-
Notifications
You must be signed in to change notification settings - Fork 0
/
ga.py
214 lines (180 loc) · 8.31 KB
/
ga.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
import operator
import os
import random
from time import time
import numpy as np
from gym_evaluator import GymEnvironment
from inidividual import Individual
from network import Network
from multiprocessing import Pool
import tensorflow as tf
import gc
import gym
import csv
gym.logger.set_level(40)
class GeneticAlgorithm:
def __init__(self, threads, env_name: str, max_episode_len: int, elite_choose_best_count: int, min_equal_steps: int,
render_each: int, logdir, nn_width: int, seed: int = 42):
self._seed = seed
gym = GymEnvironment(env_name)
self._input_shape = gym.state_shape
self._output_shape = gym.action_shape
self._threads = threads
self._env_name = env_name
self._max_episode_len = max_episode_len
self._elite_choose_best_count = elite_choose_best_count
self._min_equal_steps = min_equal_steps
self._render_each = render_each
self._logdir = logdir
self.nn_width = nn_width
def fit(self, generation_count, population_size, sigma, truncation_size, elitism_evaluations, sigma_final=None):
"""main ga cycle"""
population = self.init_population(population_size)
elite = None
output_csv_path = os.path.join(self._logdir, "metrics.csv")
if not sigma_final:
sigma_final = sigma
sigmas = np.linspace(sigma, sigma_final, generation_count)
for g in range(1, generation_count + 1):
sigma = sigmas[g]
generation_start_time = time()
print(f"GENERATION {g}", flush=True)
new_population = []
# paralelize
parents = population[-truncation_size:]
for _ in range(population_size):
offspring = self.generate_offspring(parents, sigma)
new_population.append(offspring)
print(f"Offspring generation ({time() - generation_start_time:.2f}s)", end="", flush=True)
start_time = time()
max_int = 2 ** 63 - 1
# fitnesses = [self.evaluate_fitness(
# (ind.network.get_weights(), random.randint(0, max_int))) for ind in new_population]
with Pool(self._threads) as pool:
fitnesses = pool.map(self.evaluate_fitness, [
(ind.network.get_weights(), random.randint(0, max_int)) for ind in new_population])
print(f"\rFitness computation ({time() - start_time:.2f}s)", end="", flush=True)
for index in range(len(population)):
new_population[index].fitness = fitnesses[index]
# descending sort
new_population.sort(key=lambda x: x.fitness)
start_time = time()
elite = self.get_elite(elite, new_population, elitism_evaluations)
print(f"\rElite chosen ({time() - start_time:.2f}s)", end="", flush=True)
# remove elite (if it exists) and readd it
try:
new_population.remove(elite)
except ValueError:
# if there's no elite, exclude first member (with worst fitness)
new_population.pop(0)
new_population.append(elite)
population = new_population
# specify and log metrics
fitnesses = np.array([ind.fitness for ind in population])
mean = np.mean(fitnesses)
std = np.std(fitnesses, ddof=1)
quantiles = np.quantile(fitnesses, [0.25, 0.5, 0.75])
best_fitness = elite.fitness
with open(output_csv_path, "a") as file:
writer = csv.writer(file, delimiter=",")
writer.writerow([best_fitness, mean, std, quantiles[0], quantiles[1], quantiles[2]])
# save network weights of the elite
elite.network.save_weights(os.path.join(self._logdir, f"cp-{g}.h5"))
output_str = f"best: {best_fitness:.4f}, mean: {mean:.4f}, std: {std:.4f}, q1: {quantiles[0]:.4f}, " \
f"q2(med): {quantiles[1]:.4f}, q3: {quantiles[2]:.4f}"
print(f"\rGeneration {g} ({time() - generation_start_time:.2f}s): ", output_str, flush=True)
print()
gc.collect()
tf.keras.backend.clear_session()
def generate_offspring(self, parents, sigma):
"""
Generates offspring from selected list of parents adding sample from
normal distribution with zero mean and sigma parameter.
:param parents: Parents from which offspring will be generated.
:param sigma: Sigma of normal distribution
"""
chosen_parent: Individual = random.choice(parents)
offspring = chosen_parent.clone()
self.mutate(offspring, sigma)
return offspring
def init_population(self, population_size):
max_int = 2 ** 63 - 1
population = []
for _ in range(population_size):
seed = random.randint(0, max_int)
network = Network(self._input_shape, self._output_shape, seed=seed, nn_width=self.nn_width)
ind = Individual(network)
population.append(ind)
return population
def mutate(self, individual, sigma):
"""
Mutates specified individual. It performs modifications on existing instance, doesn't create new one.
:param individual:
:param sigma:
:return:
"""
network = individual.network
weights = network.get_weights()
modified_weights = []
for w in weights:
update = tf.random.normal(shape=w.shape, stddev=sigma)
modified_weights.append(w + update)
network.set_weights(modified_weights)
def get_elite(self, elite, population, elitism_evaluations):
# elitism
# candidates - 10 best + last gen elite
max_int = 2 ** 63 - 1
choose_best_count = self._elite_choose_best_count
best_from_population = population[-choose_best_count:]
candidates = best_from_population
if elite is not None:
candidates.append(elite)
# choose best candidate according to mean in -elitism_evaluations- evals
for candidate in candidates:
# candidate_fitnesses = [self.evaluate_fitness((ind.network.get_weights(), random.randint(0, max_int)))
# for ind in [candidate] * elitism_evaluations]
with Pool(self._threads) as pool:
candidate_fitnesses = pool.map(self.evaluate_fitness, [
(ind.network.get_weights(), random.randint(0, max_int)) for ind in [candidate] * elitism_evaluations])
candidate_fitnesses.append(candidate.fitness)
candidate.fitness = np.mean(candidate_fitnesses)
new_elite = max(candidates, key=operator.attrgetter('fitness'))
return new_elite
def evaluate_fitness(self, params):
"""
Evaluates fitness of specified individual.
:return: Fitness of the individual
"""
network_weights = params[0]
seed = params[1]
network = Network(self._input_shape, self._output_shape, seed, nn_width=self.nn_width, initializer="zeros")
network.set_weights(network_weights)
gym = GymEnvironment(self._env_name, seed=seed)
state, done = gym.reset(), False
# start_time = time()
step = 0
equal_steps = 0
rewards = []
while not done:
if self._render_each and step % self._render_each == 0:
gym.render()
state = np.expand_dims(state, 0)
action = network(state).numpy()[0]
next_state, reward, done, _ = gym.step(action)
if self._min_equal_steps > 0:
if np.allclose(state, next_state):
equal_steps += 1
else:
equal_steps = 0
rewards.append(reward)
state = next_state
step += 1
if step >= self._max_episode_len:
done = True
elif self._min_equal_steps > 0 and equal_steps >= self._min_equal_steps:
done = True
# add expected reward if we waited till the episode would end
rewards.append((self._max_episode_len - step) * np.mean(rewards[-self._min_equal_steps:]))
# print(f"Total steps {step}: {time() - start_time:.4f}")
total_reward = np.sum(rewards)
return total_reward