利用梯度下降求解一元多次函数最值

刚学了TensorFlow2.0的梯度下降原理，觉得是否可以用到连续函数求最值问题中，先从简单的函数入手，一元多次函数，比如：$x^7\ast (1-x{)}^3$
把梯度下降代码稍微改了下，利用TensorFlow2.0自动求导函数

import tensorflow as tf


def func1(x):
    return (x ** 7) * (1 - x) ** 3


x1 = tf.constant(0.5)  # 初始点的值

for step in range(20001):  # 迭代20000次。
    with tf.GradientTape() as tape:
        tape.watch([x1])
        y1 = func1(x1)

    grads = tape.gradient(y1, [x1])[0]
    # print(grads)
    # 求最大值为 + ，最小值为 -
    x1 += 0.01 * grads

    if step % 2000 == 0:
        print('step {0}: x1 = {1}, y1 = {2}'.format(step, x1.numpy(), y1.numpy()))

多迭代几次，这里设置的是2万次。
此时求的是这个函数的最大值，求最小值就改为减号
结果为

step 0: x1 = 0.5000781416893005, y1 = 0.0009765625
step 2000: x1 = 0.647747278213501, y1 = 0.002091002417728305
step 4000: x1 = 0.6929699182510376, y1 = 0.0022209687158465385
step 6000: x1 = 0.6991444826126099, y1 = 0.0022235270589590073
step 8000: x1 = 0.6998952031135559, y1 = 0.00222356547601521
step 10000: x1 = 0.6999719142913818, y1 = 0.0022235659416764975
step 12000: x1 = 0.6999719142913818, y1 = 0.0022235659416764975
step 14000: x1 = 0.6999719142913818, y1 = 0.0022235659416764975
step 16000: x1 = 0.6999719142913818, y1 = 0.0022235659416764975
step 18000: x1 = 0.6999719142913818, y1 = 0.0022235659416764975
step 20000: x1 = 0.6999719142913818, y1 = 0.0022235659416764975

可以看出当x1=0.7时，y1取最大值0.0022235659416764975
x的值不断趋近最值，这就是梯度下降的原理。

要不来个更复杂的函数，例如$2\ast sin(x)+cos(x)$
用遗传算法得到最大值为：2.23606797749979
遗传算法结果为

那我们修改梯度下降的代码，求解这个函数（其实就是改了func1的表达式！！！）

import tensorflow as tf
import math


def func1(x):
    return 2 * tf.sin(x) + tf.cos(x)


x1 = tf.constant(0.5)  # 初始点的值

for step in range(20001):  # 迭代20000次。
    with tf.GradientTape() as tape:
        tape.watch([x1])
        y1 = func1(x1)

    grads = tape.gradient(y1, [x1])[0]
    # print(grads)
    # 求最大值为 + ，最小值为 -
    x1 += 0.01 * grads

    if step % 2000 == 0:
        print('step {0}: x1 = {1}, y1 = {2}'.format(step, x1.numpy(), y1.numpy()))

结果为

step 0: x1 = 0.512757420539856, y1 = 1.8364336490631104
step 2000: x1 = 1.107146143913269, y1 = 2.2360680103302
step 4000: x1 = 1.107146143913269, y1 = 2.2360680103302
step 6000: x1 = 1.107146143913269, y1 = 2.2360680103302
step 8000: x1 = 1.107146143913269, y1 = 2.2360680103302
step 10000: x1 = 1.107146143913269, y1 = 2.2360680103302
step 12000: x1 = 1.107146143913269, y1 = 2.2360680103302
step 14000: x1 = 1.107146143913269, y1 = 2.2360680103302
step 16000: x1 = 1.107146143913269, y1 = 2.2360680103302
step 18000: x1 = 1.107146143913269, y1 = 2.2360680103302
step 20000: x1 = 1.107146143913269, y1 = 2.2360680103302

求得x = 1.107146143913269时，y有最大值2.2360680103302，和遗传算法求的2.23606797749979，基本一样，但是代码量肯定比遗传算法少多了，原理也比它简单，有意思。

本站所有文章均为原创，欢迎转载，请注明文章出处：https://blog.csdn.net/weixin_45092662。百度和各类采集站皆不可信，搜索请谨慎鉴别。技术类文章一般都有时效性，本人习惯不定期对自己的博文进行修正和更新，因此请访问出处以查看本文的最新版本。