line_search_strong_wolfe.py 12.5 KB
Newer Older
1
2
3
4
5
6
7
8
9
10
11
12
# This program is free software: you can redistribute it and/or modify
# it under the terms of the GNU General Public License as published by
# the Free Software Foundation, either version 3 of the License, or
# (at your option) any later version.
#
# This program is distributed in the hope that it will be useful,
# but WITHOUT ANY WARRANTY; without even the implied warranty of
# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
# GNU General Public License for more details.
#
# You should have received a copy of the GNU General Public License
# along with this program.  If not, see <http://www.gnu.org/licenses/>.
Theo Steininger's avatar
Theo Steininger committed
13
14
15
16
17
#
# Copyright(C) 2013-2017 Max-Planck-Society
#
# NIFTy is being developed at the Max-Planck-Institut fuer Astrophysik
# and financially supported by the Studienstiftung des deutschen Volkes.
18

Martin Reinecke's avatar
Martin Reinecke committed
19
20
from __future__ import division
from builtins import range
21
22
import numpy as np
from .line_search import LineSearch
Martin Reinecke's avatar
Martin Reinecke committed
23
from .line_energy import LineEnergy
Martin Reinecke's avatar
Martin Reinecke committed
24
from .. import dobj
25
26
27


class LineSearchStrongWolfe(LineSearch):
28
    """Class for finding a step size that satisfies the strong Wolfe conditions.
29

30
    Algorithm contains two stages. It begins with a trial step length and
Martin Reinecke's avatar
Martin Reinecke committed
31
32
    keeps increasing it until it finds an acceptable step length or an
    interval. If it does not satisfy the Wolfe conditions, it performs the Zoom
33
34
35
    algorithm (second stage). By interpolating it decreases the size of the
    interval until an acceptable step length is found.

36
37
    Parameters
    ----------
38
    c1 : float
39
        Parameter for Armijo condition rule. (Default: 1e-4)
40
    c2 : float
41
        Parameter for curvature condition rule. (Default: 0.9)
42
    max_step_size : float
43
        Maximum step allowed in to be made in the descent direction.
44
45
46
47
48
        (Default: 50)
    max_iterations : integer
        Maximum number of iterations performed by the line search algorithm.
        (Default: 10)
    max_zoom_iterations : integer
49
        Maximum number of iterations performed by the zoom algorithm.
50
        (Default: 10)
51
52
53
    """

    def __init__(self, c1=1e-4, c2=0.9,
Martin Reinecke's avatar
Martin Reinecke committed
54
                 max_step_size=1000000000, max_iterations=100,
55
                 max_zoom_iterations=100):
56
57
58
59
60
61
62
63
64

        super(LineSearchStrongWolfe, self).__init__()

        self.c1 = np.float(c1)
        self.c2 = np.float(c2)
        self.max_step_size = max_step_size
        self.max_iterations = int(max_iterations)
        self.max_zoom_iterations = int(max_zoom_iterations)

65
    def perform_line_search(self, energy, pk, f_k_minus_1=None):
66
        """Performs the first stage of the algorithm.
67
68

        It starts with a trial step size and it keeps increasing it until it
69
70
        satisfies the strong Wolf conditions. It also performs the descent and
        returns the optimal step length and the new energy.
71

72
73
74
75
76
77
        Parameters
        ----------
        energy : Energy object
            Energy object from which we will calculate the energy and the
            gradient at a specific point.
        pk : Field
78
            Vector pointing into the search direction.
79
        f_k_minus_1 : float
80
            Value of the fuction (which is being minimized) at the k-1
81
            iteration of the line search procedure. (Default: None)
82

83
84
85
86
        Returns
        -------
        energy_star : Energy object
            The new Energy object on the new position.
87
        """
88
        le_0 = LineEnergy(0., energy, pk, 0.)
89
90

        # initialize the zero phis
91
        old_phi_0 = f_k_minus_1
Martin Reinecke's avatar
Martin Reinecke committed
92
        phi_0 = le_0.value
93
        phiprime_0 = le_0.directional_derivative
94
95
96
97
        if phiprime_0 == 0:
            dobj.mprint("Directional derivative is zero; assuming convergence")
            return energy
        if phiprime_0 > 0:
98
            dobj.mprint("Error: search direction is not a descent direction")
Martin Reinecke's avatar
bug fix    
Martin Reinecke committed
99
            raise ValueError("search direction must be a descent direction")
100
101
102

        # set alphas
        alpha0 = 0.
103
104
105
        phi_alpha0 = phi_0
        phiprime_alpha0 = phiprime_0

106
107
        if self.preferred_initial_step_size is not None:
            alpha1 = self.preferred_initial_step_size
108
        elif old_phi_0 is not None:
109
110
111
112
            alpha1 = min(1.0, 1.01*2*(phi_0 - old_phi_0)/phiprime_0)
            if alpha1 < 0:
                alpha1 = 1.0
        else:
Martin Reinecke's avatar
Martin Reinecke committed
113
            alpha1 = 1.0/pk.norm()
114
115

        # start the minimization loop
Theo Steininger's avatar
Theo Steininger committed
116
117
118
        iteration_number = 0
        while iteration_number < self.max_iterations:
            iteration_number += 1
119
            if alpha1 == 0:
Theo Steininger's avatar
Theo Steininger committed
120
121
                result_energy = le_0.energy
                break
122
123
124

            le_alpha1 = le_0.at(alpha1)
            phi_alpha1 = le_alpha1.value
125

Martin Reinecke's avatar
Martin Reinecke committed
126
            if (phi_alpha1 > phi_0 + self.c1*alpha1*phiprime_0) or \
127
               ((phi_alpha1 >= phi_alpha0) and (iteration_number > 1)):
128
129
130
                le_star = self._zoom(alpha0, alpha1, phi_0, phiprime_0,
                                     phi_alpha0, phiprime_alpha0, phi_alpha1,
                                     le_0)
Theo Steininger's avatar
Theo Steininger committed
131
132
                result_energy = le_star.energy
                break
133

134
            phiprime_alpha1 = le_alpha1.directional_derivative
Martin Reinecke's avatar
Martin Reinecke committed
135
            if abs(phiprime_alpha1) <= -self.c2*phiprime_0:
Theo Steininger's avatar
Theo Steininger committed
136
137
                result_energy = le_alpha1.energy
                break
138
139

            if phiprime_alpha1 >= 0:
140
141
142
                le_star = self._zoom(alpha1, alpha0, phi_0, phiprime_0,
                                     phi_alpha1, phiprime_alpha1, phi_alpha0,
                                     le_0)
Theo Steininger's avatar
Theo Steininger committed
143
144
                result_energy = le_star.energy
                break
145
146

            # update alphas
147
148
149
150
            alpha0, alpha1 = alpha1, min(2*alpha1, self.max_step_size)
            if alpha1 == self.max_step_size:
                return le_alpha1.energy

151
152
153
            phi_alpha0 = phi_alpha1
            phiprime_alpha0 = phiprime_alpha1
        else:
Martin Reinecke's avatar
Martin Reinecke committed
154
            dobj.mprint("max iterations reached")
155
            return le_alpha1.energy
Theo Steininger's avatar
Theo Steininger committed
156
        return result_energy
157
158

    def _zoom(self, alpha_lo, alpha_hi, phi_0, phiprime_0,
159
              phi_lo, phiprime_lo, phi_hi, le_0):
160
        """Performs the second stage of the line search algorithm.
161
162
163

        If the first stage was not successful then the Zoom algorithm tries to
        find a suitable step length by using bisection, quadratic, cubic
164
        interpolation.
165

166
167
168
        Parameters
        ----------
        alpha_lo : float
Martin Reinecke's avatar
Martin Reinecke committed
169
170
            A boundary for the step length interval.
            Fulfills Wolfe condition 1.
Martin Reinecke's avatar
Martin Reinecke committed
171
        alpha_hi : float
Martin Reinecke's avatar
Martin Reinecke committed
172
            The other boundary for the step length interval.
173
        phi_0 : float
174
            Value of the energy at the starting point of the line search
175
            algorithm.
176
177
178
        phiprime_0 : float
            directional derivative at the starting point of the line search
            algorithm.
179
        phi_lo : float
180
            Value of the energy if we perform a step of length alpha_lo in
181
            descent direction.
182
183
184
        phiprime_lo : float
            directional derivative at the new position if we perform a step of
            length alpha_lo in descent direction.
185
        phi_hi : float
186
            Value of the energy if we perform a step of length alpha_hi in
187
            descent direction.
188

189
190
191
192
193
        Returns
        -------
        energy_star : Energy object
            The new Energy object on the new position.
        """
Martin Reinecke's avatar
Martin Reinecke committed
194
195
        cubic_delta = 0.2  # cubic interpolant checks
        quad_delta = 0.1  # quadratic interpolant checks
Theo Steininger's avatar
Theo Steininger committed
196
197
        alpha_recent = None
        phi_recent = None
198

199
200
201
202
        if phi_lo > phi_0 + self.c1*alpha_lo*phiprime_0:
            raise ValueError("inconsistent data")
        if phiprime_lo*(alpha_hi-alpha_lo) >= 0.:
            raise ValueError("inconsistent data")
Martin Reinecke's avatar
Martin Reinecke committed
203
        for i in range(self.max_zoom_iterations):
Theo Steininger's avatar
Theo Steininger committed
204
205
            # assert phi_lo <= phi_0 + self.c1*alpha_lo*phiprime_0
            # assert phiprime_lo*(alpha_hi-alpha_lo)<0.
206
            delta_alpha = alpha_hi - alpha_lo
207
            a, b = min(alpha_lo, alpha_hi), max(alpha_lo, alpha_hi)
208
209
210
211
212
213
214
215
216
217
218
219
220

            # Try cubic interpolation
            if i > 0:
                cubic_check = cubic_delta * delta_alpha
                alpha_j = self._cubicmin(alpha_lo, phi_lo, phiprime_lo,
                                         alpha_hi, phi_hi,
                                         alpha_recent, phi_recent)
            # If cubic was not successful or not available, try quadratic
            if (i == 0) or (alpha_j is None) or (alpha_j > b - cubic_check) or\
               (alpha_j < a + cubic_check):
                quad_check = quad_delta * delta_alpha
                alpha_j = self._quadmin(alpha_lo, phi_lo, phiprime_lo,
                                        alpha_hi, phi_hi)
221
                # If quadratic was not successful, try bisection
222
223
224
225
226
                if (alpha_j is None) or (alpha_j > b - quad_check) or \
                   (alpha_j < a + quad_check):
                    alpha_j = alpha_lo + 0.5*delta_alpha

            # Check if the current value of alpha_j is already sufficient
227
            le_alphaj = le_0.at(alpha_j)
Martin Reinecke's avatar
Martin Reinecke committed
228
            phi_alphaj = le_alphaj.value
229

230
231
            # If the first Wolfe condition is not met replace alpha_hi
            # by alpha_j
232
            if (phi_alphaj > phi_0 + self.c1*alpha_j*phiprime_0) or \
233
234
235
236
               (phi_alphaj >= phi_lo):
                alpha_recent, phi_recent = alpha_hi, phi_hi
                alpha_hi, phi_hi = alpha_j, phi_alphaj
            else:
237
                phiprime_alphaj = le_alphaj.directional_derivative
238
                # If the second Wolfe condition is met, return the result
Martin Reinecke's avatar
Martin Reinecke committed
239
                if abs(phiprime_alphaj) <= -self.c2*phiprime_0:
240
                    return le_alphaj
241
242
243
244
245
246
247
248
249
250
251
                # If not, check the sign of the slope
                if phiprime_alphaj*delta_alpha >= 0:
                    alpha_recent, phi_recent = alpha_hi, phi_hi
                    alpha_hi, phi_hi = alpha_lo, phi_lo
                else:
                    alpha_recent, phi_recent = alpha_lo, phi_lo
                # Replace alpha_lo by alpha_j
                (alpha_lo, phi_lo, phiprime_lo) = (alpha_j, phi_alphaj,
                                                   phiprime_alphaj)

        else:
Martin Reinecke's avatar
Martin Reinecke committed
252
            dobj.mprint("The line search algorithm (zoom) did not converge.")
253
            return le_alphaj
254
255

    def _cubicmin(self, a, fa, fpa, b, fb, c, fc):
256
        """Estimating the minimum with cubic interpolation.
257

258
        Finds the minimizer for a cubic polynomial that goes through the
Martin Reinecke's avatar
Martin Reinecke committed
259
        points (a,a), (b,fb), and (c,fc) with derivative at point a of fpa.
260
        If no minimizer can be found return None
261

262
263
        Parameters
        ----------
Martin Reinecke's avatar
Martin Reinecke committed
264
265
266
267
268
269
        a, fa, fpa : float
            abscissa, function value and derivative at first point
        b, fb : float
            abscissa and function value at second point
        c, fc : float
            abscissa and function value at third point
270

271
272
273
274
        Returns
        -------
        xmin : float
            Position of the approximated minimum.
275
276
277
278
279
280
        """
        with np.errstate(divide='raise', over='raise', invalid='raise'):
            try:
                C = fpa
                db = b - a
                dc = c - a
281
                denom = db * db * dc * dc * (db - dc)
282
                d1 = np.empty((2, 2))
283
284
285
286
                d1[0, 0] = dc * dc
                d1[0, 1] = -(db*db)
                d1[1, 0] = -(dc*dc*dc)
                d1[1, 1] = db*db*db
287
                [A, B] = np.dot(d1, np.asarray([fb - fa - C * db,
Martin Reinecke's avatar
Martin Reinecke committed
288
                                                fc - fa - C * dc]).ravel())
289
290
291
292
293
294
295
296
297
298
299
                A /= denom
                B /= denom
                radical = B * B - 3 * A * C
                xmin = a + (-B + np.sqrt(radical)) / (3 * A)
            except ArithmeticError:
                return None
        if not np.isfinite(xmin):
            return None
        return xmin

    def _quadmin(self, a, fa, fpa, b, fb):
300
        """Estimating the minimum with quadratic interpolation.
301

302
        Finds the minimizer for a quadratic polynomial that goes through
Martin Reinecke's avatar
Martin Reinecke committed
303
        the points (a,fa), (b,fb) with derivative at point a of fpa.
304

305
306
        Parameters
        ----------
Martin Reinecke's avatar
Martin Reinecke committed
307
308
309
310
        a, fa, fpa : float
            abscissa, function value and derivative at first point
        b, fb : float
            abscissa and function value at second point
311

312
313
314
        Returns
        -------
        xmin : float
315
            Position of the approximated minimum.
316
317
318
319
        """
        with np.errstate(divide='raise', over='raise', invalid='raise'):
            try:
                db = b - a * 1.0
Martin Reinecke's avatar
Martin Reinecke committed
320
321
                B = (fb - fa - fpa * db) / (db * db)
                xmin = a - fpa / (2.0 * B)
322
323
324
325
326
            except ArithmeticError:
                return None
        if not np.isfinite(xmin):
            return None
        return xmin