-
Notifications
You must be signed in to change notification settings - Fork 35
/
csaes.py
195 lines (175 loc) · 10.6 KB
/
csaes.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
import numpy as np # engine for numerical computing
from pypop7.optimizers.es.es import ES # abstract class of all evolution strategies (ES)
from pypop7.optimizers.es.dsaes import DSAES
class CSAES(DSAES):
"""Cumulative Step-size self-Adaptation Evolution Strategy (CSAES).
.. note:: `CSAES` adapts all the *individual* step-sizes on-the-fly with a *relatively small* population,
according to the well-known `CSA <http://link.springer.com/chapter/10.1007/3-540-58484-6_263>`_ rule
(a.k.a. cumulative (evolution) path-length control) from the Evolutionary Computation community. The
default setting (i.e., using a `small` population) can result in *relatively fast* (local) convergence,
but perhaps with the risk of getting trapped in suboptima on multi-modal fitness landscape. Therefore,
it is recommended to first attempt more advanced ES variants (e.g., `LMCMA`, `LMMAES`) for large-scale
black-box optimization. Here we include `CSAES` mainly for *benchmarking* and *theoretical* purpose.
Parameters
----------
problem : dict
problem arguments with the following common settings (`keys`):
* 'fitness_function' - objective function to be **minimized** (`func`),
* 'ndim_problem' - number of dimensionality (`int`),
* 'upper_boundary' - upper boundary of search range (`array_like`),
* 'lower_boundary' - lower boundary of search range (`array_like`).
options : dict
optimizer options with the following common settings (`keys`):
* 'max_function_evaluations' - maximum of function evaluations (`int`, default: `np.inf`),
* 'max_runtime' - maximal runtime to be allowed (`float`, default: `np.inf`),
* 'seed_rng' - seed for random number generation needed to be *explicitly* set (`int`);
and with the following particular settings (`keys`):
* 'sigma' - initial global step-size, aka mutation strength (`float`),
* 'mean' - initial (starting) point, aka mean of Gaussian search distribution (`array_like`),
* if not given, it will draw a random sample from the uniform distribution whose search range is
bounded by `problem['lower_boundary']` and `problem['upper_boundary']`.
* 'n_individuals' - number of offspring, aka offspring population size (`int`, default:
`4 + int(np.floor(3*np.log(problem['ndim_problem'])))`),
* 'n_parents' - number of parents, aka parental population size (`int`, default:
`int(options['n_individuals']/4)`),
* 'lr_sigma' - learning rate of global step-size adaptation (`float`, default:
`np.sqrt(options['n_parents']/(problem['ndim_problem'] + options['n_parents']))`).
Examples
--------
Use the black-box optimizer `CSAES` to minimize the well-known test function
`Rosenbrock <http://en.wikipedia.org/wiki/Rosenbrock_function>`_:
.. code-block:: python
:linenos:
>>> import numpy # engine for numerical computing
>>> from pypop7.benchmarks.base_functions import rosenbrock # function to be minimized
>>> from pypop7.optimizers.es.csaes import CSAES
>>> problem = {'fitness_function': rosenbrock, # to define problem arguments
... 'ndim_problem': 2,
... 'lower_boundary': -5.0*numpy.ones((2,)),
... 'upper_boundary': 5.0*numpy.ones((2,))}
>>> options = {'max_function_evaluations': 5000, # to set optimizer options
... 'seed_rng': 2022,
... 'mean': 3.0*numpy.ones((2,)),
... 'sigma': 0.1} # global step-size may need to be tuned
>>> csaes = CSAES(problem, options) # to initialize the optimizer class
>>> results = csaes.optimize() # to run the optimization/evolution process
>>> # to return the number of function evaluations and best-so-far fitness
>>> print(f"CSAES: {results['n_function_evaluations']}, {results['best_so_far_y']}")
CSAES: 5000, 0.010143683086819875
For its correctness checking of coding, refer to `this code-based repeatability report
<https://tinyurl.com/2s4ctvdw>`_ for more details.
Attributes
----------
best_so_far_x : `array_like`
final best-so-far solution found during entire optimization.
best_so_far_y : `array_like`
final best-so-far fitness found during entire optimization.
lr_sigma : `float`
learning rate of global step-size adaptation.
mean : `array_like`
initial (starting) point, aka mean of Gaussian search distribution.
n_individuals : `int`
number of offspring, aka offspring population size.
n_parents : `int`
number of parents, aka parental population size.
sigma : `float`
initial global step-size, aka mutation strength.
References
----------
Hansen, N., Arnold, D.V. and Auger, A., 2015.
`Evolution strategies.
<https://link.springer.com/chapter/10.1007%2F978-3-662-43505-2_44>`_
In Springer Handbook of Computational Intelligence (pp. 871-898). Springer, Berlin, Heidelberg.
Kern, S., Müller, S.D., Hansen, N., Büche, D., Ocenasek, J. and Koumoutsakos, P., 2004.
`Learning probability distributions in continuous evolutionary algorithms–a comparative review.
<https://link.springer.com/article/10.1023/B:NACO.0000023416.59689.4e>`_
Natural Computing, 3, pp.77-112.
Ostermeier, A., Gawelczyk, A. and Hansen, N., 1994, October.
`Step-size adaptation based on non-local use of selection information.
<http://link.springer.com/chapter/10.1007/3-540-58484-6_263>`_
In International Conference on Parallel Problem Solving from Nature (pp. 189-198).
Springer, Berlin, Heidelberg.
"""
def __init__(self, problem, options):
if options.get('n_individuals') is None: # number of offspring, aka offspring population size
options['n_individuals'] = 4 + int(np.floor(3*np.log(problem['ndim_problem'])))
if options.get('n_parents') is None: # number of parents, aka parental population size
options['n_parents'] = int(options['n_individuals']/4)
if options.get('lr_sigma') is None: # learning rate of global step-size adaptation
options['lr_sigma'] = np.sqrt(options['n_parents']/(
problem['ndim_problem'] + options['n_parents']))
DSAES.__init__(self, problem, options)
self._s_1 = None
self._s_2 = None
# set E[||N(0,I)||]: expectation of chi distribution
self._e_chi = np.sqrt(self.ndim_problem)*(
1.0 - 1.0/(4.0*self.ndim_problem) + 1.0/(21.0*np.power(self.ndim_problem, 2)))
def initialize(self, is_restart=False):
self._s_1 = 1.0 - self.lr_sigma
self._s_2 = np.sqrt(self.lr_sigma*(2.0 - self.lr_sigma)*self.n_parents)
self._axis_sigmas = self._sigma_bak*np.ones((self.ndim_problem,))
z = np.empty((self.n_individuals, self.ndim_problem)) # noise for offspring population
x = np.empty((self.n_individuals, self.ndim_problem)) # offspring population
mean = self._initialize_mean(is_restart) # mean of Gaussian search distribution
s = np.zeros((self.ndim_problem,)) # evolution path
y = np.empty((self.n_individuals,)) # fitness (no evaluation)
self._list_initial_mean.append(np.copy(mean))
return z, x, mean, s, y
def iterate(self, z=None, x=None, mean=None, y=None, args=None):
for k in range(self.n_individuals): # to sample offspring population
if self._check_terminations():
return z, x, y
z[k] = self.rng_optimization.standard_normal((self.ndim_problem,))
x[k] = mean + self._axis_sigmas*z[k]
y[k] = self._evaluate_fitness(x[k], args)
return z, x, y
def _update_distribution(self, z=None, x=None, s=None, y=None):
order = np.argsort(y)[:self.n_parents]
s = self._s_1*s + self._s_2*np.mean(z[order], axis=0)
sigmas_1 = np.power(np.exp(np.abs(s)/self._e_hnd - 1.0), 1.0/(3.0*self.ndim_problem))
sigmas_2 = np.power(np.exp(np.linalg.norm(s)/self._e_chi - 1.0),
self.lr_sigma/(1.0 + np.sqrt(self.n_parents/self.ndim_problem)))
self._axis_sigmas *= (sigmas_1*sigmas_2)
mean = np.mean(x[order], axis=0)
return s, mean
def restart_reinitialize(self, z=None, x=None, mean=None, s=None, y=None):
if not self.is_restart:
return z, x, mean, s, y
min_y = np.min(y)
if min_y < self._list_fitness[-1]:
self._list_fitness.append(min_y)
else:
self._list_fitness.append(self._list_fitness[-1])
is_restart_1, is_restart_2 = np.all(self._axis_sigmas < self.sigma_threshold), False
if len(self._list_fitness) >= self.stagnation:
is_restart_2 = (self._list_fitness[-self.stagnation] - self._list_fitness[-1]) < self.fitness_diff
is_restart = bool(is_restart_1) or bool(is_restart_2)
if is_restart:
self._print_verbose_info([], y, True)
if self.verbose:
print(' ....... *** restart *** .......')
self._n_restart += 1
self._list_generations.append(self._n_generations) # for each restart
self._n_generations = 0
self.n_individuals *= 2
self.n_parents = int(self.n_individuals/4)
self.lr_sigma = np.sqrt(self.n_parents/(self.ndim_problem + self.n_parents))
z, x, mean, s, y = self.initialize(True)
self._list_fitness = [np.inf]
return z, x, mean, s, y
def optimize(self, fitness_function=None, args=None): # for all generations (iterations)
fitness = ES.optimize(self, fitness_function)
z, x, mean, s, y = self.initialize()
while True:
# sample and evaluate offspring population
z, x, y = self.iterate(z, x, mean, y, args)
if self._check_terminations():
break
self._print_verbose_info(fitness, y)
self._n_generations += 1
s, mean = self._update_distribution(z, x, s, y)
z, x, mean, s, y = self.restart_reinitialize(z, x, mean, s, y)
results = self._collect(fitness, y, mean)
results['s'] = s
results['_axis_sigmas'] = self._axis_sigmas
return results