用遗传算法训练多层神经网络.zip

import numpy as np import pygame import random import json import matplotlib.pyplot as plt from tqdm import tqdm from pygame.locals import * def relu_l(array): return np.minimum(np.maximum(array, 0), 1) class NetWork: def __init__(self): self.input = np.array([0 for _ in range(8)]) # 输入层 8个神经元 self.l2 = np.array([0 for _ in range(4)]) # 隐含层1 4个神经元 self.l3 = np.array([0 for _ in range(4)]) # 隐含层2 4个神经元 self.output = np.array([0 for _ in range(4)]) # 输出层 4个神经元 self.chromosome = np.random.randn(100) self.rebuild_from_chromosome() def run(self, input): # 输入格式： # ↑ ↗ → ↘ ↓ ↙ ← ↖ self.input = input self.l2 = relu_l(np.dot(self.w_in_to_2, self.input) + self.b_in_to_2) self.l3 = relu_l(np.dot(self.w_2_to_3, self.l2) + self.b_2_to_3) self.output = relu_l(np.dot(self.w_3_to_out, self.l3) + self.b_3_to_out) # 输出格式： # ↑ ← ↓ → return self.output def rebuild_from_chromosome(self): self.w_in_to_2 = self.chromosome[ :32 ].reshape(4, 8) self.w_2_to_3 = self.chromosome[32 :32 + 16 ].reshape(4, 4) self.w_3_to_out = self.chromosome[32 + 16 :32 + 16 + 16 ].reshape(4, 4) self.b_in_to_2 = self.chromosome[32 + 16 + 16 :32 + 16 + 16 + 4 ] self.b_2_to_3 = self.chromosome[32 + 16 + 16 + 4 :32 + 16 + 16 + 4 + 4 ] self.b_3_to_out = self.chromosome[32 + 16 + 16 + 4 + 4:32 + 16 + 16 + 4 + 4 + 4] @classmethod def crossover(cls, a, b): crosspoint = random.randint(0, 99) c_chromosome = np.array([*a.chromosome[:crosspoint], *b.chromosome[crosspoint:]]) c = cls() c.chromosome = c_chromosome c.rebuild_from_chromosome() return c class Player: def __init__(self, action_from, pos, color): self.action_from = action_from self.x, self.y = pos self.direction = [0, 0] self.fitness = 0 self.color = color def update(self, screen, food_pos, show): food_x, food_y = food_pos inputs = np.array( [ int((food_x == self.x) and (food_y < self.y)), int((food_x > self.x) and (food_y < self.y)), int((food_x > self.x) and (food_y == self.y)), int((food_x > self.x) and (food_y > self.y)), int((food_x == self.x) and (food_y > self.y)), int((food_x < self.x) and (food_y > self.y)), int((food_x < self.x) and (food_y == self.y)), int((food_x < self.x) and (food_y < self.y)) ] ) action = self.action_from.run(inputs) self.direction = [0, 0] self.direction[0] += action[3] * 2 # x+ self.direction[0] -= action[1] * 2 # x- self.direction[1] += action[2] * 2 # y+ self.direction[1] -= action[0] * 2 # y- self.x += self.direction[0] self.y += self.direction[1] if self.x > 400: self.x = 400 if self.x < 0: self.x = 0 if self.y > 400: self.y = 400 if self.y < 0: self.y = 0 distance = np.sqrt( abs((food_x - self.x) ** 2) + \ abs((food_y - self.y) ** 2)) self.fitness = 566 - distance if show: rect = pygame.Rect(self.x - 5, self.y - 5, 10, 10) pygame.draw.rect(screen, self.color, rect) def __gt__(self, b): return self.fitness > b.fitness def __int__(self): return self.fitness def main1(gen_end, show): if show: pygame.init() screen = pygame.display.set_mode((400, 400)) players = [Player(NetWork(), [200, 200], ( round(_ * 0.51), 0, round((500 - _) * 0.51) )) for _ in range(500)] font = pygame.font.SysFont(None, 50) else: screen = None players = [Player(NetWork(), [200, 200], (0, 0, 0)) for _ in range(500)] highest_fitnesses = [] average_fitnesses = [] for gen in tqdm(range(1, gen_end + 1, 1)): foodpos = [random.randint(0, 400), random.randint(0, 400)] for _ in range(200): if show: screen.fill(0) for event in pygame.event.get(): if event.type == QUIT: return 0 for player in players: player.update(screen, foodpos, show) if show: foodrect = pygame.Rect(foodpos[0] - 5, foodpos[1] - 5, 10, 10) pygame.draw.rect(screen, (0, 255, 0), foodrect) text = font.render("GEN {}".format(gen), True, (255, 255, 255)) screen.blit(text, (0, 0)) pygame.display.update() # 排序并算出最高适应度和平均适应度 players.sort() highest_fitness = players[-1].fitness highest_fitnesses.append(highest_fitness) sum_fitness = 0 for player in players: sum_fitness += player.fitness average_fitness = sum_fitness / len(players) average_fitnesses.append(average_fitness) # 选择（选择离目标最近的50个） players = players[-50:] # 交叉（随机选择450组，单点交叉后放入队列末端） for _ in range(450): p1 = random.choice(players) p2 = random.choice(players) n3 = NetWork.crossover(p1.action_from, p2.action_from) p3 = Player(n3, [200, 200], (0, 0, 0)) players.append(p3) # 变异（每个player有一定几率改变染色体中任意五个数） for i in range(len(players)): if random.randint(1, 10) == 1 and i > 0: for _ in range(5): mutationpoint = random.randint(0, 99) players[i].action_from.chromosome[mutationpoint] = np.random.rand() # 将所有玩家放回起点，重置适应值，上色 for i in range(len(players)): players[i].x = 200 players[i].y = 200 players[i].fitness = 0 if show: players[i].color = ( round(i * 0.51), 0, round((500 - i) * 0.51) ) # 演化结束，将适应度最高的玩家的染色体放入res.json players.sort() res = { 'gen': gen_end, 'chromosome': list(players[-1].action_from.chromosome) } res2 = { 'gen_end': gen_end, 'highest_fitnesses': highest_fitnesses, 'average_fitnesses': average_fitnesses } with open("res.json", "w") as f_obj: json.dump(res, f_obj) with open("fitnesses.json", 'w') as f_obj: json.dump(res2, f_obj) pygame.quit() def main2(): with open("fitnesses.json", "r") as f_obj: res = json.load(f_obj) #fig = plt.Figure() x = list(range(1, res['gen_end'] + 1)) y = res['highest_fitnesses'] y2 = res['average_fitnesses'] highest, = plt.plot(x, y, label="highest fitness") average, = plt.plot(x, y2, label="average fitness") plt.xlabel("Generation") plt.ylabel("Fitness") plt.legend(handles=[highest, average], loc='best') plt.show() def main3(): pygame.init() with open("res.json", "r") as f_obj: res = json.load(f_obj) gen = res['gen'] net = NetWork() net.chromosome = np.array(res["chromosome"]) net.rebuild_from_chromosome() player = Player(net, (200, 200), (255, 255, 255)) foodpos = [random.randint(0, 400), random.randint(0, 400)] font = pygame.font.SysFont(None, 50) screen = pygame.display.set_mode((400, 400)