scipy.optimize.shgo#

scipy.optimize.shgo(func, bounds, args=(), constraints=None, n=None, iters=1, callback=None, minimizer_kwargs=None, options=None, sampling_method='simplicial')[source]#

Finds the global minimum of a function using SHG optimization.

SHGO stands for “simplicial homology global optimization”.

Parameters:
funccallable

The objective function to be minimized. Must be in the form f(x, *args), where x is the argument in the form of a 1-D array and args is a tuple of any additional fixed parameters needed to completely specify the function.

boundssequence or Bounds

Bounds for variables. There are two ways to specify the bounds:

  1. Instance of Bounds class.

  2. Sequence of (min, max) pairs for each element in x.

argstuple, optional

Any additional fixed parameters needed to completely specify the objective function.

constraintsdict or sequence of dict, optional

Constraints definition. Function(s) R**n in the form:

g(x) >= 0 applied as g : R^n -> R^m
h(x) == 0 applied as h : R^n -> R^p

Each constraint is defined in a dictionary with fields:

typestr

Constraint type: ‘eq’ for equality, ‘ineq’ for inequality.

funcallable

The function defining the constraint.

jaccallable, optional

The Jacobian of fun (only for SLSQP).

argssequence, optional

Extra arguments to be passed to the function and Jacobian.

Equality constraint means that the constraint function result is to be zero whereas inequality means that it is to be non-negative. Note that COBYLA only supports inequality constraints.

Note

Only the COBYLA and SLSQP local minimize methods currently support constraint arguments. If the constraints sequence used in the local optimization problem is not defined in minimizer_kwargs and a constrained method is used then the global constraints will be used. (Defining a constraints sequence in minimizer_kwargs means that constraints will not be added so if equality constraints and so forth need to be added then the inequality functions in constraints need to be added to minimizer_kwargs too).

nint, optional

Number of sampling points used in the construction of the simplicial complex. Note that this argument is only used for sobol and other arbitrary sampling_methods. In case of sobol, it must be a power of 2: n=2**m, and the argument will automatically be converted to the next higher power of 2. Default is 100 for sampling_method='simplicial' and 128 for sampling_method='sobol'.

itersint, optional

Number of iterations used in the construction of the simplicial complex. Default is 1.

callbackcallable, optional

Called after each iteration, as callback(xk), where xk is the current parameter vector.

minimizer_kwargsdict, optional

Extra keyword arguments to be passed to the minimizer scipy.optimize.minimize Some important options could be:

  • methodstr

    The minimization method, the default is SLSQP.

  • argstuple

    Extra arguments passed to the objective function (func) and its derivatives (Jacobian, Hessian).

  • optionsdict, optional

    Note that by default the tolerance is specified as {ftol: 1e-12}

optionsdict, optional

A dictionary of solver options. Many of the options specified for the global routine are also passed to the scipy.optimize.minimize routine. The options that are also passed to the local routine are marked with “(L)”.

Stopping criteria, the algorithm will terminate if any of the specified criteria are met. However, the default algorithm does not require any to be specified:

  • maxfevint (L)

    Maximum number of function evaluations in the feasible domain. (Note only methods that support this option will terminate the routine at precisely exact specified value. Otherwise the criterion will only terminate during a global iteration)

  • f_min

    Specify the minimum objective function value, if it is known.

  • f_tolfloat

    Precision goal for the value of f in the stopping criterion. Note that the global routine will also terminate if a sampling point in the global routine is within this tolerance.

  • maxiterint

    Maximum number of iterations to perform.

  • maxevint

    Maximum number of sampling evaluations to perform (includes searching in infeasible points).

  • maxtimefloat

    Maximum processing runtime allowed

  • minhgrdint

    Minimum homology group rank differential. The homology group of the objective function is calculated (approximately) during every iteration. The rank of this group has a one-to-one correspondence with the number of locally convex subdomains in the objective function (after adequate sampling points each of these subdomains contain a unique global minimum). If the difference in the hgr is 0 between iterations for maxhgrd specified iterations the algorithm will terminate.

Objective function knowledge:

  • symmetrybool

    Specify True if the objective function contains symmetric variables. The search space (and therefore performance) is decreased by O(n!).

  • jacbool or callable, optional

    Jacobian (gradient) of objective function. Only for CG, BFGS, Newton-CG, L-BFGS-B, TNC, SLSQP, dogleg, trust-ncg. If jac is a boolean and is True, fun is assumed to return the gradient along with the objective function. If False, the gradient will be estimated numerically. jac can also be a callable returning the gradient of the objective. In this case, it must accept the same arguments as fun. (Passed to scipy.optimize.minmize automatically)

  • hess, hesspcallable, optional

    Hessian (matrix of second-order derivatives) of objective function or Hessian of objective function times an arbitrary vector p. Only for Newton-CG, dogleg, trust-ncg. Only one of hessp or hess needs to be given. If hess is provided, then hessp will be ignored. If neither hess nor hessp is provided, then the Hessian product will be approximated using finite differences on jac. hessp must compute the Hessian times an arbitrary vector. (Passed to scipy.optimize.minmize automatically)

Algorithm settings:

  • minimize_every_iterbool

    If True then promising global sampling points will be passed to a local minimization routine every iteration. If False then only the final minimizer pool will be run. Defaults to False.

  • local_iterint

    Only evaluate a few of the best minimizer pool candidates every iteration. If False all potential points are passed to the local minimization routine.

  • infty_constraints: bool

    If True then any sampling points generated which are outside will the feasible domain will be saved and given an objective function value of inf. If False then these points will be discarded. Using this functionality could lead to higher performance with respect to function evaluations before the global minimum is found, specifying False will use less memory at the cost of a slight decrease in performance. Defaults to True.

Feedback:

  • dispbool (L)

    Set to True to print convergence messages.

sampling_methodstr or function, optional

Current built in sampling method options are halton, sobol and simplicial. The default simplicial provides the theoretical guarantee of convergence to the global minimum in finite time. halton and sobol method are faster in terms of sampling point generation at the cost of the loss of guaranteed convergence. It is more appropriate for most “easier” problems where the convergence is relatively fast. User defined sampling functions must accept two arguments of n sampling points of dimension dim per call and output an array of sampling points with shape n x dim.

Returns:
resOptimizeResult

The optimization result represented as a OptimizeResult object. Important attributes are: x the solution array corresponding to the global minimum, fun the function output at the global solution, xl an ordered list of local minima solutions, funl the function output at the corresponding local solutions, success a Boolean flag indicating if the optimizer exited successfully, message which describes the cause of the termination, nfev the total number of objective function evaluations including the sampling calls, nlfev the total number of objective function evaluations culminating from all local search optimizations, nit number of iterations performed by the global routine.

Notes

Global optimization using simplicial homology global optimization [1]. Appropriate for solving general purpose NLP and blackbox optimization problems to global optimality (low-dimensional problems).

In general, the optimization problems are of the form:

minimize f(x) subject to

g_i(x) >= 0,  i = 1,...,m
h_j(x)  = 0,  j = 1,...,p

where x is a vector of one or more variables. f(x) is the objective function R^n -> R, g_i(x) are the inequality constraints, and h_j(x) are the equality constraints.

Optionally, the lower and upper bounds for each element in x can also be specified using the bounds argument.

While most of the theoretical advantages of SHGO are only proven for when f(x) is a Lipschitz smooth function, the algorithm is also proven to converge to the global optimum for the more general case where f(x) is non-continuous, non-convex and non-smooth, if the default sampling method is used [1].

The local search method may be specified using the minimizer_kwargs parameter which is passed on to scipy.optimize.minimize. By default, the SLSQP method is used. In general, it is recommended to use the SLSQP or COBYLA local minimization if inequality constraints are defined for the problem since the other methods do not use constraints.

The halton and sobol method points are generated using scipy.stats.qmc. Any other QMC method could be used.

References

[1] (1,2)

Endres, SC, Sandrock, C, Focke, WW (2018) “A simplicial homology algorithm for lipschitz optimisation”, Journal of Global Optimization.

[2]

Joe, SW and Kuo, FY (2008) “Constructing Sobol’ sequences with better two-dimensional projections”, SIAM J. Sci. Comput. 30, 2635-2654.

[3] (1,2)

Hoch, W and Schittkowski, K (1981) “Test examples for nonlinear programming codes”, Lecture Notes in Economics and Mathematical Systems, 187. Springer-Verlag, New York. http://www.ai7.uni-bayreuth.de/test_problem_coll.pdf

[4]

Wales, DJ (2015) “Perspective: Insight into reaction coordinates and dynamics from the potential energy landscape”, Journal of Chemical Physics, 142(13), 2015.

Examples

First consider the problem of minimizing the Rosenbrock function, rosen:

>>> import numpy as np
>>> from scipy.optimize import rosen, shgo
>>> bounds = [(0,2), (0, 2), (0, 2), (0, 2), (0, 2)]
>>> result = shgo(rosen, bounds)
>>> result.x, result.fun
(array([1., 1., 1., 1., 1.]), 2.920392374190081e-18)

Note that bounds determine the dimensionality of the objective function and is therefore a required input, however you can specify empty bounds using None or objects like np.inf which will be converted to large float numbers.

>>> bounds = [(None, None), ]*4
>>> result = shgo(rosen, bounds)
>>> result.x
array([0.99999851, 0.99999704, 0.99999411, 0.9999882 ])

Next, we consider the Eggholder function, a problem with several local minima and one global minimum. We will demonstrate the use of arguments and the capabilities of shgo. (https://en.wikipedia.org/wiki/Test_functions_for_optimization)

>>> def eggholder(x):
...     return (-(x[1] + 47.0)
...             * np.sin(np.sqrt(abs(x[0]/2.0 + (x[1] + 47.0))))
...             - x[0] * np.sin(np.sqrt(abs(x[0] - (x[1] + 47.0))))
...             )
...
>>> bounds = [(-512, 512), (-512, 512)]

shgo has built-in low discrepancy sampling sequences. First, we will input 64 initial sampling points of the Sobol’ sequence:

>>> result = shgo(eggholder, bounds, n=64, sampling_method='sobol')
>>> result.x, result.fun
(array([512.        , 404.23180824]), -959.6406627208397)

shgo also has a return for any other local minima that was found, these can be called using:

>>> result.xl
array([[ 512.        ,  404.23180824],
       [ 283.0759062 , -487.12565635],
       [-294.66820039, -462.01964031],
       [-105.87688911,  423.15323845],
       [-242.97926   ,  274.38030925],
       [-506.25823477,    6.3131022 ],
       [-408.71980731, -156.10116949],
       [ 150.23207937,  301.31376595],
       [  91.00920901, -391.283763  ],
       [ 202.89662724, -269.38043241],
       [ 361.66623976, -106.96493868],
       [-219.40612786, -244.06020508]])

>>> result.funl
array([-959.64066272, -718.16745962, -704.80659592, -565.99778097,
       -559.78685655, -557.36868733, -507.87385942, -493.9605115 ,
       -426.48799655, -421.15571437, -419.31194957, -410.98477763])

These results are useful in applications where there are many global minima and the values of other global minima are desired or where the local minima can provide insight into the system (for example morphologies in physical chemistry [4]).

If we want to find a larger number of local minima, we can increase the number of sampling points or the number of iterations. We’ll increase the number of sampling points to 64 and the number of iterations from the default of 1 to 3. Using simplicial this would have given us 64 x 3 = 192 initial sampling points.

>>> result_2 = shgo(eggholder, bounds, n=64, iters=3, sampling_method='sobol')
>>> len(result.xl), len(result_2.xl)
(12, 20)

Note the difference between, e.g., n=192, iters=1 and n=64, iters=3. In the first case the promising points contained in the minimiser pool are processed only once. In the latter case it is processed every 64 sampling points for a total of 3 times.

To demonstrate solving problems with non-linear constraints consider the following example from Hock and Schittkowski problem 73 (cattle-feed) [3]:

minimize: f = 24.55 * x_1 + 26.75 * x_2 + 39 * x_3 + 40.50 * x_4

subject to: 2.3 * x_1 + 5.6 * x_2 + 11.1 * x_3 + 1.3 * x_4 - 5     >= 0,

            12 * x_1 + 11.9 * x_2 + 41.8 * x_3 + 52.1 * x_4 - 21
                -1.645 * sqrt(0.28 * x_1**2 + 0.19 * x_2**2 +
                              20.5 * x_3**2 + 0.62 * x_4**2)       >= 0,

            x_1 + x_2 + x_3 + x_4 - 1                              == 0,

            1 >= x_i >= 0 for all i

The approximate answer given in [3] is:

f([0.6355216, -0.12e-11, 0.3127019, 0.05177655]) = 29.894378

>>> def f(x):  # (cattle-feed)
...     return 24.55*x[0] + 26.75*x[1] + 39*x[2] + 40.50*x[3]
...
>>> def g1(x):
...     return 2.3*x[0] + 5.6*x[1] + 11.1*x[2] + 1.3*x[3] - 5  # >=0
...
>>> def g2(x):
...     return (12*x[0] + 11.9*x[1] +41.8*x[2] + 52.1*x[3] - 21
...             - 1.645 * np.sqrt(0.28*x[0]**2 + 0.19*x[1]**2
...                             + 20.5*x[2]**2 + 0.62*x[3]**2)
...             ) # >=0
...
>>> def h1(x):
...     return x[0] + x[1] + x[2] + x[3] - 1  # == 0
...
>>> cons = ({'type': 'ineq', 'fun': g1},
...         {'type': 'ineq', 'fun': g2},
...         {'type': 'eq', 'fun': h1})
>>> bounds = [(0, 1.0),]*4
>>> res = shgo(f, bounds, iters=3, constraints=cons)
>>> res
 message: Optimization terminated successfully.
 success: True
     fun: 29.894378159142136
    funl: [ 2.989e+01]
       x: [ 6.355e-01  1.137e-13  3.127e-01  5.178e-02]
      xl: [[ 6.355e-01  1.137e-13  3.127e-01  5.178e-02]]
     nit: 3
    nfev: 114
   nlfev: 35
   nljev: 5
   nlhev: 0

>>> g1(res.x), g2(res.x), h1(res.x)
(-5.062616992290714e-14, -2.9594104944408173e-12, 0.0)