纯Python实现人工智能

很久以前微信流行过一个小游戏：打飞机，这个游戏简单又无聊。在2017年来临之际，我就实现一个超级弱智的人工智能（AI），这货可以躲避从屏幕上方飞来的飞机。本帖只使用纯Python实现，不依赖任何高级库。

本文的AI基于neuro-evolution，首先简单科普一下neuro-evolution。从neuro-evolution这个名字就可以看出它由两部分组成-neuro and evolution，它是使用进化算法（遗传算法是进化算法的一种）提升人工神经网络的机器学习技术，其实就是用进化算法改进并选出最优的神经网络。如果你觉得这篇文章看起来稍微还有些吃力，或者想要系统地学习人工智能，那么推荐你去看床长人工智能教程。非常棒的大神之作，教程不仅通俗易懂，而且很风趣幽默。点击这里可以查看教程。

neuro-evolution

定义一些变量：

1. import math
2. import random
4. # 神经网络3层, 1个隐藏层; 4个input和1个output
5. network = [4, [16], 1]
6. # 遗传算法相关
7. population = 50
8. elitism = 0.2
9. random_behaviour = 0.1
10. mutation_rate = 0.5
11. mutation_range = 2
12. historic = 0
13. low_historic = False
14. score_sort = -1
15. n_child = 1

• 1

定义神经网络：

1. # 激活函数
2. def sigmoid(z):
3. return 1.0/(1.0+math.exp(-z))
4. # random number
5. def random_clamped():
6. return random.random()2-1
8. # "神经元"
9. class Neuron():
10. def init(self):
11.   self.biase = 0
12.   self.weights = []
14. def init_weights(self, n):
15.   self.weights = []
16.   for i in range(n):
17.    self.weights.append(random_clamped())
18. def repr(self):
19.   return ‘Neuron weight size:{}  biase value:{}’.format(len(self.weights), self.biase)
21. # 层
22. class Layer():
23. def init(self, index):
24.   self.index = index
25.   self.neurons = []
27. def init_neurons(self, n_neuron, n_input):
28.   self.neurons = []
29.   for i in range(n_neuron):
30.    neuron = Neuron()
31.    neuron.init_weights(n_input)
32.    self.neurons.append(neuron)
34. def repr(self):
35.   return ‘Layer ID:{}  Layer neuron size:{}’.format(self.index, len(self.neurons))
37. # 神经网络
38. class NeuroNetwork():
39. def init(self):
40.   self.layers = []
42. # input:输入层神经元数 hiddens:隐藏层 output:输出层神经元数
43. def init_neuro_network(self, input, hiddens , output):
44.   index = 0
45.   previous_neurons = 0
46.   # input
47.   layer = Layer(index)
48.   layer.init_neurons(input, previous_neurons)
49.   previous_neurons = input
50.   self.layers.append(layer)
51.   index += 1
52.   # hiddens
53.   for i in range(len(hiddens)):
54.    layer = Layer(index)
55.    layer.init_neurons(hiddens[i], previous_neurons)
56.    previous_neurons = hiddens[i]
57.    self.layers.append(layer)
58.    index += 1
59.   # output
60.   layer = Layer(index)
61.   layer.init_neurons(output, previous_neurons)
62.   self.layers.append(layer)
64. def get_weights(self):
65.   data = { ‘network’:[], ‘weights’:[] }
66.   for layer in self.layers:
67.    data[‘network’].append(len(layer.neurons))
68.    for neuron in layer.neurons:
69.     for weight in neuron.weights:
70.      data[‘weights’].append(weight)
71.   return data
73. def set_weights(self, data):
74.   previous_neurons = 0
75.   index = 0
76.   index_weights = 0
78.   self.layers = []
79.   for i in data[‘network’]:
80.    layer = Layer(index)
81.    layer.init_neurons(i, previous_neurons)
82.    for j in range(len(layer.neurons)):
83.     for k in range(len(layer.neurons[j].weights)):
84.      layer.neurons[j].weights[k] = data[‘weights’][index_weights]
85.      index_weights += 1
86.    previous_neurons = i
87.    index += 1
88.    self.layers.append(layer)
90. # 输入游戏环境中的一些条件(如敌机位置), 返回要执行的操作
91. def feed_forward(self, inputs):
92.   for i in range(len(inputs)):
93.    self.layers[0].neurons[i].biase = inputs[i]
95.   prev_layer = self.layers[0]
96.   for i in range(len(self.layers)):
97.    # 第一层没有weights
98.    if i == 0:
99.     continue
100.    for j in range(len(self.layers[i].neurons)):
101.     sum = 0
102.     for k in range(len(prev_layer.neurons)):
103.      sum += prev_layer.neurons[k].biase * self.layers[i].neurons[j].weights[k]
104.     self.layers[i].neurons[j].biase = sigmoid(sum)
105.    prev_layer = self.layers[i]
107.   out = []
108.   last_layer = self.layers[-1]
109.   for i in range(len(last_layer.neurons)):
110.    out.append(last_layer.neurons[i].biase)
111.   return out
113. def print_info(self):
114.   for layer in self.layers:
115.    print(layer)

• 1

1. # "基因组"
2. class Genome():
3. def init(self, score, network_weights):
4.   self.score = score
5.   self.network_weights = network_weights
7. class Generation():
8. def init(self):
9.   self.genomes = []
11. def add_genome(self, genome):
12.   i = 0
13.   for i in range(len(self.genomes)):
14.    if score_sort < 0:
15.     if genome.score > self.genomes[i].score:
16.      break
17.    else:
18.     if genome.score < self.genomes[i].score:
19.      break
20.   self.genomes.insert(i, genome)
22.         # 杂交+突变
23. def breed(self, genome1, genome2, n_child):
24.   datas = []
25.   for n in range(n_child):
26.    data = genome1
27.    for i in range(len(genome2.network_weights[‘weights’])):
28.     if random.random() <= 0.5:
29.      data.network_weights[‘weights’][i] = genome2.network_weights[‘weights’][i]
31.    for i in range(len(data.network_weights[‘weights’])):
32.     if random.random() <= mutation_rate:
33.      data.network_weights[‘weights’][i] += random.random() * mutation_range * 2 - mutation_range
34.    datas.append(data)
35.   return datas
37.         # 生成下一代
38. def generate_next_generation(self):
39.   nexts = []
40.   for i in range(round(elitismpopulation)):
41.    if len(nexts) < population:
42.     nexts.append(self.genomes[i].network_weights)
44.   for i in range(round(random_behaviourpopulation)):
45.    n = self.genomes[0].network_weights
46.    for k in range(len(n[‘weights’])):
47.     n[‘weights’][k] = random_clamped()
48.    if len(nexts) < population:
49.     nexts.append(n)
51.   max_n = 0
52.   while True:
53.    for i in range(max_n):
54.     childs = self.breed(self.genomes[i], self.genomes[max_n], n_child if n_child > 0 else 1)
55.     for c in range(len(childs)):
56.      nexts.append(childs[c].network_weights)
57.      if len(nexts) >= population:
58.       return nexts
59.    max_n += 1
60.    if max_n >= len(self.genomes)-1:
61.     max_n = 0

• 1

1. class Generations():
2. def init(self):
3.   self.generations = []
5. def first_generation(self):
6.   out = []
7.   for i in range(population):
8.    nn = NeuroNetwork()
9.    nn.init_neuro_network(network[0], network[1], network[2])
10.    out.append(nn.get_weights())
11.   self.generations.append(Generation())
12.   return out
13.  
14. def next_generation(self):
15.   if len(self.generations) == 0:
16.    return False
18.   gen = self.generations[-1].generate_next_generation()
19.   self.generations.append(Generation())
20.   return gen
22. def add_genome(self, genome):
23.   if len(self.generations) == 0:
24.    return False
26.   return self.generations[-1].add_genome(genome)
28. class NeuroEvolution():
29. def init(self):
30.   self.generations = Generations()
32. def restart(self):
33.   self.generations = Generations()
35. def next_generation(self):
36.   networks = []
37.   if len(self.generations.generations) == 0:
38.    networks = self.generations.first_generation()
39.   else:
40.    networks = self.generations.next_generation()
42.   nn = []
43.   for i in range(len(networks)):
44.    n = NeuroNetwork()
45.    n.set_weights(networks[i])
46.    nn.append(n)
48.   if low_historic:
49.    if len(self.generations.generations) >= 2:
50.     genomes = self.generations.generations[len(self.generations.generations) - 2].genomes
51.     for i in range(genomes):
52.      genomes[i].network = None
54.   if historic != -1:
55.    if len(self.generations.generations) > historic+1:
56.     del self.generations.generations[0:len(self.generations.generations)-(historic+1)]
58.   return nn
60. def network_score(self, score, network):
61.   self.generations.add_genome(Genome(score, network.get_weights()))

• 1

1. import pygame
2. import sys
3. from pygame.locals import
4. import random
5. import math
7. import neuro_evolution
9. BACKGROUND = (200, 200, 200)
10. SCREEN_SIZE = (320, 480)
12. class Plane():
13. def init(self, plane_image):
14.   self.plane_image = plane_image
15.   self.rect = plane_image.get_rect()
17.   self.width = self.rect[2]
18.   self.height = self.rect[3]
19.   self.x = SCREEN_SIZE[0]/2 - self.width/2
20.   self.y = SCREEN_SIZE[1] - self.height
22.   self.move_x = 0
23.   self.speed = 2
25.   self.alive = True
27. def update(self):
28.   self.x += self.move_x * self.speed
30. def draw(self, screen):
31.   screen.blit(self.plane_image, (self.x, self.y, self.width, self.height))
33. def is_dead(self, enemes):
34.   if self.x < -self.width or self.x + self.width > SCREEN_SIZE[0]+self.width:
35.    return True
37.   for eneme in enemes:
38.    if self.collision(eneme):
39.     return True
40.   return False
42. def collision(self, eneme):
43.   if not (self.x > eneme.x + eneme.width or self.x + self.width < eneme.x or self.y > eneme.y + eneme.height or self.y + self.height < eneme.y):
44.    return True
45.   else:
46.    return False
48. def get_inputs_values(self, enemes, input_size=4):
49.   inputs = []
51.   for i in range(input_size):
52.    inputs.append(0.0)
54.   inputs[0] = (self.x1.0 / SCREEN_SIZE[0])
55.   index = 1
56.   for eneme in enemes:
57.    inputs[index] = eneme.x1.0 / SCREEN_SIZE[0]
58.    index += 1
59.    inputs[index] = eneme.y*1.0 / SCREEN_SIZE[1]
60.    index += 1
61.   #if len(enemes) > 0:
62.    #distance = math.sqrt(math.pow(enemes[0].x + enemes[0].width/2 - self.x + self.width/2, 2) + math.pow(enemes[0].y + enemes[0].height/2 - self.y + self.height/2, 2));
63.   if len(enemes) > 0 and self.x < enemes[0].x:
64.    inputs[index] = -1.0
65.    index += 1
66.   else:
67.    inputs[index] = 1.0
69.   return inputs
71. class Enemy():
72. def init(self, enemy_image):
73.   self.enemy_image = enemy_image
74.   self.rect = enemy_image.get_rect()
76.   self.width = self.rect[2]
77.   self.height = self.rect[3]
78.   self.x = random.choice(range(0, int(SCREEN_SIZE[0] - self.width/2), 71))
79.   self.y = 0
81. def update(self):
82.   self.y += 6
84. def draw(self, screen):
85.   screen.blit(self.enemy_image, (self.x, self.y, self.width, self.height))
87. def is_out(self):
88.   return True if self.y >= SCREEN_SIZE[1] else False
90. class Game():
91. def init(self):
92.   pygame.init()
93.   self.screen = pygame.display.set_mode(SCREEN_SIZE)
94.   self.clock = pygame.time.Clock()
95.   pygame.display.set_caption(‘是AI就躲个飞机’)
97.   self.ai = neuro_evolution.NeuroEvolution()
98.   self.generation = 0
100.   self.max_enemes = 1
101.                 # 加载飞机、敌机图片
102.   self.plane_image = pygame.image.load(‘plane.png’).convert_alpha()
103.   self.enemy_image = pygame.image.load(‘enemy.png’).convert_alpha()
105. def start(self):
106.   self.score = 0
107.   self.planes = []
108.   self.enemes = []
110.   self.gen = self.ai.next_generation()
111.   for i in range(len(self.gen)):
112.    plane = Plane(self.plane_image)
113.    self.planes.append(plane)
115.   self.generation += 1
116.   self.alives = len(self.planes)
118. def update(self, screen):
119.   for i in range(len(self.planes)):
120.    if self.planes[i].alive:
121.     inputs = self.planes[i].get_inputs_values(self.enemes)
122.     res = self.gen[i].feed_forward(inputs)
123.     if res[0] < 0.45:
124.      self.planes[i].move_x = -1
125.     elif res[0] > 0.55:
126.      self.planes[i].move_x = 1
129.     self.planes[i].update()
130.     self.planes[i].draw(screen)
132.     if self.planes[i].is_dead(self.enemes) == True:
133.      self.planes[i].alive = False
134.      self.alives -= 1
135.      self.ai.network_score(self.score, self.gen[i])
136.      if self.is_ai_all_dead():
137.       self.start()
139.  
140.   self.gen_enemes()
142.   for i in range(len(self.enemes)):
143.    self.enemes[i].update()
144.    self.enemes[i].draw(screen)
145.    if self.enemes[i].is_out():
146.     del self.enemes[i]
147.     break
149.   self.score += 1
151.   print(“alive:{}, generation:{}, score:{}”.format(self.alives, self.generation, self.score), end=’\r’)
153. def run(self, FPS=1000):
154.   while True:
155.    for event in pygame.event.get():
156.     if event.type == QUIT:
157.      pygame.quit()
158.      sys.exit()
160.    self.screen.fill(BACKGROUND)
162.    self.update(self.screen)
164.    pygame.display.update()
165.    self.clock.tick(FPS)
167. def gen_enemes(self):
168.   if len(self.enemes) < self.max_enemes:
169.    enemy = Enemy(self.enemy_image)
170.    self.enemes.append(enemy)
172. def is_ai_all_dead(self):
173.   for plane in self.planes:
174.    if plane.alive:
175.     return False
176.   return True
179. game = Game()
180. game.start()
181. game.run()

• 1

AI的工作逻辑

假设你是AI，你首先繁殖一个种群（50个个体），开始的个体大都是歪瓜裂枣（上来就被敌机撞）。但是，即使是歪瓜裂枣也有表现好的，在下一代，你会使用这些表现好的再繁殖一个种群，经过代代相传，存活下来的个体会越来越优秀。其实就是仿达尔文进化论，种群->自然选择->优秀个体->杂交、变异->种群->循环n世代。ai开始时候的表现：

经过几百代之后，ai开始娱乐的躲飞机.