to existing

solvers: SNOPT, MINOS, CPLEX, etc.
 Things I won’t be able to answer
• How does BARON handle mixed integers
• What exactly occurs in any given step
• There may be more information available in:
Available for $221.09, and not currently owned
by the Stanford Library.
What does BARON have to offer, then?
• BARON calls itself a branch and reduce
algorithm rather than a branch and bound
algorithm, and it is this “reducing” that will
provide the most insight.
• The modular nature of BARON allows it to be
tailored to your specific application.
Preprocessing on a Node

From:
Sahinidis, N. V. and M. Tawarmalani, BARON 9.0.4: Global
Optimization of Mixed-Integer Nonlinear Programs, User's
manual, 2010
Feasibility Based Range Reduction:
a “Poor Man’s LP”
• The solid outer box represents the
initial domain of the node.

• The solid inner box represents the

reduced domain by considering
each of the constraints individually.
“The Poor Man’s LP”

• The dotted boundary represents

solving for the limits on each
coordinate with all constraints active

• The shaded region is the true

feasible region.

From:
Sahinidis, N. V. and M. Tawarmalani, BARON 9.0.4: Global
Optimization of Mixed-Integer Nonlinear Programs, User's
manual, 2010
Lower Bounding

From:
Sahinidis, N. V. and M. Tawarmalani, BARON 9.0.4: Global
Optimization of Mixed-Integer Nonlinear Programs, User's
manual, 2010
 Lower Bounding
This is an overview of the lower bounding scheme used by BARON, but is a little
complex to go over. Remember that one requirement of all of our functions was
that they be factorable:

From:
M. Tawarmalani and N. V. Sahinidis
(2004), ‘Global optimization of mixed-
integer
nonlinear programs: A theoretical and
computational study’, Math. Program.,
DOI 10.1007/s10107-003-0467-6

Note: there are other bounding

schemes discussed, but this appears to
be the one that is used
Lower Bounding Subroutines
Lower Bounding Bilinearities
xi x j  0
( xiU  xi )( xUj  x j )  0
xiU xUj  xiU x j  xiU x j  xi x j  0
xi x j  xiU x j  xiU x j  xiU xUj
xi x j  xiL x j  xiL x j  xiL x Lj
So create a new variable to replace the bilinearity
wij  0
wij  xiU x j  xiU x j  xiU xUj
wij  xiL x j  xiL x j  xiL x Lj
A similar procedure can be performed in the case xi x j  0
and for upper bounding.
 Lower Bounding Linear Fractional
Terms
• The same inequalities used in the bilinear case can be
rearranged, with some added technicalities on sign to
produce the liner fractional inequalities:
 Further Lower Bounding
• Once a lower bound is created, a linear
approximation of this bound will often be
created to further speed computation.

From:
M. Tawarmalani and N. V. Sahinidis
(2004), ‘Global optimization of mixed-
integer
nonlinear programs: A theoretical and
computational study’, Math. Program.,
DOI 10.1007/s10107-003-0467-6
Upper Bounding

From:
Sahinidis, N. V. and M. Tawarmalani, BARON 9.0.4: Global
Optimization of Mixed-Integer Nonlinear Programs, User's
manual, 2010
Post Processing on a Node

From:
Sahinidis, N. V. and M. Tawarmalani, BARON 9.0.4: Global
Optimization of Mixed-Integer Nonlinear Programs, User's
manual, 2010
Optimality Based Range Reduction
• We will first define three different problems:
Original problem

Relaxed problem

Perturbed problem
 Optimality Based Range Reduction:
on an active constraint
In this example the upper bound range constraint is active on x

From:
Sahinidis, N. V. and M. Tawarmalani, BARON
9.0.4: Global Optimization of Mixed-Integer
Nonlinear Programs, User's manual, 2010

L: the solution of the relaxed problem R

U: an upper bound on the original problem P
Φ: Φ(y) unknown solution of R(y) but is L when y = 0, as R(0) = R.
Κj*: the range reduction if Φ(y)
Κj: the range reduction based on the supporting hyperplane of Φ(y), z
Z: a supporting hyperplane based on the Lagrange multiplier in the solution of R.
 Optimality Based Range Reduction:
on an inactive constraint
In this example there is no range constraint active.

From:
Sahinidis, N. V. and M. Tawarmalani, BARON
9.0.4: Global Optimization of Mixed-Integer
Nonlinear Programs, User's manual, 2010

L: the solution of the relaxed problem R

U: an upper bound on the original problem P
zL: linear underestimator found by fixing the value of x to be xjL
zU: linear underestimator found by fixing the value of x to xjU
Κj*/πj*: the range reduction if the entire value function were known.
Κj/ πj: the range reduction based on the underestimators z.
Branching

From:
Sahinidis, N. V. and M. Tawarmalani, BARON 9.0.4: Global
Optimization of Mixed-Integer Nonlinear Programs, User's
manual, 2010
 Branching: Variable Selection
• BARON uses a rectangular subdivision
scheme, so only a single variable is chosen for
each branching.
• The variable chosen to branch is the variable
that contributes the most to the “relaxation
gap”.
• This is a non trivial process as many additional
variables were introduced in our
underestimating step
 Branching: Point Selection
• Periodically, branch at the midpoint
• Otherwise, branch at the solution of the lower
bounding problem.
• This decision is stored, but not executed until
the node is selected again.
Node Selection

From:
Sahinidis, N. V. and M. Tawarmalani, BARON 9.0.4: Global
Optimization of Mixed-Integer Nonlinear Programs, User's
manual, 2010
 Node Selection
• Default BARON implementation is a composite
value based on lower bound, violation (sum of
violations of all variables) and order of
creation.
• If memory limits are approached BARON
switches to FIFO.
• As with all steps, customizable by the user.
 For More Information
• Sahinidis, N. V. and M. Tawarmalani, BARON 9.0.4: Global Optimization of
Mixed-Integer Nonlinear Programs, User's manual, 2010

• M. Tawarmalani and N. V. Sahinidis , ‘Global optimization of mixed-integer

nonlinear programs: A theoretical and computational study, Math Program,
DOI 10.1007/s10107-003-0467-6, 2004

• M. Tawarmalani and N. V. Sahinidis, Convexification and global optimization in

continuous and mixed-integer nonlinear programming: theory, algorithms,
software, and applications, Springer, 2002