Mark Schmidt (2005-2012)
minFunc is a Matlab function for unconstrained optimization of differentiable
real-valued multivariate functions using line-search methods. It uses an
interface very similar to the Matlab Optimization Toolbox function fminunc,
and can be called as a replacement for this function.
On many problems, minFunc requires fewer function evaluations to converge than
fminunc (or minimize.m). Further it
can optimize problems with a much larger number of variables (fminunc is
restricted to several thousand variables), and uses
a line search that is robust to several common function pathologies.
The default parameters of minFunc
call a quasi-Newton strategy, where limited-memory BFGS updates with Shanno-Phua scaling are used in
computing the step direction, and a bracketing line-search for a point
satisfying the strong Wolfe conditions is used to compute the step direction. In the
line search, (safeguarded) cubic
interpolation is used to generate trial values, and the method switches to an
Armijo back-tracking line search on iterations where the objective function
enters a region where the
parameters do not produce a real valued output (i.e. complex, NaN, or Inf).
of the non-default features present in minFunc:
- Step directions can be computed based on: Exact Newton (requires
user-supplied Hessian), full quasi-Newton approximation (uses a dense Hessian approximation),
limited-memory BFGS (uses a low-rank Hessian approximation - default),
(preconditioned) Hessian-free Newton (uses Hessian-vector
products), (preconditioned) conjugate gradient (uses only previous step and a vector beta),
Barzilai and Borwein (uses only previous step), or (cyclic) steepest descent.
- Step lengths can be computed based on either the (non-monotone) Armijo or Wolfe
conditions, and trial values can be generated by either backtracking/bisection,
or polynomial interpolation. Several strategies are available for selecting
the initial trial value.
- Numerical differentiation and derivative checking are available, including
an option for automatic differentiation using complex-step differentials (if the objective
function code handles complex inputs).
- Most methods have user-modifiable parameters, such as the number of
corrections to store for L-BFGS, modification options for Hessian matrices that
are not positive-definite in the pure Newton method, choice of
preconditioning and Hessian-vector product functions for
the Hessian-free Newton method, choice of
update method;scaling;preconditioning for the non-linear conjugate gradient method, the type of Hessian approximation to use in the quasi-Newton iteration, number
of steps to look back for the non-monotone Armijo condition, the parameters of the line search
algorithm, the parameters of the termination criteria, etc.
minFunc uses an interface very similar to Matlab's fminunc. If you currently
call 'fminunc(@myFunc,x0,options,myFuncArg1,myFuncArg2)', then you can use
minFunc instead by simply
replacing 'fminunc' with 'minFunc'. Note that by default minFunc assumes that
the gradient is supplied, unless the 'numDiff' option is set to 1 (for forward-differencing) or 2 (for central-differencing). minFunc
supports many of the same parameters as fminunc (but not all), but has some differences in naming and also has many
parameters that are not available for fminunc. 'help minFunc' will give a list
of parameters and their explanation.
'example_minFunc' gives an example of running the various limited-memory
solvers in minFunc with default options. It requires rosenbrock.m
(the 2D Rosenbrock "banana" function)
from the minimize.m
webpage (minimize.m is also included in the demo if it is found on the path).
To run the example, in Matlab type:
>> cd minFunc_2012
Running the example should produce the following output:
Result after 25 evaluations of limited-memory solvers on 2D rosenbrock:
x1 = 0.0000, x2 = 0.0000 (starting point)
x1 = 1.0000, x2 = 1.0000 (optimal solution)
x1 = 0.8725, x2 = 0.7569 (minimize.m by C. Rasmussen)
x1 = 0.3654, x2 = 0.1230 (minFunc with steepest descent)
x1 = 0.4974, x2 = 0.2452 (minFunc with cyclic steepest descent)
x1 = 0.8756, x2 = 0.7661 (minFunc with spectral gradient descent)
x1 = 0.5840, x2 = 0.3169 (minFunc with Hessian-free Newton)
x1 = 0.7478, x2 = 0.5559 (minFunc with preconditioned Hessian-free Newton)
x1 = 1.0010, x2 = 1.0020 (minFunc with conjugate gradient)
x1 = 0.7907, x2 = 0.6256 (minFunc with scaled conjugate gradient)
x1 = 0.9794, x2 = 0.9491 (minFunc with preconditioned conjugate gradient)
x1 = 1.0000, x2 = 1.0000 (minFunc with limited-memory BFGS - default)
A note on the above performance results: The default
parameters of minFunc were tuned based on log-linear statistical models (not on standard optimization problems like Rosenbrock's function).
To illustrate the use of Hessian-vector products and preconditioning, running example_minFunc_LR will show how to train a logistic regression classifier under various configurations of the Hessian-vector product and preconditioning function.
A very useful function that comes with the minFunc package is the derivativeCheck function that numerically checks the user-supplied gradient and/or Hessian, as well as the fastDerivativeCheck function which is more suitable for high-dimensional problems. Running example_derivativeCheck will show how to call these functions.
I have also prepared a more extensive set of examples of using minFunc on the webpage Examples of using minFunc.
The complete set of .m and .mex files for the current version of minFunc are available here
2012: A bunch of minor changes in response to feedback and some other changes that I've been wanting to make for a while. Most notably, the core L-BFGS code is now faster, and several nice enhancements to the derivative-checking functionality were added.
2009: Another set of minor changes in response to feedback (including mex files for some 64-bit architectures), but also some non-trivial changes to the conjugate gradient and dense quasi-Newton methods (this version is available here). Some highlights include:
- The L-BFGS code was completely re-written to be faster, to require less memory, and to be more stack-friendly.
- Excised the derivative checking code into a new function, derivativeCheck, that can be called directly to check the user-supplied gradient and/or Hessian. This function now also uses central-differencing by default, in order to return higher-accuracy results.
Also added the fastDerivativeCheck function, which uses a a random direction to quickly test the quality of the user-supplied gradient using only two or three function evaluations (thanks to Nicolas Le Roux for this suggestion). This is useful for high-dimensional problems where many function evaluations would be costly, but it has the disadvantage that it does not return an approximation of the gradient in order to help debugging the gradient code.
- The LS option was replaced by the LS_type and LS_interp options. The LS_type option specifies whether the Armijo or the Wolfe line-search will be used, and the LS_interp function specifies the type of interpolation/extrapolation strategy.
- Added the LS_multi option. If this is turned on then the Armijo line-search will use more than two points to build a higher-order interpolating function (quadratic, cubic, quartic, or quintic depending on the situation).
- Changed the names of the confusing-to-me tolX and tolFun variables to the more-simple-to-me optTol and progTol. The optTol variable specifies a first-order optimality tolerance, and the progTol variable specifies a lack-of-progress tolerance.
- The extrapolation procedure was modified to suppress the (harmless) condition-number warning message than many users encountered.
- Fixed a problem where the Wolfe line-search could prematurely detect a lack of progress.
- The user callback function can now tell minFunc to stop.
- The 64-bit windows mex files were added, thanks to Rainer Heintzmann for these files.
2008: I put an updated version of minFunc on this webpage (this version is available here).
Most of the changes made were minor and were made in response to user
feedback. Here are some highlights:
- Removed bounding of the initial step length in the line search (except on the first iteration). This sometimes caused slow convergence for functions with large gradient values.
- Changed the optimality tolerance parameters; tolFun is now only used to measure first-order optimality conditions, while tolX is now used for measures of lack of progress.
- Added the 'cyclic' steepest descent method, which performs an accurate line search, but then performs several iterations with the same step size under an inaccurate line search.
- To prevent the non-linear conjugate gradient method from restarting so often, this method was modified to accept the conjugate gradient step whenever a sufficient decrease condition is satisfied.
- Added the 'scaled' conjugate gradient method, where a Hessian-vector product is used to give a good initialization of the line search.
- Added the option to use the eigenvector associated with the most negative eigenvalue as a negative curvature direction, as an alternative to Hessian modification strategies.
- When the linear conjugate gradient method terminates because of negative
curvature in the newton0 method, the option is now available to use the negative curvature direction as the search direction (now the default).
- Added a preconditioned non-linear conjugate gradient method.
- Added the ability to call user-specified preconditioners.
- Changed the name of the 'bfgs' method to 'qnewton', as there are now a bunch of Hessian approximations available via the 'qnUpdate' option: BFGS updates, SR1 updates (falls back on BFGS if it doesn't maintain positive-definiteness), Hoshino updates, Self-Scaling BFGS (now the default), Oren's Self Scaling Variable Metric Method, and the McCormick-Huang asymmetric update.
- Removed the constrained optimization codes. For constrained optimization,
- The 64-bit linux mex file 'lbfgsC.mexa64' was added. Thanks to Emt Khan for this file.
- The 64-bit windows mex files 'lbfgsC.mexw64' and 'mcholC.mexw64' were added, as well as an updated version of 'mcholC.c' that allows this file to be compiled under 64-Windows machines. Thanks to Rainer Heintzmann for these files.
- The mex files 'lbfgsC.mexmaci', 'lbfgsC.mexmaci64', and 'mcholC.mexmaci64' were added. Thanks to Jascha Sohl-Dickstein for these files.
2007: Since a bunch of people have been using it, I put a version of minFunc online at this webpage (this first version is still available here).
- Input options are now case insensitive.
- The Wolfe line search now switches to an Armijo line search if the
function returns an illegal (complex, NaN, or Inf) value (previously, this was
one of the biggest causes of problems and you had to make the switch manually
for any function that you suspected might overflow or have other problems).
- The Wolfe line search is now more robust (thanks to Glenn Fung, Mirela
Andronescu, and Frank Hutter for helping me find some pathological cases).
- Added the ability to have user-specified Hessian-vector product functions
in the Hessian-Free Newton method.
- Added a mex version of the LDL^T factorization (often the fastest
Hessian modification strategy).
- Made finite-differencing the default numerical differentiation technique (this is
less accurate but more broadly applicable than the complex-step derivative approach).
- Added a Tensor method that uses a Tensor of 3rd derivatives to potentially
give improved convergence (this is mainly for entertainment purposes since I
can't imagine it being practically useful).
- Added three constrained optimization codes that use some of minFunc's
functionality to solve constrained optimization problems. 'minFuncBC' solves
problems with (upper/lower) bounds on the variables, 'minFuncEC' solves
problems with linear equality constraints,
and 'minFuncIC' solves problems with linear inequality constraints (this one
is not as efficient).
There are several situations that account for the majority of
the unsatisfactory behaviour of minFunc (with the first being the most common):
- The user-supplied gradient code is wrong. It is recommended to use
the derivativeCheck options (or function) to make sure that the gradient code roughly matches
numerically computed derivatives.
- minFunc runs out of memory. By default, minFunc uses a large number
of corrections in the L-BFGS
method. For problems with a very large number of variables, the 'Corr'
parameter should be decreased (i.e. if you run out of memory on the 10th
iteration, try setting 'Corr' to 9 or lower).
- minFunc returns before a solution of satisfactory accuracy is
In this case, decreasing optTol and/or progTol will typically fix the problem,
in some cases the objective function or variables need to be re-scaled.
- Each iteration requires many function evaluations.
In this case, changing the line search initialization (options.LS_init) from
its default value can sometimes improve performance (i.e. setting it to 2 will
use a quadratic approximation to initialize the line search).
- You get the error "??? Undefined function or method 'lbfgsAddC'". This occurs if
Matlab can't find the mex file for your operating system. If this occurs, you can either: (i) set 'options.useMex' to 0, or (ii) compile the mex file for your operating system by running the mexAll function. If you end up compiling a file in minFunc for your operating system, please send it to me and I will add to the distribution.
- The objective function is not appropriate (this is a modeling issue,
an optimization problem). You need to make sure that a minimum of the
actually the result you want, and that you don't need constraints.
- When using the pure Newton method, the default options in minFunc assume that the Hessian is positive-definite or close to
If this is not the case, the 'hessianModify' option may need to be
(to 1, 2, or 3, for example) in order to obtain good performance.
Constraints, L1-Regularization, Derivative-Free Methods, Trust Region, Coordinate Descent
Previous versions of minFunc included extra functions that solved certain types of constrained optimization problems (such as 'minFuncBC', that optimized with lower and/or upper bounds on the variables).
These functions are no longer included but several constrained optimization
methods are available in minConf.
Alternately, Michael Zibulevsky has written a penalty/barrier multiplier method
for large-scale constrained optimization that uses minFunc as a sub-routine,
available here (I have
not used this code). For non-smooth optimization where the only source of the
non-differentiability is due to the presence of an L1-regularization term, see
I have also written trust-region, derivative-free, and coordinate descent codes.
In some scenarios, the second-order trust-region code requires
evaluations than minFunc (based on line-searches), while some of the
derivative-free methods can be applied to problems that are continuous but
necessarily differentiable (i.e. if the function uses absolute values).
The coordinate descent method can use additional structure present in
the problem to (for very structured problems) give better performance.
methods are not available in the on-line package, contact me if you want these
Running on Graphics Cards
Rainer Heintzmann has a written package that allows some Matlab operations to be performed very efficiently on graphics cards. This package and a discussion of modifying minFunc to use this package are available here (I have not used this code).
Sometimes, we want to minimize a function of x that has the interface
instead of myFunc(x,arg1,arg2), such as when using Matlab's object oriented
programming functionality. To switch the argument order, you can use an
myObj = @(x)f(y,x,arg1,arg2);
To reference minFunc in a publication, please include my name and a link to this website. You may also want to include the date, since I may update the software in the future.
Mark Schmidt > Software > minFunc