Go to the first, previous, next, last section, table of contents.
This chapter describes routines for finding minima of arbitrary multidimensional functions. The library provides low level components for a variety of iterative minimizers and convergence tests. These can be combined by the user to achieve the desired solution, while providing full access to the intermediate steps of the algorithms. Each class of methods uses the same framework, so that you can switch between minimizers at runtime without needing to recompile your program. Each instance of a minimizer keeps track of its own state, allowing the minimizers to be used in multi-threaded programs. The minimization algorithms can be used to maximize a function by inverting its sign.
The header file `gsl_multimin.h' contains prototypes for the minimization functions and related declarations.
The problem of multidimensional minimization requires finding a point x such that the scalar function,
f(x_1, ..., x_n)
takes a value which is lower than at any neighboring point. For smooth functions the gradient g = \nabla f vanishes at the minimum. In general there are no bracketing methods available for the minimization of n-dimensional functions. The algorithms proceed from an initial guess using a search algorithm which attempts to move in a downhill direction.
Algorithms making use of the gradient of the function perform a one-dimensional line minimisation along this direction until the lowest point is found to a suitable tolerance. The search direction is then updated with local information from the function and its derivatives, and the whole process repeated until the true n-dimensional minimum is found.
Algorithms which do not require the gradient of the function use different strategies. For example, the Nelder-Mead Simplex algorithm maintains n+1 trial parameter vectors as the vertices of a n-dimensional simplex. On each iteration it tries to improve the worst vertex of the simplex by geometrical transformations. The iterations are continued until the overall size of the simplex has decreased sufficiently.
Both types of algorithms use a standard framework. The user provides a high-level driver for the algorithms, and the library provides the individual functions necessary for each of the steps. There are three main phases of the iteration. The steps are,
Each iteration step consists either of an improvement to the
line-minimisation in the current direction or an update to the search
direction itself. The state for the minimizers is held in a
gsl_multimin_fdfminimizer
struct or a
gsl_multimin_fminimizer
struct.
Note that the minimization algorithms can only search for one local minimum at a time. When there are several local minima in the search area, the first minimum to be found will be returned; however it is difficult to predict which of the minima this will be. In most cases, no error will be reported if you try to find a local minimum in an area where there is more than one.
It is also important to note that the minimization algorithms find local minima; there is no way to determine whether a minimum is a global minimum of the function in question.
The following function initializes a multidimensional minimizer. The minimizer itself depends only on the dimension of the problem and the algorithm and can be reused for different problems.
GSL_ENOMEM
.
gsl_multimin_fdfminimizer_set
initializes the minimizer s to minimize the function
fdf starting from the initial point x. The size of the
first trial step is given by step_size. The accuracy of the line
minimization is specified by tol. The precise meaning of this
parameter depends on the method used. Typically the line minimization
is considered successful if the gradient of the function g is
orthogonal to the current search direction p to a relative
accuracy of tol, where
dot(p,g) < tol |p| |g|. A tol value of 0.1 is
suitable for most purposes, since line minimization only needs to
be carried out approximately. Note that setting tol to zero will
force the use of "exact" line-searches, which are extremely expensive.
The function gsl_multimin_fminimizer_set
initializes the minimizer s to minimize the function
f, starting from the initial point
x. The size of the initial trial steps is given in vector
step_size. The precise meaning of this parameter depends on the
method used.
printf ("s is a '%s' minimizer\n", gsl_multimin_fdfminimizer_name (s));
would print something like s is a 'conjugate_pr' minimizer
.
You must provide a parametric function of n variables for the minimizers to operate on. You may also need to provide a routine which calculates the gradient of the function and a third routine which calculates both the function value and the gradient together. In order to allow for general parameters the functions are defined by the following data types:
double (* f) (const gsl_vector * x, void * params)
GSL_NAN
should be returned.
void (* df) (const gsl_vector * x, void * params, gsl_vector * g)
void (* fdf) (const gsl_vector * x, void * params, double * f, gsl_vector * g)
size_t n
void * params
double (* f) (const gsl_vector * x, void * params)
GSL_NAN
should be returned.
size_t n
void * params
The following example function defines a simple two-dimensional paraboloid with five parameters,
/* Paraboloid centered on (p[0],p[1]), with scale factors (p[2],p[3]) and minimum p[4] */ double my_f (const gsl_vector *v, void *params) { double x, y; double *p = (double *)params; x = gsl_vector_get(v, 0); y = gsl_vector_get(v, 1); return p[2] * (x - p[0]) * (x - p[0]) + p[3] * (y - p[1]) * (y - p[1]) + p[4]; } /* The gradient of f, df = (df/dx, df/dy). */ void my_df (const gsl_vector *v, void *params, gsl_vector *df) { double x, y; double *p = (double *)params; x = gsl_vector_get(v, 0); y = gsl_vector_get(v, 1); gsl_vector_set(df, 0, 2.0 * p[2] * (x - p[0])); gsl_vector_set(df, 1, 2.0 * p[3] * (y - p[1])); } /* Compute both f and df together. */ void my_fdf (const gsl_vector *x, void *params, double *f, gsl_vector *df) { *f = my_f(x, params); my_df(x, params, df); }
The function can be initialized using the following code,
gsl_multimin_function_fdf my_func; /* Paraboloid center at (1,2), scale factors (10, 20), minimum value 30 */ double p[5] = { 1.0, 2.0, 10.0, 20.0, 30.0 }; my_func.n = 2; /* number of function components */ my_func.f = &my_f; my_func.df = &my_df; my_func.fdf = &my_fdf; my_func.params = (void *)p;
The following function drives the iteration of each algorithm. The function performs one iteration to update the state of the minimizer. The same function works for all minimizers so that different methods can be substituted at runtime without modifications to the code.
GSL_ENOPROG
signifies that
the minimizer is unable to improve on its current estimate, either due
to numerical difficulty or because a genuine local minimum has been
reached.
The minimizer maintains a current best estimate of the minimum at all times. This information can be accessed with the following auxiliary functions,
A minimization procedure should stop when one of the following conditions is true:
The handling of these conditions is under user control. The functions below allow the user to test the precision of the current result.
GSL_SUCCESS
if the following condition is achieved,
|g| < epsabs
and returns GSL_CONTINUE
otherwise. A suitable choice of
epsabs can be made from the desired accuracy in the function for
small variations in x. The relationship between these quantities
is given by
\delta f = g \delta x.
GSL_SUCCESS
if the size is smaller than tolerance,
otherwise GSL_CONTINUE
is returned.
There are several minimization methods available. The best choice of algorithm depends on the problem. The algorithms described in this section use the value of the function and its gradient at each evaluation point.
An initial search direction p is chosen using the gradient, and line minimization is carried out in that direction. The accuracy of the line minimization is specified by the parameter tol. The minimum along this line occurs when the function gradient g and the search direction p are orthogonal. The line minimization terminates when dot(p,g) < tol |p| |g|. The search direction is updated using the Fletcher-Reeves formula p' = g' - \beta g where \beta=-|g'|^2/|g|^2, and the line minimization is then repeated for the new search direction.
The bfgs2
version of this minimizer is the most efficient
version available, and is a faithful implementation of the line
minimization scheme described in Fletcher's Practical Methods of
Optimization, Algorithms 2.6.2 and 2.6.4. It supersedes the original
bfgs
routine and requires substantially fewer function and
gradient evaluations. The user-supplied tolerance tol
corresponds to the parameter \sigma used by Fletcher. A value
of 0.1 is recommended for typical use (larger values correspond to
less accurate line searches).
The algorithms described in this section use only the value of the function at each evaluation point.
p_0 = (x_0, x_1, ... , x_n) p_1 = (x_0 + s_0, x_1, ... , x_n) p_2 = (x_0, x_1 + s_1, ... , x_n) ... = ... p_n = (x_0, x_1, ... , x_n + s_n)
These vectors form the n+1 vertices of a simplex in n dimensions. On each iteration the algorithm uses simple geometrical transformations to update the vector corresponding to the highest function value. The geometric transformations are reflection, reflection followed by expansion, contraction and multiple contraction. Using these transformations the simplex moves through the space towards the minimum, where it contracts itself.
After each iteration, the best vertex is returned. Note, that due to the nature of the algorithm not every step improves the current best parameter vector. Usually several iterations are required.
The minimizer-specific characteristic size is calculated as the
average distance from the geometrical center of the simplex to all its
vertices. This size can be used as a stopping criteria, as the
simplex contracts itself near the minimum. The size is returned by the
function gsl_multimin_fminimizer_size
.
The nmsimplex2
version of this minimiser is a new O(N) operations
implementation of the earlier O(N^2) operations nmsimplex
minimiser. It uses the same underlying algorithm, but the simplex
updates are computed more efficiently for high-dimensional problems.
In addition, the size of simplex is calculated as the RMS
distance of each vertex from the center rather than the mean distance,
allowing a linear update of this quantity on each step. The memory usage is
O(N^2) for both algorithms.
nmsimplex2
which initialises the
simplex around the starting point x using a randomly-oriented
set of basis vectors instead of the fixed coordinate axes. The
final dimensions of the simplex are scaled along the coordinate axes by the
vector step_size. The randomisation uses a simple deterministic
generator so that repeated calls to gsl_multimin_fminimizer_set
for
a given solver object will vary the orientation in a well-defined way.
This example program finds the minimum of the paraboloid function defined earlier. The location of the minimum is offset from the origin in x and y, and the function value at the minimum is non-zero. The main program is given below, it requires the example function given earlier in this chapter.
int main (void) { size_t iter = 0; int status; const gsl_multimin_fdfminimizer_type *T; gsl_multimin_fdfminimizer *s; /* Position of the minimum (1,2), scale factors 10,20, height 30. */ double par[5] = { 1.0, 2.0, 10.0, 20.0, 30.0 }; gsl_vector *x; gsl_multimin_function_fdf my_func; my_func.n = 2; my_func.f = my_f; my_func.df = my_df; my_func.fdf = my_fdf; my_func.params = par; /* Starting point, x = (5,7) */ x = gsl_vector_alloc (2); gsl_vector_set (x, 0, 5.0); gsl_vector_set (x, 1, 7.0); T = gsl_multimin_fdfminimizer_conjugate_fr; s = gsl_multimin_fdfminimizer_alloc (T, 2); gsl_multimin_fdfminimizer_set (s, &my_func, x, 0.01, 1e-4); do { iter++; status = gsl_multimin_fdfminimizer_iterate (s); if (status) break; status = gsl_multimin_test_gradient (s->gradient, 1e-3); if (status == GSL_SUCCESS) printf ("Minimum found at:\n"); printf ("%5d %.5f %.5f %10.5f\n", iter, gsl_vector_get (s->x, 0), gsl_vector_get (s->x, 1), s->f); } while (status == GSL_CONTINUE && iter < 100); gsl_multimin_fdfminimizer_free (s); gsl_vector_free (x); return 0; }
The initial step-size is chosen as 0.01, a conservative estimate in this case, and the line minimization parameter is set at 0.0001. The program terminates when the norm of the gradient has been reduced below 0.001. The output of the program is shown below,
x y f 1 4.99629 6.99072 687.84780 2 4.98886 6.97215 683.55456 3 4.97400 6.93501 675.01278 4 4.94429 6.86073 658.10798 5 4.88487 6.71217 625.01340 6 4.76602 6.41506 561.68440 7 4.52833 5.82083 446.46694 8 4.05295 4.63238 261.79422 9 3.10219 2.25548 75.49762 10 2.85185 1.62963 67.03704 11 2.19088 1.76182 45.31640 12 0.86892 2.02622 30.18555 Minimum found at: 13 1.00000 2.00000 30.00000
Note that the algorithm gradually increases the step size as it successfully moves downhill, as can be seen by plotting the successive points.
The conjugate gradient algorithm finds the minimum on its second direction because the function is purely quadratic. Additional iterations would be needed for a more complicated function.
Here is another example using the Nelder-Mead Simplex algorithm to minimize the same example object function, as above.
int main(void) { double par[5] = {1.0, 2.0, 10.0, 20.0, 30.0}; const gsl_multimin_fminimizer_type *T = gsl_multimin_fminimizer_nmsimplex2; gsl_multimin_fminimizer *s = NULL; gsl_vector *ss, *x; gsl_multimin_function minex_func; size_t iter = 0; int status; double size; /* Starting point */ x = gsl_vector_alloc (2); gsl_vector_set (x, 0, 5.0); gsl_vector_set (x, 1, 7.0); /* Set initial step sizes to 1 */ ss = gsl_vector_alloc (2); gsl_vector_set_all (ss, 1.0); /* Initialize method and iterate */ minex_func.n = 2; minex_func.f = my_f; minex_func.params = par; s = gsl_multimin_fminimizer_alloc (T, 2); gsl_multimin_fminimizer_set (s, &minex_func, x, ss); do { iter++; status = gsl_multimin_fminimizer_iterate(s); if (status) break; size = gsl_multimin_fminimizer_size (s); status = gsl_multimin_test_size (size, 1e-2); if (status == GSL_SUCCESS) { printf ("converged to minimum at\n"); } printf ("%5d %10.3e %10.3e f() = %7.3f size = %.3f\n", iter, gsl_vector_get (s->x, 0), gsl_vector_get (s->x, 1), s->fval, size); } while (status == GSL_CONTINUE && iter < 100); gsl_vector_free(x); gsl_vector_free(ss); gsl_multimin_fminimizer_free (s); return status; }
The minimum search stops when the Simplex size drops to 0.01. The output is shown below.
1 6.500e+00 5.000e+00 f() = 512.500 size = 1.130 2 5.250e+00 4.000e+00 f() = 290.625 size = 1.409 3 5.250e+00 4.000e+00 f() = 290.625 size = 1.409 4 5.500e+00 1.000e+00 f() = 252.500 size = 1.409 5 2.625e+00 3.500e+00 f() = 101.406 size = 1.847 6 2.625e+00 3.500e+00 f() = 101.406 size = 1.847 7 0.000e+00 3.000e+00 f() = 60.000 size = 1.847 8 2.094e+00 1.875e+00 f() = 42.275 size = 1.321 9 2.578e-01 1.906e+00 f() = 35.684 size = 1.069 10 5.879e-01 2.445e+00 f() = 35.664 size = 0.841 11 1.258e+00 2.025e+00 f() = 30.680 size = 0.476 12 1.258e+00 2.025e+00 f() = 30.680 size = 0.367 13 1.093e+00 1.849e+00 f() = 30.539 size = 0.300 14 8.830e-01 2.004e+00 f() = 30.137 size = 0.172 15 8.830e-01 2.004e+00 f() = 30.137 size = 0.126 16 9.582e-01 2.060e+00 f() = 30.090 size = 0.106 17 1.022e+00 2.004e+00 f() = 30.005 size = 0.063 18 1.022e+00 2.004e+00 f() = 30.005 size = 0.043 19 1.022e+00 2.004e+00 f() = 30.005 size = 0.043 20 1.022e+00 2.004e+00 f() = 30.005 size = 0.027 21 1.022e+00 2.004e+00 f() = 30.005 size = 0.022 22 9.920e-01 1.997e+00 f() = 30.001 size = 0.016 23 9.920e-01 1.997e+00 f() = 30.001 size = 0.013 converged to minimum at 24 9.920e-01 1.997e+00 f() = 30.001 size = 0.008
The simplex size first increases, while the simplex moves towards the minimum. After a while the size begins to decrease as the simplex contracts around the minimum.
The conjugate gradient and BFGS methods are described in detail in the following book,
A brief description of multidimensional minimization algorithms and more recent references can be found in,
The simplex algorithm is described in the following paper,
Go to the first, previous, next, last section, table of contents.