Increasing solve speed
CVXGEN is designed to let you solve convex optimization problems particularly
fast. Here are some tips that may help you improve performance if solve time is
critical.
Reduce the size of your problem. Where possible, reduce the number of
variables, constraints or objective terms. With model predictive control
problems, for example, see the
paper by Wang and Boyd for
suggestions. Reducing the time horizon will provide an approximately linear
improvement in performance.
Use the most recent compiler for your platform, with appropriate compiler
optimizations. We recommend gcc4.4, with the Os option. This can make a
surprising difference. (A typical improvement with good optimization settings
is .) Using Os is appropriate, since it aims to reduce code size,
and CVXGEN problems often have relatively large code.
Relax constraint satisfaction and duality gap specifications. (See the settings
section below.) This is particularly important if average solve times are more
important than maximum solve times, as it lets CVXGEN return quickly if a good
solution is found. Often a nearoptimal point is found early, with subsequent
iterations merely confirming the point's quality.
Lower the fixed iteration limit. (Again, see the settings.) This is particularly
important if the maximum solve time is more important than the average solve
time. It may lead to a reducedquality (or even infeasible) solution, and
should be used with care.
Adjust the initial values for the slack and dual variables. This advanced
option is detailed in the settings, below.
Settings
The C code generated by CVXGEN allows several customizations. These can be made
by changing the settings structure defined in solver.h. When using the
Matlab interface, they may also be modified via the settings structure in
Matlab. Here is an explanation of the various available options.
Primary settings
settings.eps Default: . CVXGEN will not declare a problem
converged until the duality gap is known to be bounded by eps.
settings.resid_tol. Default: . CVXGEN will not declare a
problem converged until the norm of the equality and inequality residuals are
both less than resid_tol.
settings.max_iters. Default: 25. CVXGEN will exit early if eps and
resid_tol are satisfied. It will also exit when it has performed max_iters
iterations, regardless of the quality of the point it finds.
Most problems should require fewer than 20 iterations.
settings.verbose. Default: 1. When set to 0, CVXGEN will output
information about each iteration, including residual norms, duality gap bounds
and step sizes.
Additional settings related to numerical performance
These advanced settings are recommended only for experienced users. More details
will be provided in a later paper.
settings.kkt_reg. Default: . Rather than solve the exact KKT system
to find the search direction, CVXGEN regularizes the KKT matrix with a small
positive or negative term on each part of the diagonal. Increasing kkt_reg
yields more stable numerical characteristics, at the possible cost of more
iterations. Increase if some valid problem instances fail to converge;
decrease cautiously to potentially improve performance.
settings.refine_steps. Default: 1. To compensate for the KKT regularization,
and further improve numerical characteristics, CVXGEN performs refine_steps
of iterative refinement. Usually 1 step is sufficient. Increase for
potentially improved reliability and fewer outer iterations, at a 10–20%
cost per extra refine step; decrease to 0 for a risky solver that may not work, but
may be significantly faster if it does.
settings.verbose_refinement. Default: 1. Provides more information about the
iterative refinement. For debugging only.
