line_search_strong_wolfe.py 12.7 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
54
55
56
57
    Attributes
    ----------
    c1 : float
        Parameter for Armijo condition rule.
    c2 : float
        Parameter for curvature condition rule.
58
    max_step_size : float
59
        Maximum step allowed in to be made in the descent direction.
60
61
62
63
    max_iterations : integer
        Maximum number of iterations performed by the line search algorithm.
    max_zoom_iterations : integer
        Maximum number of iterations performed by the zoom algorithm.
64
65
66
    """

    def __init__(self, c1=1e-4, c2=0.9,
Martin Reinecke's avatar
Martin Reinecke committed
67
                 max_step_size=1000000000, max_iterations=100,
68
                 max_zoom_iterations=100):
69
70
71
72
73
74
75
76
77

        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)

78
    def perform_line_search(self, energy, pk, f_k_minus_1=None):
79
        """Performs the first stage of the algorithm.
80
81

        It starts with a trial step size and it keeps increasing it until it
82
83
        satisfies the strong Wolf conditions. It also performs the descent and
        returns the optimal step length and the new energy.
84

85
86
87
88
89
90
        Parameters
        ----------
        energy : Energy object
            Energy object from which we will calculate the energy and the
            gradient at a specific point.
        pk : Field
91
            Vector pointing into the search direction.
92
        f_k_minus_1 : float
93
            Value of the fuction (which is being minimized) at the k-1
94
            iteration of the line search procedure. (Default: None)
95

96
97
98
99
        Returns
        -------
        energy_star : Energy object
            The new Energy object on the new position.
100
        """
101
        le_0 = LineEnergy(0., energy, pk, 0.)
102
103

        # initialize the zero phis
104
        old_phi_0 = f_k_minus_1
Martin Reinecke's avatar
Martin Reinecke committed
105
        phi_0 = le_0.value
106
        phiprime_0 = le_0.directional_derivative
Theo Steininger's avatar
Theo Steininger committed
107
        if phiprime_0 >= 0:
108
            dobj.mprint("Error: search direction is not a descent direction")
Martin Reinecke's avatar
Martin Reinecke committed
109
            raise RuntimeError("search direction must be a descent direction")
110
111
112

        # set alphas
        alpha0 = 0.
113
114
115
        phi_alpha0 = phi_0
        phiprime_alpha0 = phiprime_0

116
117
        if self.preferred_initial_step_size is not None:
            alpha1 = self.preferred_initial_step_size
118
        elif old_phi_0 is not None:
119
120
121
122
            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
123
            alpha1 = 1.0/pk.norm()
124
125

        # start the minimization loop
Theo Steininger's avatar
Theo Steininger committed
126
127
128
        iteration_number = 0
        while iteration_number < self.max_iterations:
            iteration_number += 1
129
            if alpha1 == 0:
Theo Steininger's avatar
Theo Steininger committed
130
131
                result_energy = le_0.energy
                break
132
133
134

            le_alpha1 = le_0.at(alpha1)
            phi_alpha1 = le_alpha1.value
135

Martin Reinecke's avatar
Martin Reinecke committed
136
            if (phi_alpha1 > phi_0 + self.c1*alpha1*phiprime_0) or \
137
               ((phi_alpha1 >= phi_alpha0) and (iteration_number > 1)):
138
139
140
                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
141
142
                result_energy = le_star.energy
                break
143

144
            phiprime_alpha1 = le_alpha1.directional_derivative
Martin Reinecke's avatar
Martin Reinecke committed
145
            if abs(phiprime_alpha1) <= -self.c2*phiprime_0:
Theo Steininger's avatar
Theo Steininger committed
146
147
                result_energy = le_alpha1.energy
                break
148
149

            if phiprime_alpha1 >= 0:
150
151
152
                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
153
154
                result_energy = le_star.energy
                break
155
156

            # update alphas
157
158
159
160
            alpha0, alpha1 = alpha1, min(2*alpha1, self.max_step_size)
            if alpha1 == self.max_step_size:
                return le_alpha1.energy

161
162
163
            phi_alpha0 = phi_alpha1
            phiprime_alpha0 = phiprime_alpha1
        else:
Martin Reinecke's avatar
Martin Reinecke committed
164
            dobj.mprint("max iterations reached")
165
            return le_alpha1.energy
Theo Steininger's avatar
Theo Steininger committed
166
        return result_energy
167
168

    def _zoom(self, alpha_lo, alpha_hi, phi_0, phiprime_0,
169
              phi_lo, phiprime_lo, phi_hi, le_0):
170
        """Performs the second stage of the line search algorithm.
171
172
173

        If the first stage was not successful then the Zoom algorithm tries to
        find a suitable step length by using bisection, quadratic, cubic
174
        interpolation.
175

176
177
178
        Parameters
        ----------
        alpha_lo : float
Martin Reinecke's avatar
Martin Reinecke committed
179
180
            A boundary for the step length interval.
            Fulfills Wolfe condition 1.
Martin Reinecke's avatar
Martin Reinecke committed
181
        alpha_hi : float
Martin Reinecke's avatar
Martin Reinecke committed
182
            The other boundary for the step length interval.
183
        phi_0 : float
184
            Value of the energy at the starting point of the line search
185
            algorithm.
186
187
188
        phiprime_0 : float
            directional derivative at the starting point of the line search
            algorithm.
189
        phi_lo : float
190
            Value of the energy if we perform a step of length alpha_lo in
191
            descent direction.
192
193
194
        phiprime_lo : float
            directional derivative at the new position if we perform a step of
            length alpha_lo in descent direction.
195
        phi_hi : float
196
            Value of the energy if we perform a step of length alpha_hi in
197
            descent direction.
198

199
200
201
202
203
        Returns
        -------
        energy_star : Energy object
            The new Energy object on the new position.
        """
Martin Reinecke's avatar
Martin Reinecke committed
204
205
        cubic_delta = 0.2  # cubic interpolant checks
        quad_delta = 0.1  # quadratic interpolant checks
Theo Steininger's avatar
Theo Steininger committed
206
207
        alpha_recent = None
        phi_recent = None
208

Martin Reinecke's avatar
Martin Reinecke committed
209
        assert phi_lo <= phi_0 + self.c1*alpha_lo*phiprime_0
Theo Steininger's avatar
Theo Steininger committed
210
        assert phiprime_lo*(alpha_hi-alpha_lo) < 0.
Martin Reinecke's avatar
Martin Reinecke committed
211
        for i in range(self.max_zoom_iterations):
Theo Steininger's avatar
Theo Steininger committed
212
213
            # assert phi_lo <= phi_0 + self.c1*alpha_lo*phiprime_0
            # assert phiprime_lo*(alpha_hi-alpha_lo)<0.
214
            delta_alpha = alpha_hi - alpha_lo
215
            a, b = min(alpha_lo, alpha_hi), max(alpha_lo, alpha_hi)
216
217
218
219
220
221
222
223
224
225
226
227
228

            # 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)
229
                # If quadratic was not successful, try bisection
230
231
232
233
234
                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
235
            le_alphaj = le_0.at(alpha_j)
Martin Reinecke's avatar
Martin Reinecke committed
236
            phi_alphaj = le_alphaj.value
237

238
239
            # If the first Wolfe condition is not met replace alpha_hi
            # by alpha_j
240
            if (phi_alphaj > phi_0 + self.c1*alpha_j*phiprime_0) or \
241
242
243
244
               (phi_alphaj >= phi_lo):
                alpha_recent, phi_recent = alpha_hi, phi_hi
                alpha_hi, phi_hi = alpha_j, phi_alphaj
            else:
245
                phiprime_alphaj = le_alphaj.directional_derivative
246
                # If the second Wolfe condition is met, return the result
Martin Reinecke's avatar
Martin Reinecke committed
247
                if abs(phiprime_alphaj) <= -self.c2*phiprime_0:
248
                    return le_alphaj
249
250
251
252
253
254
255
256
257
258
259
                # 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
260
            dobj.mprint("The line search algorithm (zoom) did not converge.")
261
            return le_alphaj
262
263

    def _cubicmin(self, a, fa, fpa, b, fb, c, fc):
264
        """Estimating the minimum with cubic interpolation.
265

266
        Finds the minimizer for a cubic polynomial that goes through the
Martin Reinecke's avatar
Martin Reinecke committed
267
        points (a,a), (b,fb), and (c,fc) with derivative at point a of fpa.
268
        If no minimizer can be found return None
269

270
271
        Parameters
        ----------
Martin Reinecke's avatar
Martin Reinecke committed
272
273
274
275
276
277
        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
278

279
280
281
282
        Returns
        -------
        xmin : float
            Position of the approximated minimum.
283
284
285
286
287
288
        """
        with np.errstate(divide='raise', over='raise', invalid='raise'):
            try:
                C = fpa
                db = b - a
                dc = c - a
289
                denom = db * db * dc * dc * (db - dc)
290
                d1 = np.empty((2, 2))
291
292
293
294
                d1[0, 0] = dc * dc
                d1[0, 1] = -(db*db)
                d1[1, 0] = -(dc*dc*dc)
                d1[1, 1] = db*db*db
295
                [A, B] = np.dot(d1, np.asarray([fb - fa - C * db,
Martin Reinecke's avatar
Martin Reinecke committed
296
                                                fc - fa - C * dc]).ravel())
297
298
299
300
301
302
303
304
305
306
307
                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):
308
        """Estimating the minimum with quadratic interpolation.
309

310
        Finds the minimizer for a quadratic polynomial that goes through
Martin Reinecke's avatar
Martin Reinecke committed
311
        the points (a,fa), (b,fb) with derivative at point a of fpa.
312

313
314
        Parameters
        ----------
Martin Reinecke's avatar
Martin Reinecke committed
315
316
317
318
        a, fa, fpa : float
            abscissa, function value and derivative at first point
        b, fb : float
            abscissa and function value at second point
319

320
321
322
        Returns
        -------
        xmin : float
323
            Position of the approximated minimum.
324
325
326
327
        """
        with np.errstate(divide='raise', over='raise', invalid='raise'):
            try:
                db = b - a * 1.0
Martin Reinecke's avatar
Martin Reinecke committed
328
329
                B = (fb - fa - fpa * db) / (db * db)
                xmin = a - fpa / (2.0 * B)
330
331
332
333
334
            except ArithmeticError:
                return None
        if not np.isfinite(xmin):
            return None
        return xmin