自研进化模拟器生物往复晃动不趋近食物 代码问题排查求助

存在的核心问题 #

• 训练目标设定错误：原代码中move函数内计算的ny、nx作为训练目标完全不符合预期，导致网络学习方向完全错误
• 权重更新逻辑错误：Think模块的adjust方法中更新weightbx时错误使用了y方向的误差erry，而非x方向的errx
• 数值计算错误：主模块move函数中计算nx时误写为减dely，应该减delx
• 学习率过大：初始学习率lr=1会导致权重更新幅度过大，模型参数反复振荡，表现为生物来回移动
• 类属性误用：Creature的brain被定义为类属性，所有实例会共用同一个神经网络
• 代码缩进错误：多处Python缩进不符合规范，会直接导致运行报错

修复后代码 #

主程序模块 #

import turtle, random, Think, time

world = turtle.Screen()
world.setup(400, 400) # 坐标范围O(0,0),X(180,0),X'(-185,0),Y(0,185),Y'(0,-180)

def growfood():
    food = turtle.Turtle()
    food.color("green")
    food.shape("circle")
    food.shapesize(0.4)
    food.penup()
    food.speed(0)
    food.goto(random.randint(-185, 180), random.randint(-180, 185))
    return food

class Creature:
    body = None
    def __init__(self):
        self.body = turtle.Turtle()
        self.body.penup() # 取消注释避免留下移动轨迹
        self.body.color("blue")
        self.body.shape("square")
        self.body.shapesize(0.5)
        self.brain = Think.neuron() # 改为实例属性，每个生物单独一个脑

    def move(self, y, x, dely, delx):
        self.body.goto(self.body.xcor() + x*5, self.body.ycor() + y*5)
        # 修正训练目标：目标输出为距离的符号，即向食物方向移动
        target_y = 1 if dely > 0 else (-1 if dely < 0 else 0)
        target_x = 1 if delx > 0 else (-1 if delx < 0 else 0)
        self.brain.adjust(dely, delx, target_y, target_x, y, x)

    def locatefood(self):
        delx = food.xcor() - self.body.xcor()
        dely = food.ycor() - self.body.ycor()
        return dely, delx

    def think(self):
        dely, delx = self.locatefood()
        y, x = self.brain.think(dely, delx)
        print(f"(x,y)=({x},{y})")
        self.move(y, x, dely, delx)

crt = Creature()      
food = growfood()
run = 0
while run < 120:
    world.update()
    crt.think()
    if crt.body.distance(food) < 15:
        food = growfood()
    time.sleep(0.2) # 调小等待时间运行更流畅

Think神经网络模块 #

from math import tanh
from random import random

class neuron:
    def __init__(self):
        self.biasy, self.biasx = 1, 1
        self.lr = 0.0001 # 调小学习率避免振荡
        self.weighty, self.weightx, self.weightby, self.weightbx = random(), random(), random(), random()
    
    def adjust(self, dely, delx, target_y, target_x, y, x):
        # 修正y方向权重更新
        erry = target_y - y
        self.weighty += erry * dely * self.lr
        self.weightby += erry * self.biasy * self.lr
        # 修正x方向权重更新，使用x方向误差errx
        errx = target_x - x
        self.weightx += errx * delx * self.lr
        self.weightbx += errx * self.biasx * self.lr # 修正此处误用erry的问题

    def think(self, dely, delx):
        y = dely * self.weighty + self.biasy * self.weightby
        x = delx * self.weightx + self.biasx * self.weightbx
        y = tanh(y)
        if y > 0:
            y = 1
        elif y < 0:
            y = -1
        else:
            y = 0
        x = tanh(x)
        if x > 0:
            x = 1
        elif x < 0:
            x = -1
        else:
            x = 0
        return y, x

内容的提问来源于stack exchange，提问作者Himel