Differential Evolution: A Survey of the State-of-the-Art

概
主要内容
- DE/rand/1/bin
- DE/?/?/?
- 超参数的选择
- 一些连续变体
  - A
  - B
  - C
  - D
  - E
  - G
- 一些缺点
代码

Das S, Suganthan P N. Differential Evolution: A Survey of the State-of-the-Art[J]. IEEE Transactions on Evolutionary Computation, 2011, 15(1): 4-31.

@article{das2011differential,

title={Differential Evolution: A Survey of the State-of-the-Art},

author={Das, Swagatam and Suganthan, P N},

journal={IEEE Transactions on Evolutionary Computation},

volume={15},

number={1},

pages={4--31},

year={2011}}

概

这是一篇关于Differential Evolution (DE) 的综述, 由于对这类方法并不熟悉, 只能简单地做个记录.

主要内容

考虑如下问题,

\[\min \: f(X),
\]

其中\(X=(x_1,\ldots,x_D)\).

我所知的, 如梯度下降方法, 贝叶斯优化可以用来处理这类问题, 但是还有诸如 evolutionary algorithm (EA), evolutionary programming (EP), evolution strategies(ESs), genetic algorithm (GA), 以及本文介绍的 DE (后面的基本都不了解).

DE/rand/1/bin

先给出最初的形式, 称之为DE/rand/1/bin:

Input: scale factor \(F\), crossover rate \(Cr\), population size \(NP\).

1: 令\(G=0\), 并随机初始化\(P_G=\{ X_{1,G},\ldots, X_{NP,G}\}\).

2: While the stopping criterion is not satisfied Do:

For \(i=1,\ldots, NP\) do:

Mutation step:

\[V_{i,G} = X_{r_1^i,G} + F \cdot (X_{r_2^i,G} - X_{r_3^i,G}).
\]

Crossover step: 按照如下方式生成\(U_{i,G}=(u_{1,i,G},\ldots, u_{D,i,G})\)

\[u_{j,i,G} =
\left \{ \begin{array}{ll}
v_{j,i,G} & if \: \mathrm{rand}[0,1] \le Cr \: or \: j=j_{rand} \\
x_{j,i,G} & otherwise.
\end{array} \right.
\]

Selection step:

\[X_{i,G} =
\left \{ \begin{array}{ll}
U_{i,G} & if \: f(U_{i,G}) \le f(X_{i,G}) \\
X_{i,G} & otherwise.
\end{array} \right.
\]

End For.
\(G=G+1\).

End While.

其中\(X_{i,G}=(x_{j,i,G}, \ldots, x_{D,i,G})\), \(j_{rand}\)是预先随机生成的一个属于\([1,D]\)的整数, 以保证\(U\)相对于\(X\)至少有些许变化产生, \(X_{r_1^i,G}, X_{r_2^i,G},X_{r_3^i,G}\)是从\(P_G\)中随机抽取且互异的.

在接下来我们可以发现很多变种, 而这些变种往往是Mutation step 和 Crossover step的变体.

DE/?/?/?

DE/rand/1/exp

这是crossover step步的的一个变种:

随机从\([1, D]\)中抽取整数\(n\)和\(L\), 然后

\[u_{j,i,G} =
\left \{ \begin{array}{ll}
v_{j,i,G} & if \: n \le j \le n+L-1\\
x_{j,i,G} & otherwise.
\end{array} \right.
\]

\(L\)可以通过下面的步骤生成

\(L=0\)
while \(\mathrm{rand}[0,1] \le Cr\) and \(L\le D\):

\[L=L+1.
\]

DE/best/1

\[V_{i,G}=X_{best,G} + F\cdot (X_{r_1^i,G} - X_{r_2^i,G}),
\]

其中\(X_{best,G}\)是\(P_{G}\)中的最优的点.

DE/best/2

\[V_{i,G}=X_{best,G} + F\cdot (X_{r_1^i,G} - X_{r_2^i,G}) + F\cdot (X_{r_3^i,G} - X_{r_4^i,G}).
\]

DE/rand/2

\[V_{i,G}=X_{r_1^i,G} + F\cdot (X_{r_2^i,G} - X_{r_3^i,G}) + F\cdot (X_{r_4^i,G} - X_{r_5^i,G}),
\]

超参数的选择

真的没有细看, 文中粗略地介绍了几处, 还有很多需要查原文.

\(F\)的选择

有的推荐\([0.4, 1]\)(最佳0.5), 有的推荐\(0.6\), 有的推荐\([0.4, 0.95]\)(最佳0.9).

还有一些自适应的选择, 如

\[F =
\left \{ \begin{array}{ll}
\max \{l_{\min}, 1- |\frac{f_{\max}}{f_{\min}}|\} & if : |\frac{f_{\max}}{f_{\min}}|<1 \\
\max \{l_{\min}, 1- |\frac{f_{\max}}{f_{\min}}|\} & otherwise,
\end{array} \right.
\]

我比较疑惑的是难道\(|\frac{f_{\max}}{f_{\min}}|\)不是大于等于1吗?

\[F_{i,G+1} =
\left \{ \begin{array}{ll}
\mathrm{rand}[F_l, F_{u}] & with \: probability \:\tau \\
F_{i,G} & else,
\end{array} \right.
\]

其中\(F_l\), \(F_u\)分别为\(F\)取值的下界和上界.

\(NP\)的选择

有的推荐\([5D,10D]\), 有的推荐\([3D, 8D]\).

\(Cr\)的选择

有的推荐\([0.3, 0.9]\).

还有

\[Cr_{i,G+1} =
\left \{ \begin{array}{ll}
\mathrm{rand}[0, 1] & with \: probability \:\tau \\
Cr_{i,G} & else,
\end{array} \right.
\]

一些连续变体

A

\[p=f(X_{r_1})+f(X_{r_2}) + f(X_{r_3}) \\
p_1 = f(X_{r_1})/p \\
p_2 = f(X_{r_2}) / p \\
p_3 = f(X_{r_1}) / p.
\]

如果\(\mathrm{rand}[0,1] < \Gamma\)(\(\Gamma\)是给定的):

\[\begin{array}{ll}
V_{i,G+1} = & (X_{r_1}+X_{r_2}+X_{r_3})/3 +(p_2-p_1)(X_{r_1}-X_{r_2}) \\
&+ (p_3-p_2)(X_{r_2} - X_{r_3}) + (p_1-p_3) (X_{r_3}- X_{r_1}),
\end{array}
\]

否则

\[V_{i,G+1} = X_{r_1} + F \cdot (X_{r_2}-X_{r_3}).
\]

B

\[U_{i,G}=X_{i, G}+k_i \cdot (X_{r_1,G}-X_{i,G})+F' \cdot (X_{r_2,G}-X_{r_3, G}),
\]

其中\(k_i\)给定, \(F'=k_i \cdot F\).

C

D

即在考虑\(x\)的时候, 还需要考虑其反\(a+b-x\), 假定\(x \in [a, b]\), \([a,b]\)为我们给定范围, \(X\)的反类似的构造.

E

其中\(X_{n_{best},G}\)表示在\(X_{i,G}\)的\(n\)的近邻中的最优点, \(p, q\in [i-k,i+k]\).

其中\(X_{g_{best},G}\)为\(P_G\)中的最优点.

\[V_{i,G}= w \cdot g_{i, G} + (1-w) \cdot L_{i, G}.
\]

G

剩下的在复杂环境下的应用就不记录了(只是单纯讲了该怎么做).

一些缺点

高维问题不易处理;
容易被一些问题欺骗, 而现如局部最优解;
对不能分解的函数效果不是很好;
路径往往不会太大(即探索不充分);
缺少收敛性的理论的保证.

代码

\(f(x,y)=x^2+50y^2\).

{

  "dim": 2,

  "F": 0.5,

  "NP": 5,

  "Cr": 0.35

}



"""

de.py

"""

import numpy as np

from scipy import stats

import random

class Parameter:

    def __init__(self, dim, xmin, xmax):

        self.dim = dim

        self.xmin = xmin

        self.xmax = xmax

        self.initial()

    def initial(self):

        self.para = stats.uniform.rvs(

            self.xmin, self.xmax - self.xmin

        )

    @property

    def data(self):

        return self.para

    def __getitem__(self, item):

        return self.para[item]

    def __setitem__(self, key, value):

        self.para[key] = value

    def __len__(self):

        return len(self.para)

    def __add__(self, other):

        return self.para + other

    def __mul__(self, other):

        return self.para * other

    def __pow__(self, power):

        return self.para ** power

    def __neg__(self):

        return -self.para

    def __sub__(self, other):

        return self.para - other

    def __truediv__(self, other):

        return self.para / other

class DE:

    def __init__(self, func, dim ,F=0.5, NP=50,

                 Cr=0.35, xmin=-10, xmax=10,

                 require_history=True):

        self.func = func

        self.dim = dim

        self.F = F

        self.NP = NP

        self.Cr = Cr

        self.xmin = np.array(xmin)

        self.xmax = np.array(xmax)

        assert all(self.xmin <= self.xmax), "Invalid xmin or xmax"

        self.require_history = require_history

        self.init_x()

        if self.require_history:

            self.build_history()

    def init_x(self):

        self.paras = [Parameter(self.dim, self.xmin, self.xmax)

                      for i in range(self.NP)]

    @property

    def data(self):

        return [para.data for para in self.paras]

    def build_history(self):

        self.paras_history = [self.data]

    def add_history(self):

        self.paras_history.append(self.data)

    def choose(self, size=3):

        return random.sample(self.paras, k=size)

    def mutation(self):

        x1, x2, x3 = self.choose(3)

        return x1 + self.F * (x2 - x3)

    def crossover(self, v, x):

        u = np.zeros_like(v)

        for i, _ in enumerate(v):

            jrand = random.randint(0, self.dim)

            if np.random.rand() < self.Cr or i is jrand:

                u[i] = v[i]

            else:

                u[i] = x[i]

            u[i] = v[i] if np.random.rand() < self.Cr else x[i]

        return u

    def selection(self, u, x):

        if self.func(u) < self.func(x):

            x.para = u

        else:

            pass

    def step(self):

        donors = [self.mutation()

                  for i in range(self.NP)]

        for i, donor in enumerate(donors):

            x = self.paras[i]

            u = self.crossover(donor, x)

            self.selection(u, x)

        if self.require_history:

            self.add_history()

    def multi_steps(self, times):

        for i in range(times):

            self.step()

class DEbest1(DE):

    def bestone(self):

        y = np.array([self.func(para)

             for para in self.paras])

        return self.paras[np.argmax(y)]

    def mutation(self, bestone):

        x1, x2 = self.choose(2)

        return bestone + self.F * (x1 - x2)

    def step(self):

        bestone = self.bestone()

        donors = [self.mutation(bestone)

                  for i in range(self.NP)]

        for i, donor in enumerate(donors):

            x = self.paras[i]

            u = self.crossover(donor, x)

            self.selection(u, x)

        if self.require_history:

            self.add_history()

class DEbest2(DEbest1):

    def mutation(self, bestone):

        x1, x2, x3, x4 = self.choose(4)

        return bestone + self.F * (x1 - x2) \

                + self.F * (x3 - x4)

class DErand2(DE):

    def mutation(self):

        x1, x2, x3, x4, x5 = self.choose(5)

        return x1 + self.F * (x2 - x3) \

                + self.F * (x4 - x5)

class DErandTM(DE):

    def mutation(self):

        x = self.choose(3)

        y = np.array(list(map(self.func, x)))

        p = y / y.sum()

        part1 = (x[0] + x[1] + x[2]) / 3

        part2 = (p[1] - p[0]) * (x[0] - x[1])

        part3 = (p[2] - p[1]) * (x[2] - x[1])

        part4 = (p[0] - p[2]) * (x[2] - x[0])

        return part1 + part2 + part3 + part4

巴特西