in my impression, the gradient descent is for finding the independent variable that can get the minimum/maximum value of an objective function. So we need an obj. function: \(\mathcal{L}\)

  • an obj. function: \(\mathcal{L}\)
  • The gradient of \(\mathcal{L}: 2x+2\)
  • \(\Delta x\) , The value of idependent variable needs to be updated: \(x \leftarrow x+\Delta x\)

1. the \(\mathcal{L}\) is a context function: \(f(x)=x^2+2x+1\)

how to find the \(x_0\) that makes the \(f(x)\) has the minimum value, via gradient descent?

Start with an arbitrary \(x\), calculate the value of \(f(x)\) :

import random
def func(x):
return x*x + 2*x +1
def gred(x): # the gradient of f(x)
return 2*x + 2 x = random.uniform(-10.0,10.0) #randomly pick a float in interval of (-10, 10)
# x = 10
print('x starts at:', x) y0 = func(x) #first cal
delta = 0.5 #the value of delta_x, each iteration
x = x + delta # === interation ===
for i in range(100):
print('i=',i)
y1 = func(x)
delta = -0.08*gred(x)
print(' delta=',delta)
if y1 > y0:
print(' y1>y0')
# if gred(x) is positive, the x should decrease.
# if gred(x) is negative, the x should increase.
else:
print(' y1<=y0')
# if gred(x) is positive, the x should increase.
# if gred(x) is negative, the x should decrease.
x = x+delta
y0 = y1
print(' x=', x, 'f(x)=', y1)

Let's disscuss how to determin the some_value in the psudo code above.

if \(y_1-y_0\) has a large positive difference, i.e. \(y1 >> y0\), the x should shift backward heavily. so the some_value can be a ratio of \((y_1-y_0)\times(-gradient)\) , Let's say, some_value: \(\lambda = r \times\) gred(x) , here, \(r=0.08\) is the step-size.

The basic gradient descent has many shortcomings which can be found by search the 'shortcoming of gd'.

Another problem of GD algorithm is , What if the \(\mathcal{L}\) does not have explicit expression of its gradient?

Stochastic Gradient Descent(SGD) is another GD algorithm.

The component and implementation of a basic gradient descent in python的更多相关文章

  1. (转)Introduction to Gradient Descent Algorithm (along with variants) in Machine Learning

    Introduction Optimization is always the ultimate goal whether you are dealing with a real life probl ...

  2. Logistic Regression and Gradient Descent

    Logistic Regression and Gradient Descent Logistic regression is an excellent tool to know for classi ...

  3. (转) An overview of gradient descent optimization algorithms

    An overview of gradient descent optimization algorithms Table of contents: Gradient descent variants ...

  4. 机器学习-随机梯度下降(Stochastic gradient descent)

    sklearn实战-乳腺癌细胞数据挖掘(博主亲自录制视频) https://study.163.com/course/introduction.htm?courseId=1005269003& ...

  5. An overview of gradient descent optimization algorithms

    原文地址:An overview of gradient descent optimization algorithms An overview of gradient descent optimiz ...

  6. 机器学习数学基础- gradient descent算法(上)

    为什么要了解点数学基础 学习大数据分布式计算时多少会涉及到机器学习的算法,所以理解一些机器学习基础,有助于理解大数据分布式计算系统(比如spark)的设计.机器学习中一个常见的就是gradient d ...

  7. flink 批量梯度下降算法线性回归参数求解(Linear Regression with BGD(batch gradient descent) )

    1.线性回归 假设线性函数如下: 假设我们有10个样本x1,y1),(x2,y2).....(x10,y10),求解目标就是根据多个样本求解theta0和theta1的最优值. 什么样的θ最好的呢?最 ...

  8. 梯度下降(Gradient Descent)小结

    在求解机器学习算法的模型参数,即无约束优化问题时,梯度下降(Gradient Descent)是最常采用的方法之一,另一种常用的方法是最小二乘法.这里就对梯度下降法做一个完整的总结. 1. 梯度 在微 ...

  9. 机器学习基础——梯度下降法(Gradient Descent)

    机器学习基础--梯度下降法(Gradient Descent) 看了coursea的机器学习课,知道了梯度下降法.一开始只是对其做了下简单的了解.随着内容的深入,发现梯度下降法在很多算法中都用的到,除 ...

随机推荐

  1. 1.继承(extends)、超类(superClass)、子类(subClass)

    注意:继承主要使用的is-a关系 在子类中用一个新的方法来覆盖超类中的方法(override),需要注意的是如果子类之中的方法或者域  被覆盖时,仍然想访问superClass中的方法和域,此时必须使 ...

  2. Linux下安装Python3的django并配置mysql作为django默认数据库(转载)

    我的操作系统为centos6.5 1  首先选择django要使用什么数据库.django1.10默认数据库为sqlite3,本人想使用mysql数据库,但为了测试方便顺便要安装一下sqlite开发包 ...

  3. Uva10562——Undraw the Trees

    上来一看感觉难以下手,仔细想想就是dfs啊!!!! #include <cstdio> #include<iostream> #include<iomanip> # ...

  4. 窗函数法设计FIR滤波器参数特征表

  5. macbook远程连接报错no matching cipher found

    在.ssh/目录下添加config文件 Host xxx.xxx.xxx.xxx Ciphers 3des-cbc KexAlgorithms +diffie-hellman-group1-sha1 ...

  6. 【java高级编程】JDK和CGLIB动态代理区别

    转载:https://blog.csdn.net/yhl_jxy/article/details/80635012 前言 JDK动态代理实现原理(jdk8):https://blog.csdn.net ...

  7. 6_7_8_10html-css

    Ps:  1.标准流   2.浮动  3.定位  CCS重点 <!DOCTYPE html> <html lang="en"> <head> & ...

  8. Defender 防卫者

    发售年份 1980 平台 街机 开发商 Williams 类型 射击 https://www.youtube.com/watch?v=gss3lxeqCok

  9. day061 cookie和session

    一. cookie 1.cookie 的原理 工作原理是:浏览器访问服务端,带着一个空的cookie,然后由服务器产生内容, 浏览器收到相应后保存在本地:当浏览器再次访问时,浏览器会自动带上Cooki ...

  10. Hbase数据库

        1.简介 HBase从诞生至今将近10年,在apache基金会的孵化下,已经变成一个非常成熟的项目,也有许多不同的公司支持着许多不同的分支版本,如cloudra等等. HBase不同于一般的关 ...