-
Notifications
You must be signed in to change notification settings - Fork 0
/
trainingStrategy.py
541 lines (461 loc) · 22.5 KB
/
trainingStrategy.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
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
#!/usr/bin/python
import random
import math
import patternSet
import Network
class TrainingStrategyType:
EvolutionStrategy, GeneticAlgorithm, DifferentialGA = range(3)
@classmethod
def desc(self, x):
if x == 3:
raise("Instance of an Abstract Class... Bad Juju!")
return {self.EvolutionStrategy: "EvolutionStrategy",
self.GeneticAlgorithm: "GeneticAlgorithm",
self.DifferentialGA: "DifferentialGA"}[x]
def recordItems(recordString):
with open('Strategy.log', 'a') as file:
file.write(recordString + '\n')
def avgSigma(population):
sigSum = 0.0
sigCount = 0
for member in population:
ms = Network.vectorizeMatrix(member.sigmas)
for singleSigma in ms:
sigSum = sigSum + singleSigma
sigCount = sigCount + 1
return sigSum/sigCount
def memberVarience(population, alpha):
diffSum = 0.0
for member in population:
mg = Network.vectorizeMatrix(member.genome)
ag = Network.vectorizeMatrix(alpha.genome)
diffSum = diffSum + Network.outputError(mg, ag)
return diffSum/len(population)
def averageFitness(population):
return sum(x.fitness for x in population)/len(population)
class Member():
memberIdInc = 0
genomeTemplate = [] # example [3, 3, 3, 3, 4, 4] Input layer has 3 nodes, Hidden has 4, output has 2
def __init__(self, geneMin, geneMax, includeStrategyParameters, strategyMax):
self.id = Member.memberIdInc
Member.memberIdInc = Member.memberIdInc + 1
self.genome = [[float(random.randrange(geneMin*10000, geneMax*10000))/10000 for _ in range(n)] for n in Member.genomeTemplate]
self.highPerformers = [0 for _ in Member.genomeTemplate]
self.sigmas = []
if includeStrategyParameters:
self.sigmas = [[float(random.randrange(strategyMax*10000))/10000-(strategyMax/2) for _ in range(n)] for n in Member.genomeTemplate]
self.fitness = 0.0
self.categoryCoverage = []
def adjustFitness(self, value):
self.fitness = self.fitness + value
def successFeedback(self, feedbackVector):
"""The Feedback Vector represents the categories from which this member was able to choose correctly, therefore we will give preference to the cooresponding genes during combination"""
for i, fbv in enumerate(feedbackVector):
self.highPerformers[-1*len(feedbackVector) + i] = fbv
# print("FV:" + str(feedbackVector))
# print("HP:" + str(self.highPerformers))
def getGenesAtPosition(self, neuronNumber):
if len(self.genome) > len(Member.genomeTemplate):
return {'genes':self.genome[neuronNumber], 'strategy parameters':self.genome[-1]}
return {'genes':self.genome[neuronNumber]}
# Used by the MLP during backprop
def setGenesAtPosition(self, neuronNumber, values):
if len(values) == len(self.genome[neuronNumber]):
self.genome[neuronNumber] = values
else:
raise("Number of Genes does not match value set length")
class TrainingStrategy(object):
def __init__(self):
Member.memberIdInc = 0
self.strategy = 4
self.trainingMode = Network.PatternType.Train
self.generation = 0
self.runningChildren = False
self.currentMember = 0
self.population = []
self.runningChildren = False
self.currentChildMember = 0
self.childPopulation = []
self.alphas = [] # a list of the best members from each generation current best is [-1]
self.highestCurrentMemberId = 0
self.childSuccess = 0.0
self.fitnessThreshold = .00005
self.maxPopulationSize = 0
self.lam = 1.0
self.useSigmas = False
self.sigmaMax = 1
self.maxGenerations = 10
@classmethod
def getTrainingStrategyOfType(self, type=3):
if type == TrainingStrategyType.EvolutionStrategy:
return EvolutionStrategy()
elif type == TrainingStrategyType.GeneticAlgorithm:
return GeneticAlgorithm()
elif type == TrainingStrategyType.DifferentialGA:
return DifferentialGA()
def initPopulation(self, pop, gRange):
self.maxPopulationSize = pop
self.population = []
for p in range(pop):
self.population.append(Member(gRange[0], gRange[-1], self.useSigmas, self.sigmaMax))
if self.lam <= 1:
self.lam = int(self.lam*self.maxPopulationSize)
print("Member Genome Sample:")
print("[" + ", ".join("[" + str(len(a)) + "]" for a in self.population[0].genome) + "]")
# print("[" + ", ".join("[" + " ".join(str(b) for b in a) + "]" for a in self.population[0].genome) + "]")
self.currentMember = 0
def avgSigma(self):
return avgSigma(self.population)
def memberVarience(self):
return memberVarience(self.population, self.alphas[0])
def averageFitness(self):
return averageFitness(self.population)
def epsilon(self):
"""epsilon"""
return 0.15
#---- Net Interface Methods ---------------------------------------------
def fitnessThresholdMet(self):
if self.generation > self.maxGenerations:
return True
if len(self.alphas) < 1:
return False
if self.alphas[0].fitness <= self.fitnessThreshold:
return True
return False
def moreMembers(self):
"""In order to calulate fitness on the chilren we'll do selection, crossover, and mutation
at the end of regular pop run, and then continue for the next set"""
if not self.runningChildren and self.currentMember < len(self.population):
return True
if not self.runningChildren:
# Once all current members have been evaluated, produce children mutate and calculate fitness for them
parents = self.select()
self.childPopulation = self.crossover(parents)
self.mutate()
self.runningChildren = True
if len(self.childPopulation) > 0:
return True
return False
if self.runningChildren and self.currentChildMember < len(self.childPopulation):
return True
return False
def continueToNextMember(self):
if self.trainingMode == Network.PatternType.Test:
self.currentAlphaMember = self.currentAlphaMember + 1
else:
if not self.runningChildren:
self.currentMember = self.currentMember + 1
else:
self.currentChildMember = self.currentChildMember + 1
def continueToNextGeneration(self):
self.repopulate()
#print("G:" + str(self.generation) + " F[" + ", ".join(str(int(m.fitness)) for m in self.population) + "] Alph:" + str(int(self.alphas[0].fitness)) + " Avg: " + str(int(self.averageFitness())) + " P:" + str(round(self.childSuccess, 3)))
self.generation = self.generation + 1
# self.currentMember = 0
self.runningChildren = False
self.currentChildMember = 0
self.childPopulation = []
def updateMemberFitness(self, error):
if not self.runningChildren:
return self.population[self.currentMember].adjustFitness(error)
return self.childPopulation[self.currentChildMember].adjustFitness(error)
def getCurrentMemberWeightsForNeuron(self, neuronNumber):
"""Get method Neurons use to fetch their coorisponding weights from the current member's genome"""
if self.trainingMode == Network.PatternType.Test:
return self.alphas[0].getGenesAtPosition(neuronNumber)
if not self.runningChildren:
return self.population[self.currentMember].getGenesAtPosition(neuronNumber)
return self.childPopulation[self.currentChildMember].getGenesAtPosition(neuronNumber)
def setCurrentMemberWeightsForNeuron(self, neuronNumber, weights):
"""Set method Neurons use to change their coorisponding weights within the current member's genome"""
if self.trainingMode == Network.PatternType.Test:
return self.alphas[0].setGenesAtPosition(neuronNumber)
if not self.runningChildren:
return self.population[self.currentMember].setGenesAtPosition(neuronNumber)
return self.childPopulation[self.currentChildMember].setGenesAtPosition(neuronNumber)
# def moreMembers(self):
# if self.currentMember < len(self.population):
# return True
# return False
# def continueToNextMember(self):
# self.currentMember = self.currentMember + 1
# def continueToNextGeneration(self):
# print("Average Fitness: " + str(int(self.averageFitness())))
# parents = self.select()
# for p in parents:
# print(str(p[0].fitness) + " " + str(p[1].fitness))
# self.childPopulation = self.crossover(parents)
# self.childPopulation = self.mutate(self.childPopulation)
# self.repopulate()
# self.generation = self.generation + 1
# self.currentMember = 0
# self.currentChildMember = 0
# self.childPopulation = []
# def getCurrentMemberWeightsForNeuron(self, neuronNumber):
# """Get method Neurons use to fetch their coorisponding weights from the current member's genome"""
# if self.trainingMode == Network.PatternType.Test:
# return self.alphas[0].getGenesAtPosition(neuronNumber)
# return self.population[self.currentMember].getGenesAtPosition(neuronNumber)
# def setCurrentMemberWeightsForNeuron(self, neuronNumber, weights):
# """Set method Neurons use to change their coorisponding weights within the current member's genome"""
# if self.trainingMode == Network.PatternType.Test:
# return self.alphas[0].setGenesAtPosition(neuronNumber)
# return self.population[self.currentMember].setGenesAtPosition(neuronNumber, weights)
#---- Genetic Algorithm Methods -----------------------------------------
def select(self):
"""Returns a list of parents chosen for crossover"""
raise("Instance of an Abstract Class... Bad Juju!")
def crossover(self, parents):
"""Returns a list of newly minted children"""
raise("Instance of an Abstract Class... Bad Juju!")
def mutate(self):
"""Go through all members of the population mutating at chance"""
raise("Instance of an Abstract Class... Bad Juju!")
def evaluateFitness(self):
raise("Instance of an Abstract Class... Bad Juju!")
def repopulate(self):
"""Given the current population and the population of children, combine to produce the next Generation"""
raise("Instance of an Abstract Class... Bad Juju!")
class EvolutionStrategy(TrainingStrategy):
def __init__(self):
super(self.__class__, self).__init__()
self.strategy = TrainingStrategyType.EvolutionStrategy
self.lam = 1.0
self.useSigmas = True
self.sigmaMax = .001
self.strongerParentPreference = .5
self.alphaCategoryMap = {}
def select(self):
moms = random.sample(self.population, self.lam)
dads = random.sample(self.population, self.lam)
for i in range(self.lam):
yield [moms[i], dads[i]]
def crossover(self, parents):
#Uniform Crossover, produces 1 child per pair of parents
#We use highestCurrentMemberId to check which of the members of the next generation are children of this generation
self.highestCurrentMemberId = Member.memberIdInc
#print(len(parents))
for pair in parents:
pair = list(pair)
pair.sort(key=lambda x: x.fitness, reverse=False)
child = Member(0, 1, self.useSigmas, self.sigmaMax)
# by gene we also mean sigmas as the crossover for these is the same
for g, gene in enumerate(child.genome):
geneSuccessMod = self.strongerParentPreference*pair[0].highPerformers[g] - self.strongerParentPreference*pair[1].highPerformers[g]
# print("Mod:" + str(geneSuccessMod) + "[" + str(pair[0].highPerformers[g]) + "][" + str(pair[1].highPerformers[g]) + "]")
for w, singleWeight in enumerate(gene):
if random.random() <= self.strongerParentPreference + geneSuccessMod:
child.genome[g][w] = pair[0].genome[g][w]
else:
child.genome[g][w] = pair[1].genome[g][w]
if random.random() <= self.strongerParentPreference + geneSuccessMod:
child.sigmas[g][w] = pair[0].sigmas[g][w]
else:
child.sigmas[g][w] = pair[1].sigmas[g][w]
self.childPopulation.append(child)
return self.childPopulation
def mutate(self):
# Use 1/5 rule for sigma mutation
sigmaMod = 1.0
if self.childSuccess > 0.2:
sigmaMod = 1.225
else:
sigmaMod = 0.816
for child in self.childPopulation:
for g, gene in enumerate(child.genome):
for w, singleWeight in enumerate(gene):
child.sigmas[g][w] = child.sigmas[g][w] * sigmaMod
child.genome[g][w] = child.genome[g][w] + random.gauss(0, child.sigmas[g][w])
def evaluateFitness(self):
return 0
def repopulate(self):
# (m+l)-ES
oldAndNew = self.population + self.childPopulation
oldAndNew.sort(key=lambda x: x.fitness, reverse=False)
self.population = oldAndNew[:self.maxPopulationSize]
if self.population[0] not in self.alphas and (len(self.alphas) == 0 or self.population[0].fitness < self.alphas[0].fitness):
self.alphas.append(self.population[0])
self.alphas.sort(key=lambda x: x.fitness, reverse=False)
recordItems(", ".join(str(int(m.fitness)) for m in self.population) + ", " + str(self.childSuccess))
self.childSuccess = 0.0
for member in self.population:
if member.id > self.highestCurrentMemberId:
self.childSuccess = self.childSuccess + 1
self.childSuccess = self.childSuccess/self.maxPopulationSize
# print("G:" + str(self.generation) + " CC:[" + ", ".join(str(m.categoryCoverage) + ":" + str(round(m.fitness, 4)) for m in self.population) + "] AvgSig:" + str(round(self.avgSigma(), 3)) + " MemVar:" + str(round(self.memberVarience(), 4)) + " Alph:" + str(round(self.alphas[0].fitness, 4)) + " Avg: " + str(round(self.averageFitness(), 4)) + " P:" + str(round(self.childSuccess, 4)))
print("G:" + str(self.generation) + " CC:[" + ", ".join(str(m.categoryCoverage) for m in self.population) + "] AvgSig:" + str(round(self.avgSigma(), 3)) + " MemVar:" + str(round(self.memberVarience(), 4)) + " Alph:" + str(round(self.alphas[0].fitness, 4)) + " Avg: " + str(round(self.averageFitness(), 4)) + " P:" + str(round(self.childSuccess, 4)))
class GeneticAlgorithm(TrainingStrategy):
def __init__(self):
super(self.__class__, self).__init__()
self.strategy = TrainingStrategyType.GeneticAlgorithm
self.childSuccess = 0.0
self.highestCurrentMemberId = 0
self.maxGenerations = 80
self.childPopulation = []
def select(self):
self.population.sort(key=lambda x: x.fitness, reverse=False)
if not self.alphas:
self.alphas.append(self.population[0])
else:
self.alphas[0] = self.population[0]
bestMembers = self.population[:len(self.population)/2]
otherMembers = self.population[len(self.population)/2:]
for i in range(len(bestMembers)):
yield [bestMembers[i], otherMembers[i]]
def crossover(self, parents):
"""For the """
self.highestCurrentMemberId = Member.memberIdInc
for pair in parents:
pair = list(pair)
child = Member(0, 1, self.useSigmas, self.sigmaMax)
j = 0
for g, gene in enumerate(child.genome):
for w, singleWeight in enumerate(gene):
if j % 2 == 0:
child.genome[g][w] = pair[0].genome[g][w]
else:
child.genome[g][w] = pair[1].genome[g][w]
j = j + 1
self.childPopulation.append(child)
return self.childPopulation
# parents = list(parents)
# self.highestCurrentMemberId = Member.memberIdInc
# for i, pair in enumerate(parents[0]):
# child = Member(0, 1, self.useSigmas, self.sigmaMax)
# for j, weight in enumerate(gene):
# if j % 2 == 0:
# child.genome.append(parents[0].genome[i][j])
# else:
# child.genome.append(parents[1].genome[i][j])
# self.childPopulation.append(child)
# return self.childPopulation
def mutate(self):
for member in self.population:
for i, gene in enumerate(member.genome):
for j, elem in enumerate(gene):
if self.mutation():
if random.choice([True, False]):
member.genome[i][j] += self.epsilon()
else:
member.genome[i][j] -= self.epsilon()
def mutation(self):
# member = self.population[0]
numberOfElements = len(Member.genomeTemplate)
probability = 1 / float(numberOfElements)
elem = random.uniform(0,numberOfElements) * probability
choice = random.uniform(0, 1)
diff = math.fabs(elem - choice)
if diff > probability:
return False
return True
def evaluateFitness(self, child):
# fitness = 0
# for pattern in patternSet.patterns:
# Network.Layer.setInputs(Network.Net[0], pattern['p'])
# Network.Layer.feedforward(Network.Net[0])
# fitness += Network.Net.calculateConvError(Network.Net, pattern['t'])
# child.fitness = fitness
return 0
# def repopulate(self, contendors):
# bestFit = 0
# nextFit = 0
# for member in contendors:
# if member.fitness > bestFit:
# bestFitMember = member
# elif member.fitness > nextFit:
# nextFitMember = member
# self.population.append(bestFitMember)
# self.population.append(nextFitMember)
def repopulate(self):
self.repopulateGA()
def repopulateGA(self):
oldAndNew = self.population + self.childPopulation
oldAndNew.sort(key=lambda x: x.fitness, reverse=False)
self.population = oldAndNew[:self.maxPopulationSize]
if not self.population:
self.alphas.append(self.population[0])
self.alphas.sort(key=lambda x: x.fitness, reverse=False)
recordItems(", ".join(str(int(m.fitness)) for m in self.population) + ", " + str(self.childSuccess))
self.childSuccess = 0.0
for member in self.population:
if member.id > self.highestCurrentMemberId:
self.childSuccess = self.childSuccess + 1
self.childSuccess = self.childSuccess/self.maxPopulationSize
class DifferentialGA(TrainingStrategy):
def __init__(self):
super(self.__class__, self).__init__()
self.strategy = TrainingStrategyType.DifferentialGA
self.mask = []
self.x = 'alpha' #way of selecting target: alpha, random
self.y = 2 #number of difference vectors
self.z = 'binomial' #crossover operator: mask, binomial. exponential
self.beta = 0.2
self.probability = 0.3
self.mask
self.useSigmas = False
self.childSuccess = 0.0
self.highestCurrentMemberId = 0
self.maxGenerations = 40
def select(self):
return 0
def selectTwo(self):
return random.sample(self.population, 2)
def createMask(self, target):
mask = []
for gene in target.genome:
for elem in gene:
prob = random.uniform(0, 1)
if prob < self.probability:
mask.append(elem)
self.mask = mask
def crossover(self, parents):
self.population.sort(key=lambda x: x.fitness, reverse=False)
if not self.alphas:
self.alphas.append(self.population[0])
else:
self.alphas[0] = self.population[0]
self.createMask(self.alphas[0])
self.highestCurrentMemberId = Member.memberIdInc
self.childPopulation = []
for member in self.population:
trial = self.mutateDiff()
child = Member(0, 1, self.useSigmas, self.sigmaMax)
for i, gene in enumerate(member.genome):
for j, elem in enumerate(gene):
if elem not in self.mask:
child.genome[i][j] = elem
else:
child.genome[i][j] = trial.genome[i][j]
self.childPopulation.append(child)
return self.childPopulation
def mutateDiff(self):
target = self.alphas[0]
randoms = self.selectTwo()
random1 = randoms[0]
random2 = randoms[-1]
trial = Member(0,1,self.useSigmas, self.sigmaMax)
for i, gene in enumerate(random1.genome):
for j, elem in enumerate(gene):
trial.genome[i][j] = target.genome[i][j] + (self.beta * (elem - random2.genome[i][j]))
return trial
def evaluateFitness(self):
return 0
def repopulate(self):
self.repopulateDGA()
def repopulateDGA(self):
newPopulation = []
for i, member in enumerate(self.population):
if member.fitness > self.childPopulation[i]:
newPopulation.append(member)
else:
newPopulation.append(self.childPopulation[i])
newPopulation.sort(key=lambda x: x.fitness, reverse=False)
self.population = newPopulation
self.childSuccess = 0.0
for member in self.population:
if member.id > self.highestCurrentMemberId:
self.childSuccess = self.childSuccess + 1
self.childSuccess = self.childSuccess/self.maxPopulationSize
#print("G:" + str(self.generation) + " CC:[" + ", ".join(str(m.categoryCoverage) + ":" + str(round(m.fitness, 4)) for m in self.population) + "] AvgSig:" + str(round(self.avgSigma(), 3)) + " MemVar:" + str(round(self.memberVarience(), 4)) + " Alph:" + str(round(self.alphas[0].fitness, 4)) + " Avg: " + str(round(self.averageFitness(), 4)) + " P:" + str(round(self.childSuccess, 4)))
def mutate(self):
return 0