source: sasview/park-1.2.1/park/diffev.py @ 162c778

ESS_GUIESS_GUI_DocsESS_GUI_batch_fittingESS_GUI_bumps_abstractionESS_GUI_iss1116ESS_GUI_iss879ESS_GUI_iss959ESS_GUI_openclESS_GUI_orderingESS_GUI_sync_sascalccostrafo411magnetic_scattrelease-4.1.1release-4.1.2release-4.2.2release_4.0.1ticket-1009ticket-1094-headlessticket-1242-2d-resolutionticket-1243ticket-1249ticket885unittest-saveload
Last change on this file since 162c778 was 3570545, checked in by Mathieu Doucet <doucetm@…>, 13 years ago

Adding park Part 2

  • Property mode set to 100644
File size: 8.5 KB
RevLine 
[3570545]1from random import randint, random
2import numpy
3
4class DE:
5    """
6    Differential evolution test implementation
7    Implements the Scheme_DE_rand_1 variant
8
9       F: float
10           weighting factor which determines the magnitude of perturbation
11           occurring in each generation.
12
13       crossover_rate:  float
14           In general, 0 < crossover_rate < 1.  Usually
15           considerably lower than 1.  Convergence slows as the value
16           increases, but higher crossover rates may be necessary when
17
18       func: w = f(p)
19           The function to be minimized, of the form w = f(p), where p
20           is a vector of length func_dim and w is a scalar
21
22       func_dim: int
23           The dimension of the objective function
24
25       pop: array
26           The initial population.  This should be an iterable composed of
27           Rank-1 numpy arrays.  The population size must be at least 4,
28           preferably much greater.
29
30       l_bound, u_bound: vector
31           arrays of size func_dim representing the upper and lower bounds
32           for the parameters in the solution vectors
33
34       tol: float
35           if (max(pop_values) - min(pop_values) <= conv), the population
36           has converged and the evolution will stop
37       """
38
39    def __init__(self, F, crossover_rate, func, func_dim, pop, l_bound,
40                 u_bound, tol=1e-7):
41        self.pop_size          = len(pop)
42        self.dimension         = func_dim
43        self.F_orig            = F
44        self.F                 = F
45        self.crossover_rate    = crossover_rate
46        self.func              = func
47        self.tol               = tol
48        self.l_bound           = l_bound
49        self.u_bound           = u_bound
50        self.population        = pop
51        self.generations       = 0
52        self.pop_values        = [self.func(p) for p in self.population]
53        self.best_gene,self.best_value = self.get_best_gene()
54        self.best_gene_history = [self.best_gene]
55        self.best_val_history  = [self.best_value]
56        self.ncalls            = 0
57
58    #//////////////////////////////////////////////////
59    def evolve(self, numgens=1000, monitor=None):
60        '''Evolve the population for numgens generations, or until it converges.
61            Returns the best vector from the run'''
62        try:
63            import psyco
64            psyco.full()
65        except ImportError:
66            pass
67
68        start_gen = self.generations
69        for gen in xrange(self.generations + 1, self.generations + numgens + 1):
70            candidate_pop = []
71
72            for i in xrange(self.pop_size):
73                (r1, r2, r3) = self.select_participants(i)
74
75                #perturbation/mutation
76                candidate = self.population[r1] + self.F*(self.population[r2]
77                                                          - self.population[r3])
78
79                #crossover
80                candidate = self.crossover(candidate, i)
81
82                #check bound constraints
83                if not self.is_within_bounds(candidate):
84                    candidate = self.get_random_gene()
85
86                #test fitness to determine membership in next gen
87                candidate_val = self.func(candidate)
88                if candidate_val < self.pop_values[i]:
89                    candidate_pop.append(candidate)
90                    self.pop_values[i] = float(candidate_val)
91                else:
92                    candidate_pop.append(self.population[i])
93            self.ncalls += self.pop_size
94
95            self.population  = candidate_pop
96            x,fx = self.get_best_gene()
97            if fx < self.best_value:
98                self.best_gene, self.best_value = x,fx
99                if monitor is not None:
100                    monitor.improvement(x,fx,self.ncalls)
101            self.best_val_history.append(self.best_value)
102            self.best_gene_history.append(self.best_gene)
103            self.generations = gen
104
105            if monitor is not None:
106                monitor.progress(gen-start_gen,numgens)
107
108            if self.is_converged():
109                break
110
111        return self.best_gene
112
113
114    #////////////////////////////////////////////////
115    def get_random_gene(self):
116        '''Generate a random gene within the bounds constraints.
117           Used to replace out-of-bounds genes'''
118        result = [numpy.random.uniform(low, high)
119                  for low, high in zip(self.l_bound, self.u_bound)]
120        return numpy.array(result)
121
122    #////////////////////////////////////////////////
123    def is_within_bounds(self, gene):
124        '''Determine whether the gene meets the bounds constraints'''
125        return numpy.all( (self.l_bound < gene) & (self.u_bound > gene) )
126
127    #////////////////////////////////////////////////
128    #This appears to be a total failure.  I'll leave the code in,
129    #but it's not called from the main evolution loop anymore.
130    def adjust_F(self):
131        '''Adjust F to account for stagnation of the population '''
132        #has the best vector improved this generation?
133        idx = len(self.best_val_history)
134        if self.best_val_history[idx - 1] == self.best_val_history[idx - 2]:
135            self.stagnant_gens += 1
136        else:
137            self.stagnant_gens  = 0
138
139        #adapt F to account for stagnation
140        self.F = min( (self.F_orig + (self.stagnant_gens * 0.01)), 2.0)
141
142    #////////////////////////////////////////////////
143    def get_best_gene(self):
144        '''find the most fit gene in the current population'''
145        #print "pop", self.pop_values
146
147        best_index = numpy.argmin(self.pop_values)
148        return self.population[best_index],self.pop_values[best_index],
149
150    #////////////////////////////////////////////////
151    def select_participants(self, i):
152        '''generate r1, r2, and r3 randomly from the range [0, NP-1]
153            such that they are distinct values not equal to i'''
154        choices = [i]
155        for choice in xrange(3):
156            while True:
157                j = numpy.random.randint(0, self.pop_size-1)
158                if j not in choices:
159                    break
160            choices.append(j)
161        return choices[1:]
162    #////////////////////////////////////////////////
163    def crossover(self, candidate, i):
164        '''Perform a crossover between the candidate and the ith member of
165            the previous generation.'''
166        result = []
167
168        #generate lower bound of crossover (this bound is inclusive)
169        low = randint(0, self.dimension-1)
170
171        #determine the size of the crossover
172        L = 0
173        while random() < self.crossover_rate and L < (self.dimension - low):
174            L += 1
175
176        #calculate the upper bound of crossover (this bound is exclusive)
177        high = low + L
178
179        high  = numpy.repeat(high, self.dimension)
180        low   = numpy.repeat(low,  self.dimension)
181        seq   = numpy.arange(0, self.dimension)
182
183        result = numpy.choose( ((seq >= low)&(seq < high)),
184                             [candidate, self.population[i]] )
185        return result
186
187    def is_converged(self):
188        '''check for convergence'''
189        return max(self.pop_values) - min(self.pop_values) <= self.tol
190
191def diffev(fn, bounds, x0=None, pop_scale=4, crossover_rate=0.8,
192           Fscale=1, tolerance=1e-5, maxiter=1000, monitor=None):
193    """
194    Run differential evolution, returning x,fx,ncalls
195    """
196    lo, hi = numpy.asarray(bounds[0]), numpy.asarray(bounds[1])
197    dim = len(lo)
198    pop = gen_pop(dim*pop_scale, lo, hi, dim)
199    if x0 is not None:
200        pop[0] = numpy.asarray(x0)
201    evolver = DE(Fscale, crossover_rate, fn, dim, pop, lo, hi, tolerance)
202    evolver.evolve(maxiter, monitor=monitor)
203    return evolver.best_gene, evolver.best_value, evolver.ncalls
204
205
206
207
208########################################################################
209#end DE def
210########################################################################
211def gen_pop(size, l_bound, u_bound, dimension):
212    '''generate a random population of vectors within the given bounds.  dimension
213       indicates the length of the vectors.  l_bound and u_bound should be vectors
214       of length dimension (any iterable container should work)'''
215    result = []
216    for i in xrange(size):
217        entry = []
218        for j in xrange(dimension):
219            entry.append( ((u_bound[j] - l_bound[j]) * random()) + l_bound[j] )
220        result.append(numpy.array(entry))
221    return result
Note: See TracBrowser for help on using the repository browser.