-
Notifications
You must be signed in to change notification settings - Fork 0
/
functions.py
564 lines (528 loc) · 24 KB
/
functions.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
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
import pylab
import numpy as np
from numpy.random import choice, randint
from random import random, randint
import random as rd
from copy import deepcopy
import networkx as nx
from networkx.drawing.nx_agraph import graphviz_layout
import matplotlib.pyplot as plt
import math
from time import time
from multiprocessing import Pool
import operator
#library of elements; there are three types: value, function (+,-,*,/) or graph (which will be added at the end of the file)
lib = [{"type":"value","symbol":"v"},
{"type":"function","function":lambda a:a[0]+a[1],"symbol":"+"},
{"type":"function","function":lambda a:a[0]-a[1],"symbol":"-"},
{"type":"function","function":lambda a:a[0]*a[1],"symbol":"*"},
{"type":"function","function":lambda a:a[0]/a[1],"symbol":"/"}]
class Tree:
def __init__(self,graph = None,values=None,variable_position=0):
self.fitness = 1
if graph == None:
self.graph = []
self.values = []
self.variable_position = 0
else:
self.graph = graph
if values == None:
n_args = self.number_of_arguments()
self.values = [random() for n in range(n_args)]
else:
self.values = values
self.variable_position = variable_position
#print("new graph created")
#print(self.values)
def clone(self):
return Tree(self.graph)
#function that takes in a graph and a list of input numbers and evaluates its result
def evaluate(self,values,index=-1):
#this case corresponds to
if index == -1:
index = self.find_output_node()
node = self.graph[index]
lib_entry = lib[node["lib_id"]]
node_type = lib_entry["type"]
if node_type == "value":
return values[node["arg_index"]]
else:
args = []
for argument in node["input"]:
arg = self.evaluate(values,argument)
args.append(arg)
if node_type == "function":
return lib_entry["function"](args)
elif node_type == "graph":
return eval(lib_entry["graph"](args),args)
#returns the position of the output node within the graph
def find_output_node(self):
referenced = []
for node in self.graph:
referenced += node["input"]
for n, node in enumerate(self.graph):
if n not in referenced:
return n
#returns the number of arguments of a graph
def number_of_arguments(self):
n_arg = 0
for n, node in enumerate(self.graph):
lib_entry = lib[node["lib_id"]]
node_type = lib_entry["type"]
if node_type == "value":
n_arg += 1
return n_arg
#returns a list with elements of the type [position of the argument in the graph,argument index in the argument vector]
def input_info(self):
args = []
for n, node in enumerate(self.graph):
lib_entry = lib[node["lib_id"]]
node_type = lib_entry["type"]
if node_type == "value":
args.append([n,node["arg_index"]])
n_args = len(args)
arguments = [0 for i in range(n_args)]
for arg in args:
arguments[arg[1]] = arg[0]
return arguments
def merge_input_entries(self,i=-1,j=-1):
n_args = self.number_of_arguments()
if n_args <= 1:
return Tree(self.graph)
arg_vector = [k for k in range(0,n_args)]
# if i is out of range pick a random value within range
i = i if i in arg_vector else randint(0,n_args-1)
# if j is out of range pick a random value within range which is different from i
arg_vector.pop(i)
j = j if j in arg_vector else choice(arg_vector)
#index to keep
min_index = min(sorted([i,j]))
#index to replace with min_index
max_index = max(sorted([i,j]))
input_indexes = self.input_info()
new_graph = deepcopy(self.graph)
#remove node corresponding to max_index in the input vector
kept_node_index = input_indexes[min_index]
scrapped_node_index = input_indexes[max_index]
del new_graph[scrapped_node_index]
#update all references within graph:
#create a mapping for the updated positions in the input vector
input_mapping = [index if index < max_index else (min_index if index == max_index else index-1) for index in range(n_args)]
#create a mapping for the updated positions in the graph
if kept_node_index < scrapped_node_index:
graph_mapping = [index if index < scrapped_node_index else (kept_node_index if index == scrapped_node_index else index-1) for index in range(len(self.graph))]
elif kept_node_index > scrapped_node_index:
graph_mapping = [index if index < scrapped_node_index else (kept_node_index - 1 if index == scrapped_node_index else index-1) for index in range(len(self.graph))]
#update the references
for node in new_graph:
node["input"] = [ graph_mapping[n] for n in node["input"] ]
for n in node["input"]:
n = graph_mapping[n]
#if it's an input node, update arg_index
if len(node["input"]) == 0:
node["arg_index"] = input_mapping[node["arg_index"]]
return Tree(new_graph)
#this function draws the graph
def draw(self):
G = nx.MultiDiGraph()
for n, node in enumerate(self.graph):
G.add_node(n,attr_dict=node)
for n1, node in enumerate(self.graph):
for i, n2 in enumerate(node["input"]):
G.add_edge(n2, n1, attr_dict={"arg":i})
labels = {}
for n, node in enumerate(G.nodes(data=True)):
symbol = lib[node[1]["lib_id"]]["symbol"]
labels[n] = symbol if symbol != "v" else node[1]["arg_index"]
plt.close()
nx.draw(G,labels = labels,with_labels = True,pos=nx.spring_layout(G))
plt.show()
#plot a 1D graph of the function given some parameters
def plot_specimen(self,values=None,x_range=[-10,10],variable_position=-1):
#pylab.close()
x = np.linspace(x_range[0],x_range[1],100)
n_args = self.number_of_arguments()
if values == None:
values = self.values
x_pos = variable_position if variable_position in range(n_args) else self.variable_position
#input_size*number_of_data_points matrix that contains a list of input vectors like [parameter1, parameter2, x_value, ...]
value_matrix = [[(xe if v == x_pos else value) for v, value in enumerate(values)] for xe in x]
y = np.array([ self.evaluate(value_matrix[i]) for i in range(len(value_matrix)) ])
pylab.plot(x,y)
pylab.show()
#plot a 1D graph of the function given some parameters
def plot_with_data(self,data,x_range=[-10,10]):
#pylab.close()
x = np.linspace(x_range[0],x_range[1],100)
n_args = self.number_of_arguments()
x_pos = self.variable_position
values = self.values
#input_size*number_of_data_points matrix that contains a list of input vectors like [parameter1, parameter2, x_value, ...]
value_matrix = [[(xe if v == x_pos else value) for v, value in enumerate(values)] for xe in x]
y = np.array([ self.evaluate(value_matrix[i]) for i in range(len(value_matrix)) ])
pylab.plot(x,y)
x = data["x"]
y = data["y"]
pylab.scatter(x,y)
pylab.show()
#given a set of parameters and a graph, compute the error relative to a dataset x, y
def error(self,data,parameters=None,variable_position=None):
if parameters == None:
parameters = self.values
if variable_position == None:
variable_position = self.variable_position
#print("computing error...")
#print("current parameters: {0}".format(self.values))
X = data["x"]
Y = data["y"]
e = 0
arg = deepcopy(parameters)
for x,y in zip(X,Y):
arg[variable_position] = x
e += (y-self.evaluate(arg))**2
return e if e == e else 10**10
def optimize_once(self,data):
EPSILON = 0.001
n_args = self.number_of_arguments()
#print("number of arguments: {0}".format(n_args))
best_error = self.error(data,variable_position=0)
best_parameters = {"X":0,"values":self.values,"error":best_error}
for variable_position in range(n_args):
grad = gradient(lambda a: self.error(data,parameters=a,variable_position=variable_position),self.values)
current_error = self.error(data,variable_position=variable_position)
new_values = self.values - EPSILON*grad
new_error = self.error(data,parameters = new_values,variable_position = variable_position)
if new_error < best_error:
best_parameters = {"X":variable_position,"values":new_values,"error":new_error}
best_error = new_error
self.values = best_parameters["values"]
self.variable_position = best_parameters["X"]
return best_parameters
#optimizes a given graph on a set of data points x and y
def optimize(self,data,variable_position=-1,timeout=False):
start = time()
n_args = self.number_of_arguments()
#select a random position in the input vector for the x value
if variable_position == -1:
variable_position = randint(0,n_args-1)
#seed input
values = np.random.rand(n_args)
#do this for a thousand steps
EPSILON = 1
TOLERANCE = 0.01
iterations = 0
MAX_ITERATIONS = 50
while True:
grad = gradient(lambda a: self.error(a,variable_position,data),values)
current_error = self.error(values,variable_position,data)
#if the accuracy goal has been reached, stop and return the input vector
if current_error < TOLERANCE:
break
while True:
iterations += 1
enough = (time() - start > 10) or (iterations > MAX_ITERATIONS) if timeout else iterations > MAX_ITERATIONS
#if iterations > MAX_ITERATIONS:
#if (time() - start > 10) or (iterations > MAX_ITERATIONS):
if enough:
return {"X": variable_position, "parameters": values, "iterations": iterations, "error": current_error}
next_error = self.error(values - EPSILON*grad,variable_position,data)
#if the change increases the error, reduce the size of the step
if next_error > current_error:
EPSILON *= 0.5
#else just move on
else:
break
#print("epsilon: {0}".format(EPSILON))
values += -EPSILON*grad
return {"X": variable_position, "parameters": values, "iterations": iterations, "error": current_error}
def full_optimize(self,data):
min_error = 10**10
n_args = self.number_of_arguments()
for n in range(n_args):
optimization = self.optimize(data,variable_position=n)
if optimization["error"] < min_error:
best = optimization
return optimization
#plot the optimal fit os a graph onto a set of data points
def plot_optimized(self,data):
pylab.plot(data["x"],data["y"])
error = 10**10
for i in range(self.number_of_arguments()):
result = self.optimize(data,i)
if result["error"] < error:
error = result["error"]
values = result["parameters"]
x_pos = result["X"]
self.plot_specimen(values,[min(data["x"]),max(data["x"])],x_pos)
#fitness of a given graph
def unfitness(self,data):
start = time()
result = self.full_optimize(data)
end = time()
return result["error"] + 1*(end-start)
################################################################################################################################
#g = {"type":"graph","graph":Tree([{"lib_id":0,"input":[],"arg_index":0}])}
#lib.append(g)
#Add a graph to the lib for each operator
#for n in range(1,5):
# g = {"type":"graph","graph":Tree([{"lib_id":0,"input":[],"arg_index":0},
# {"lib_id":0,"input":[],"arg_index":1},
# {"lib_id":n,"input":[0,1]}])}
# lib.append(g)
################################################################################################################################
class Population:
def __init__(self,maxpop = 100):
self.BIRTH_PROBABILITY = 0.02
self.MUTATION_PROBABILITY = 0.02
self.DEATH_PROBABILITY = 0.02
self.MAXIMUM_POPULATION_SIZE = maxpop
self.population = []
for n, item in enumerate(lib):
if item["type"] == "value":
tree = Tree(graph=[{"lib_id":0,"input":[],"arg_index":0}],values=[random()])
self.population.append(tree)
elif item["type"] == "function":
input_vector = [random() for i in range(2)]
tree = Tree(graph=[{"lib_id":0,"input":[],"arg_index":0},
{"lib_id":0,"input":[],"arg_index":1},
{"lib_id":n,"input":[0,1]}],values=input_vector)
self.population.append(tree)
def evolve(self,data,generations=100):
for n in range(generations):
print("{0}-th year, population size = {1}".format(n,len(self.population)))
self.evolution_step(data)
return self.best_specimen(data)
def birth(self,sorted_population=None):
if sorted_population == None:
recipient = rd.choice(self.population)
donor = rd.choice(self.population)
else:
recipient = rd.choice(sorted_population[int(0.8*len(sorted_population)):])
recipient = self.population[recipient[0]]
donor = rd.choice(sorted_population[:int(0.8*len(sorted_population))])
donor = self.population[donor[0]]
baby = insert_at(donor,recipient)
return baby
def mutate(self,sorted_population=None):
if sorted_population == None:
mutant = rd.choice(self.population).merge_input_entries()
else:
mutant = rd.choice(sorted_population[int(0.8*len(sorted_population)):])
mutant = self.population[mutant[0]]
return mutant
def kill(self,sorted_population=None):
if len(self.population) < 10:
return None
if sorted_population == None:
mark = randint(0,len(self.population))
else:
mark = rd.choice(sorted_population[:int(0.2*len(sorted_population))])
mark = mark[0]
return self.population.pop(mark) if 4 < mark < len(self.population) else None
def evolution_step(self,data,n_steps=1):
for specimen in self.population:
#print("optimizing... ")
#specimen.draw()
result = specimen.optimize_once(data)
#print("result: {0}".format(result))
sorted_population = self.sort_population(data)
birth_coin = random()
if birth_coin <= self.BIRTH_PROBABILITY:
#print("birth:")
baby = self.birth(sorted_population=sorted_population)
#baby.draw()
self.population.append(baby)
mutation_coin = random()
if mutation_coin <= self.MUTATION_PROBABILITY:
#print("mutation:")
mutant = self.mutate(sorted_population=sorted_population)
#mutant.draw()
self.population.append(mutant)
death_coin = random()
if death_coin <= self.DEATH_PROBABILITY:
self.kill(sorted_population=sorted_population)
def best_specimen(self,data):
best_error = 10**10
best = 0
for specimen in self.population:
error = specimen.error(data)
if error <= best_error:
best = specimen
best_error = error
return best
def worst_specimen(self,data):
worst_error = 0
worst = 10**10
for specimen in self.population:
error = specimen.error(data)
if error >= worst_error:
worst = specimen
worst_error = error
return worst
def sort_population(self,data):
errors = {}
for n, specimen in enumerate(self.population):
error = specimen.error(data)
errors[n] = error
sorted_population = sorted(errors.items(), key=operator.itemgetter(1))
def plot_population(self):
return False
#################################################################################################################################
#function that combines two graphs by using the output node of the second graph as one of the inputs of the first
def insert_at(donor_graph,recipient_graph,site=-1):
recipient_args = recipient_graph.input_info()
donor_args = donor_graph.input_info()
if site == -1:
site = randint(0,len(recipient_args)-1)
site_index = recipient_args[site]
site_arg_index = recipient_graph.graph[site_index]["arg_index"]
donor_output = donor_graph.find_output_node()
new_graph = [deepcopy(node) for node in recipient_graph.graph]
new_graph.pop(site_index)
for n, node in enumerate(new_graph):
lib_entry = lib[node["lib_id"]]
node_type = lib_entry["type"]
if node_type == "function" or node_type == "graph":
node_input = [ni for ni in node["input"]]
for s, subnode in enumerate(node_input):
if subnode > site_index:
node_input[s] = subnode - 1
elif subnode == site_index:
node_input[s] = donor_output + len(new_graph)
node["input"] = node_input
elif node_type == "value":
node["arg_index"] -= (1 if node["arg_index"] > site_arg_index else 0)
processed_donor = [deepcopy(node) for node in donor_graph.graph]
for n, node in enumerate(processed_donor):
lib_entry = lib[node["lib_id"]]
node_type = lib_entry["type"]
if node_type == "function" or node_type == "graph":
node_input = [ni for ni in node["input"]]
node["input"] = [ni + len(new_graph) for ni in node_input]
elif node_type == "value":
node["arg_index"] += len(recipient_args) - 1
return Tree(new_graph+processed_donor)
def print2Dmatrix(matrix):
print('\n'.join([''.join(['{:4}'.format(item) for item in row])
for row in matrix]))
#makes updates the parameters of a graph
def gradient(function,values):
EPSILON = 0.000000001
value_here = np.array([ function(values) for i in range(len(values)) ])
dx = np.array([[ EPSILON if i == j else 0 for i in range(len(values)) ] for j in range(len(values)) ])
value_there = np.array([ function(values+dx[i]) for i in range(len(values)) ])
grad = (value_there-value_here)/EPSILON
return grad
#takes the items from the library and mixes them to make new graphs until it finds one that fits the data
def evolution(data):
ACCURACY_GOAL = 0.01
POPULATION_SIZE = 300
MUTATION_RATE = 0.1
BIRTH_RATE = 0.1
CLONING_RATE = 0.05
DEATH_RATE = MUTATION_RATE + BIRTH_RATE + CLONING_RATE
population = []
#create the initial population
for i in range(POPULATION_SIZE):
graph = choice(lib[5:])["graph"].clone()
population.append(graph)
#compute weights
errors = np.array([ specimen.unfitness(data) for specimen in population ])
#stop if minimum error is below the accuracy goal
min_error = np.ndarray.min(errors)
#save the best
best_error_so_far = 10**10
winner = np.argmin(errors)
population[winner].draw()
population[winner].plot_optimized(data)
if min_error < best_error_so_far:
best_error_so_far = min_error
fittest_specimen = population[winner]
if min_error < ACCURACY_GOAL:
winner = np.argmin(errors)
return {"specimen": population[winner], "error": errors[winner]}
weights = errors/sum(errors)
#weights = np.array([ 1/POPULATION_SIZE for i in range(POPULATION_SIZE) ])
#evolve
for generation in range(20):
print('evolving generation {0} of 20'.format(generation))
#compute weights
new_population = deepcopy(population)
temp_weights = deepcopy(weights)
#DEATH ROUND
#kill int(POPULATION_SIZE*DEATH_RATE) graphs
dead = 0
n_dead = int(POPULATION_SIZE*DEATH_RATE)
for kill in range(n_dead):
to_die = choice(len(new_population),p=temp_weights)
temp_weights = np.delete(temp_weights,to_die)
temp_weights = temp_weights/np.sum(temp_weights)
new_population.pop(to_die)
dead += 1
#print("killed :{0}".format(dead))
#CLONING ROUND
#copy int(POPULATION_SIZE*CLONING_RATE) new graphs
clones = 0
n_clones = int(POPULATION_SIZE*CLONING_RATE)
for i in range(n_clones):
clone_index = choice(len(population))
clone = population[clone_index].clone()
new_population.append(clone)
clones+=1
#print("clones :{0}".format(clones))
#MUTATION ROUND
#add int(POPULATION_SIZE*MUTATION_RATE) new graphs
mutants = 0
n_mutants = int(POPULATION_SIZE*MUTATION_RATE)
for i in range(n_mutants):
mutant_index = choice(len(population))
mutant = population[mutant_index].clone()
mutant = mutant.merge_input_entries()
new_population.append(mutant)
mutants+=1
#print("mutants :{0}".format(mutants))
#REPRODUCTION ROUND
#add int(POPULATION_SIZE*BIRTH_RATE) new graphs
babies = 0
n_babies = n_dead - (n_clones+n_mutants)
for i in range(n_babies):
mother_index = choice(len(population))#,p=(1-weights)/sum(1-weights))#[0]["graph"]
mother = population[mother_index]
father_index = choice(len(population),p=(1-weights)/sum(1-weights))#[0]["graph"]
father = population[father_index]
child = insert_at(father,mother)
new_population.append(child)
babies+=1
#print("babies :{0}".format(babies))
population = deepcopy(new_population)
#recompute the errors for the newcomers
errors = deepcopy(errors)
first_new_index = POPULATION_SIZE-int(POPULATION_SIZE*DEATH_RATE)
for s, specimen in enumerate(population[first_new_index:]):
err = specimen.unfitness(data)
errors[s+first_new_index] = err#specimen.unfitness(data)
if err > 10**7:
return specimen
#stop if minimum error is below the accuracy goal
min_error = np.ndarray.min(errors)
#save the best
winner = np.argmin(errors)
#population[winner].draw()
population[winner].plot_optimized(data)
if min_error < best_error_so_far:
best_error_so_far = min_error
fittest_specimen = population[winner]
if min_error < ACCURACY_GOAL:
winner = np.argmin(errors)
return {"specimen": population[winner], "error": errors[winner]}
weights = errors/sum(errors)
#print("population size : {0}".format(len(population)))
print("min error at generation n."+str(generation)+": "+str(np.ndarray.min(errors)))
#recompute final errors
errors = np.array([ specimen.unfitness(data) for specimen in population ])
winner = np.argmin(errors)
if errors[winner] < best_error_so_far:
best_error_so_far = errors[winner]
fittest_specimen = population[winner]
return {"specimen": fittest_specimen, "error": best_error_so_far}
################################################################################################################################