[论文复现]演化博弈方法用于多智能体系统最优资源分配

原文 演化博弈方法用于多智能体系统最优资源分配 -CSDN博客
https://ieeexplore.ieee.org/document/8243778/

问题描述

  有2种资源分配给6个个体，2种资源的总量分别为 $y_1=545,y_2=467$，不同的个体收到不同的资源会产生不同的收益，分配的资源量均大于等于0。以第3个个体为例，它接受的资源为 ${\vec{x}}_3=[x_{31},x_{32}{]}^{\text{T}}$，两种资源的上限分别为 $x_{31}^{\text{max}}=66,x_{32}^{\text{max}}=120$。可行资源组合 ${\mathcal{R}}_3={\vec{x}}_3$（这我没看懂 ${\mathcal{R}}_3$ 有什么用），收到2种资源的收益为
$$\begin{array}{rl}{u}_3({\vec{x}}_3)=& \frac{x_{31}(2x_{31}^{\text{max}}-x_{31})}{c_{31}{x}_{31}^{\text{max}}}+\frac{x_{32}(2x_{32}^{\text{max}}-x_{32})}{c_{32}{x}_{32}^{\text{max}}}\\ =& \frac{x_{31}(2\cdot 66-x_{31})}{0.4\cdot 66}+\frac{x_{32}(2\cdot 120-x_{32})}{0.5\cdot 120}\end{array}$$
需要求的是在
$$\begin{gathered} x_{11}+x_{21}+x_{31}+x_{41}+x_{51}+x_{61}=545 \\ {x}_{12}+x_{22}+x_{32}+x_{42}+x_{52}+x_{62}=467 \end{gathered}$$
的约束条件下，求
$$u_1+u_2+u_3+u_4+u_5+u_6$$
的最大值。

一个证明的解读

原文中求和不变性的证明式(16)中
$$\begin{array}{rl}{\dot{x}}_{31}=& x_{31}\left(p_{31}\frac{x_{21}+x_{41}}{y_1}-\frac{x_{21}{p}_{21}+x_{41}{p}_{41}}{y_1}\right)\\ p_{31}=& \frac{\partial u_3}{\partial x_{31}}=\frac{2(x_{31}^{\text{max}}-x_{31})}{c_{31}{x}_{31}^{\text{max}}}\end{array}$$
假设对个体3和它的邻居2，包含这两个个体的式子为
$$\begin{gathered} {\dot{x}}_{31}=x_{31}\left(p_{31}\frac{x_{21}+x_{41}}{y_1}-\frac{x_{21}{p}_{21}+x_{41}{p}_{41}}{y_1}\right) \\ {\dot{x}}_{21}=x_{21}\left(p_{21}\frac{x_{11}+x_{31}}{y_1}-\frac{x_{11}{p}_{11}+x_{31}{p}_{31}}{y_1}\right) \end{gathered}$$
取出所有包含 $x_{21}$ 和 $x_{31}$ 的项就是式(16)中
$$(x_{31}{p}_{31}{x}_{21}-x_{31}{x}_{21}{p}_{21})+(x_{21}{p}_{21}{x}_{31}-x_{21}{x}_{31}{p}_{31})=0$$

与拉格朗日乘子法比较

  原论文中给出的例子的收益函数里两个资源的收益相互独立，因此本文只分析第一种资源的收益。除了原文介绍的演化博弈法以外，注意到这个问题就是一个约束优化问题。可以试一下拉格朗日乘子法。为便于说明只写3个变量，代码里完整给出。
$$U=\frac{x_1(2x_{m1}-x_1)}{c_1{x}_{m1}}+\frac{x_2(2x_{m2}-x_2)}{c_2{x}_{m2}}+\frac{x_3(2x_{m3}-x_3)}{c_3{x}_{m3}}$$
$$\begin{gathered} \frac{\partial L}{\partial x_i}=\frac{x_{mi}-x_i}{k_i}+\lambda =0 \\ {x}_i-k_i\lambda =x_{mi} \end{gathered}$$
其中 $k_i=\frac{c_i{x}_{mi}}{2}$。写成矩阵形式
$$\left[\begin{array}{cccc}1& & & k_1\\ & 1& & k_2\\ & & 1& k_3\\ 1& 1& 1& 0\end{array}\right]\left[\begin{array}{c}{x}_1\\ x_2\\ x_3\\ \lambda \end{array}\right]=\left[\begin{array}{c}{x}_{m1}\\ x_{m2}\\ x_{m3}\\ 545\end{array}\right]$$
解得
$$X=\left[\begin{array}{c}167.604914004914\\ 45.1985257985260\\ 62.6270270270270\\ 104.645700245700\\ 93.6117936117936\\ 71.3120393120392\\ 0.255528255528253\end{array}\right]$$
演化博弈法的仿真结果为
$$X=\left[\begin{array}{c}167.75139\\ 45.21687\\ 62.42745\\ 104.58382\\ 93.432434\\ 71.588005\end{array}\right]$$
可见结果基本相同。

仿真

计算出的最大值是 $3615.688$，论文里没给出最大值的具体值但结果差不多。

参考链接

1. 演化博弈方法用于多智能体系统最优资源分配 -CSDN博客
2. https://ieeexplore.ieee.org/document/8243778/
3. 演化博弈、复制动态方程与仿真 -CSDN博客

附代码

拉格朗日乘子法的matlab代码：

clear;clc;
Cij = [0.2;0.3;0.4;0.1;0.5;0.85];
Xmaxij = [172;47;66;106;100;80];
K = zeros(6,1);
for n = 1:6
    K(n) = Cij(n)*Xmaxij(n)*0.5;
end
A = [eye(6), K; ones(1, 6), 0];
B = [Xmaxij; 545];
X = pinv(A)*B

演化博弈法的python代码：

import matplotlib.pyplot as plt
import numpy as np

def Marginal_Utility(i, j):
    return 2*(Xmaxij[i, j] - Xij[i, j]) / (Cij[i, j]*Xmaxij[i, j])
def StateSum():
    sum = 0
    for n in range(Xij.shape[0]):
        sum += Xij[n, 0]
    for n in range(Xij.shape[0]):
        sum += Xij[n, 1]
    return sum
def Utility():
    u = 0
    for j in range(2):
        for i in range(6):
            u += Xij[i, j]*(2*Xmaxij[i, j] - Xij[i, j]) / Cij[i, j] / Xmaxij[i, j]
    return u

Cij = np.array([
    [0.2 , 0.85],
    [0.3 , 0.4 ],
    [0.4 , 0.5 ],
    [0.1 , 0.2 ],
    [0.5 , 0.4 ],
    [0.85, 0.8 ],
], dtype=np.float32)
Xmaxij = np.array([
    [172, 86 ],
    [47 , 70 ],
    [66 , 120],
    [106, 45 ],
    [100, 90 ],
    [80 , 100],
], dtype=np.float32)
Xij = np.array([
    [10, 10],
    [10, 10],
    [10, 10],
    [10, 10],
    [10, 10],
    [10, 10],
], dtype=np.float32)
Y = np.array([545, 467])
sum = 0
for n in range(Xij.shape[0] - 1):
    sum += Xij[n, 0]
Xij[-1, 0] = Y[0] - sum
sum = 0
for n in range(Xij.shape[0] - 1):
    sum += Xij[n, 1]
Xij[-1, 1] = Y[1] - sum
Xnew = Xij.copy()
nvec = []
Uvec = []
Xvec = []
for i in range(6):
    Xvec.append([])
for n in range(100):
    for j in range(2):
        for i in range(6):
            Ineigh0 = (i+5) % 6
            Ineigh1 = (i+1) % 6
            f = Marginal_Utility(i, j) * (Xij[Ineigh0, j] + Xij[Ineigh1, j])
            fbar = Xij[Ineigh0, j] * Marginal_Utility(Ineigh0, j)
            fbar += Xij[Ineigh1, j] * Marginal_Utility(Ineigh1, j)
            deltax = Xij[i, j] * (f - fbar) / Y[j]
            Xnew[i, j] += deltax*0.1
    Xij = Xnew.copy()
    for i in range(6):
        Xvec[i].append([Xij[i, 0]])
    # print(StateSum())
    nvec.append(n)
    Uvec.append(Utility())
print(Uvec[-1])
# plt.plot(nvec, Uvec)
for i in range(6):
    plt.plot(nvec, Xvec[i])
plt.legend([str(i) for i in range(6)])
plt.show()