找到坐标x和y列表的优化位置

时间:2017-11-06 23:56:19

标签: python numpy optimization mathematical-optimization nonlinear-optimization

我是编程新手,特别是python,但我正在努力学习它,到目前为止我发现它非常迷人。

我有30个固定坐标x和y的列表。

x = np.array([[13,10,12,13,11,12,11,13,12,13,14,15,15,16,18,2,3,4,6,9,1,3,6,7,8,10,12,11,10,30]])
y = np.array([[12,11,10,9,8,7,6,6,7,8,11,12,13,15,14,18,12,11,10,13,15,16,18,17,16,15,14,13,12,3]])

我想找到一个优化的(集中式)位置,通过查找最小距离,最多可以连接10个固定坐标。

我厌倦了使用优化算法,但我只能得到一个坐标列表的最佳位置(即我不能将位置约束到10个坐标)。

任何帮助将不胜感激。

def objective(x):
    px = x[0]
    py = x[1]
    wdistance = sum(((px - wx)**2 + (py - wy)**2)**0.5)
    tdistance = (((px - tx)**2 + (py - ty)**2)**0.5)
    cost = wdistance * 2.5 + tdistance * 5 
    return cost

# initial estimate of the centralized location 
x0 = [10,10]

# boundary values for the centralized locations 
bnds = ((0,15), (0,15))
sol = minimize(objective, x0, bounds=bnds, method='SLSQP')

print(sol)

1 个答案:

答案 0 :(得分:2)

正如我的评论中所指出的,我认为这个问题是NP-hard,因此解决它并不容易。

这是某种基数约束的问题,这意味着,通常很难/不可能使用连续优化(假设几乎所有scipy.optimize中的所有内容)。

此外,最小化最小距离并不是特别顺利,几乎所有内容都在scipy中。但重新制定是可能的(至少在支持约束的情况下!)。

如果没有基数约束(这个问题的离散性的核心),它就会崩溃到Smallest Enclosing Circle Problem 问题中,很容易在线性时间内解决。更一般wiki-article。但这对于给定的问题没有多大帮助。

这是一个演示,可能不适合胆小的人(取决于背景知识):

  • 基于Mixed Integer Second Order Cone Programming

    • 将在无限时间内解决最优性(epsilon-approximation)(忽略数值问题; fp-math)
    • 没有超过NP硬度;但可能善于利用数据中的某种结构
    • 对于基于欧氏距离(标准)的问题感到自然
  • 使用cvxpy作为建模工具

  • 使用ECOS_BB作为解算器
    • 警告:这是一个概念验证/玩具解决方案(我们将在结果中看到)
    • MISOCP / MICP是一个活跃的研究领域,但非商业专业解决方案的数量很少!
      • 可能应该使用Pajarito,而不是ECOS_BB,可能使用Bonmin或商业广告(Gurobi,CPLEX,Mosek和co。)
      • Pajarito,看起来很有趣julia based和(imho)与python一起使用并不好玩;因此,请将我的代码视为演示解算器的演示!

以下代码将解决20个问题中的10个问题(前20个问题;为了更好的可视化,丢弃了一个复制品!)

对于完整的问题它将失败(将返回相当不错的近似值;但不是最佳的!)。我责备ECOS_BB(但同时非常感谢Han Wang的代码!)。通过更好的替代方案,您的问题肯定会很容易解决。但是不要指望任何解算器能够做到1000分: - )。

代码:

import numpy as np
import cvxpy as cvx
import matplotlib.pyplot as plt
from scipy.spatial.distance import pdist

""" (Reduced) Data """
# removed a duplicate point (for less strange visualization)
x = np.array([13,10,12,13,11,12,11,13,13,14,15,15,16,18,2,3,4,6,9,1])#,3,6,7,8,10,12,11,10,30])
y = np.array([12,11,10,9,8,7,6,6,8,11,12,13,15,14,18,12,11,10,13,15])#,16,18,17,16,15,14,13,12,3])
N = x.shape[0]
M = 10

""" Mixed-integer Second-order cone problem """
c = cvx.Variable(2)                                # center position to optimize
d = cvx.Variable(N)              # helper-var: lower-bound on distance to center
d_prod = cvx.Variable(N) # helper-var: lower-bound on distance * binary/selected
b = cvx.Bool(N)                          # binary-variables used for >= M points

dists = pdist(np.vstack((x,y)).T)
U = np.amax(dists)    # upper distance-bound (not tight!) for bigM-linearization

helper_expr_c_0 = cvx.vec(c[0] - x)
helper_expr_c_1 = cvx.vec(c[1] - y)

helper_expr_norm = cvx.norm(cvx.hstack(helper_expr_c_0, helper_expr_c_1), axis=1)

constraints = []
constraints.append(cvx.sum_entries(b) >= M)           # M or more points covered
constraints.append(d >= helper_expr_norm)             # lower-bound of distances
constraints.append(d_prod <= U * b)                   # linearization of product
constraints.append(d_prod >= 0)                       # """
constraints.append(d_prod <= d)                       # """
constraints.append(d_prod >= d - U * (1-b))           # """

objective = cvx.Minimize(cvx.max_entries(d_prod))

problem = cvx.Problem(objective, constraints)
problem.solve(verbose=False)
print(problem.status)

# Visualization

center = np.array(c.value.flat)
tmp = np.array(np.round(b.value.T.flat), dtype=int)
selected_points = np.where(tmp == 1)[0]
non_selected_points = np.where(tmp == 0)[0]

ax = plt.subplot(aspect='equal')
ax.set_title('Optimal solution radius: {0:.2f}'.format(problem.value))
ax.scatter(x[non_selected_points], y[non_selected_points], c='blue', s=70, alpha=0.9)
ax.scatter(x[selected_points], y[selected_points], c='magenta', s=70, alpha=0.9)
ax.scatter(c[0].value, c[1].value, c='red', s=70, alpha=0.9)
circ = plt.Circle((center[0], center[1]), radius=problem.value, color='g', fill=True, alpha=0.1)
ax.add_patch(circ)
plt.tight_layout()
plt.show()

输出:

optimal

简介:

enter image description here

修改

我将上面的代码移植到julia,以便能够使用上面提到的求解器Pajarito。尽管过去对我来说没有太多的julia编程,但它现在正在运作:

  • 使用Convex.jl作为建模工具
    • 上面基于cvxpy的代码的近1:1映射!
  • 使用3个开源求解器的组合:
    • Pajarito作为MISOCP / MICP求解器
      • 使用GLPK作为内部MIP解算器(硬币CBC通常要好得多;但由于Pajarito中可能存在错误而无法为我工作)
      • 使用ECOS作为内部Convex-solver
  • 重复使用上面基于matplotlib的绘图代码;手工复制的结果值: - )

这种方法正在解决您的完整问题(我再次删除了重复点)以达到最佳状态!

它输出一些数值不稳定警告,但声称结果是最佳的! (Pajarito仍然是相当新的研究软件,我没有调整众多选项;我们使用的开源解算器与商业替代品相比并不那么强大)

我对结果非常满意,我很高兴我试试了Julia / Convex.jl / Pajarito!

代码:

using Convex, Pajarito, ECOS, GLPKMathProgInterface

# DATA
x = [13 10 12 13 11 12 11 13 13 14 15 15 16 18 2 3 4 6 9 1 3 6 7 8 10 12 11 10 30]
y = [12 11 10 9 8 7 6 6 8 11 12 13 15 14 18 12 11 10 13 15 16 18 17 16 15 14 13 12 3]
N = size(x)[2]
M = 10

# MISOCP
c = Variable(2)
d = Variable(N)
d_prod = Variable(N)
b = Variable(N, :Bin)

U = 100  # TODO

# Objective
problem = minimize(maximum(d_prod))

# Constraints
problem.constraints += sum(b) >= M
problem.constraints += d_prod <= U*b
problem.constraints += d_prod >= 0
problem.constraints += d_prod <= d
problem.constraints += d_prod >= d - U * (1-b)
for i = 1:N
    problem.constraints += d[i] >= norm([c[1] - x[i], c[2] - y[i]])  # ugly
end

# Solve
mip_solver_drives = false
log_level = 2
rel_gap = 1e-8
mip_solver = GLPKSolverMIP()
cont_solver = ECOSSolver(verbose=false)
solver = PajaritoSolver(
    mip_solver_drives=mip_solver_drives,
    log_level=log_level,
    rel_gap=rel_gap,
    mip_solver=mip_solver,
    cont_solver=cont_solver,
    init_exp=true,
)
# option taken from https://github.com/JuliaOpt/Pajarito.jl/blob/master/examples/gatesizing.jl
# not necessarily optimal

solve!(problem, solver)

# Print out results
println(problem.status)
println(problem.optval)
println(b.value)
println(c.value)

输出:

Problem dimensions:
       variables |       120
     constraints |       263
   nonzeros in A |       466

Cones summary:
Cone             | Count     | Min dim.  | Max dim.
    Second order |        29 |         3 |         3

Variable types:
      continuous |        91
          binary |        29

Transforming data...               1.10s

Creating conic subproblem...       0.52s

Building MIP model...              2.35s

Solving conic relaxation...        1.38s

Starting iterative algorithm
    0 |           +Inf |  +3.174473e-11 |         Inf |   6.434e+00
Warning: numerical instability (primal simplex, phase I)

Iter. | Best feasible  | Best bound     | Rel. gap    | Time (s)
    1 |  +2.713537e+00 |  +2.648653e+00 |   2.391e-02 |   1.192e+01
Warning: numerical instability (primal simplex, phase I)
Warning: numerical instability (primal simplex, phase I)
Warning: numerical instability (dual simplex, phase II)
... some more ...
    2 |  +2.713537e+00 |  +2.713537e+00 |   5.134e-12 |   1.341e+01

Iterative algorithm summary:
 - Status               =        Optimal
 - Best feasible        =  +2.713537e+00
 - Best bound           =  +2.713537e+00
 - Relative opt. gap    =      5.134e-12
 - Total time (s)       =       1.34e+01

Outer-approximation cuts added:
Cone             | Relax.    | Violated  | Nonviol.
    Second order |        58 |        32 |        24

0 numerically unstable cone duals encountered

Distance to feasibility (negative indicates strict feasibility):
Cone             | Variable  | Constraint
          Linear |        NA |  5.26e-13
    Second order |        NA |  2.20e-11

Distance to integrality of integer/binary variables:
          binary |  0.00e+00

Optimal
2.7135366682753825
[1.0; 1.0; 1.0; 1.0; 0.0; 0.0; 0.0; 0.0; 0.0; 1.0; 1.0; 1.0; 0.0; 0.0; 0.0; 0.0; 0.0; 0.0; 0.0; 0.0; 0.0; 0.0; 0.0; 0.0; 0.0; 1.0; 1.0; 1.0; 0.0]
[12.625; 11.6875]

可视化:

enter image description here

还有一个选择23的情节:

enter image description here