Multi-objective 优化示例 Pyomo
Multi-objective optimization example Pyomo
Pyomo 中的多 objective 优化的任何示例?
我正在尝试最小化 4 个目标(非线性)并且我想使用 pyomo 和 ipopt。也可以访问 Gurobi。
我想看一个非常简单的例子,我们尝试针对决策变量列表(不仅仅是一维,也可能是向量)优化两个或多个 objective(一个最小化和一个最大化) .
我拥有的 Pyomo 书 (https://link.springer.com/content/pdf/10.1007%2F978-3-319-58821-6.pdf) 没有提供任何线索。
据我所知,虽然 Pyomo 支持多 objective 模型的表达,但它还没有自动模型转换来为您生成常见的多 objective 优化公式。
也就是说,您仍然可以自己创建这些公式。查看 epsilon-constraint、1-norm 和 infinity norm 以获得一些想法。
使用 Pyomo,您必须自己实施。我现在正在做。最好的方法是增强的 epsilon 约束方法。它总是高效的,并且总能找到全局帕累托最优。最好的例子在这里:
Effective implementation of the epsilon-constraint method in Multi-Objective Mathematical Programming problems, Mavrotas, G, 2009
.
编辑:我在 pyomo 中编写了上述论文中的示例:
它将首先最大化 f1,然后最大化 f2。然后它将应用正常的 epsilon 约束并绘制低效的 Pareto 前沿,然后它将应用增强的 epsilon 约束,这最终是要采用的方法!
from pyomo.environ import *
import matplotlib.pyplot as plt
# max f1 = X1 <br>
# max f2 = 3 X1 + 4 X2 <br>
# st X1 <= 20 <br>
# X2 <= 40 <br>
# 5 X1 + 4 X2 <= 200 <br>
model = ConcreteModel()
model.X1 = Var(within=NonNegativeReals)
model.X2 = Var(within=NonNegativeReals)
model.C1 = Constraint(expr = model.X1 <= 20)
model.C2 = Constraint(expr = model.X2 <= 40)
model.C3 = Constraint(expr = 5 * model.X1 + 4 * model.X2 <= 200)
model.f1 = Var()
model.f2 = Var()
model.C_f1 = Constraint(expr= model.f1 == model.X1)
model.C_f2 = Constraint(expr= model.f2 == 3 * model.X1 + 4 * model.X2)
model.O_f1 = Objective(expr= model.f1 , sense=maximize)
model.O_f2 = Objective(expr= model.f2 , sense=maximize)
model.O_f2.deactivate()
solver = SolverFactory('cplex')
solver.solve(model);
print( '( X1 , X2 ) = ( ' + str(value(model.X1)) + ' , ' + str(value(model.X2)) + ' )')
print( 'f1 = ' + str(value(model.f1)) )
print( 'f2 = ' + str(value(model.f2)) )
f2_min = value(model.f2)
# ## max f2
model.O_f2.activate()
model.O_f1.deactivate()
solver = SolverFactory('cplex')
solver.solve(model);
print( '( X1 , X2 ) = ( ' + str(value(model.X1)) + ' , ' + str(value(model.X2)) + ' )')
print( 'f1 = ' + str(value(model.f1)) )
print( 'f2 = ' + str(value(model.f2)) )
f2_max = value(model.f2)
# ## apply normal $\epsilon$-Constraint
model.O_f1.activate()
model.O_f2.deactivate()
model.e = Param(initialize=0, mutable=True)
model.C_epsilon = Constraint(expr = model.f2 == model.e)
solver.solve(model);
print('Each iteration will keep f2 lower than some values between f2_min and f2_max, so [' + str(f2_min) + ', ' + str(f2_max) + ']')
n = 4
step = int((f2_max - f2_min) / n)
steps = list(range(int(f2_min),int(f2_max),step)) + [f2_max]
x1_l = []
x2_l = []
for i in steps:
model.e = i
solver.solve(model);
x1_l.append(value(model.X1))
x2_l.append(value(model.X2))
plt.plot(x1_l,x2_l,'o-.');
plt.title('inefficient Pareto-front');
plt.grid(True);
# ## apply augmented $\epsilon$-Constraint
# max f2 + delta*epsilon <br>
# s.t. f2 - s = e
model.del_component(model.O_f1)
model.del_component(model.O_f2)
model.del_component(model.C_epsilon)
model.delta = Param(initialize=0.00001)
model.s = Var(within=NonNegativeReals)
model.O_f1 = Objective(expr = model.f1 + model.delta * model.s, sense=maximize)
model.C_e = Constraint(expr = model.f2 - model.s == model.e)
x1_l = []
x2_l = []
for i in range(160,190,6):
model.e = i
solver.solve(model);
x1_l.append(value(model.X1))
x2_l.append(value(model.X2))
plt.plot(x1_l,x2_l,'o-.');
plt.title('efficient Pareto-front');
plt.grid(True);
免责声明:我是 pymoo 的主要开发人员,Python 中的多 objective 优化框架。
您可能需要考虑 Python 中专注于多 objective 优化的其他框架。例如,在 pymoo 中,上面提到的相当简单的测试问题的定义或多或少是直截了当的。您可以在下面找到它的实现。设计和 objectives space 中的结果如下所示:
pymoo 有据可查,并提供入门指南,演示如何定义您自己的优化问题、获得一组接近最优的解决方案并对其进行分析:https://pymoo.org/getting_started.html
该框架的重点是与多 objective 优化相关的任何事情,包括可视化和决策制定。
import matplotlib.pyplot as plt
import numpy as np
from pymoo.algorithms.nsga2 import NSGA2
from pymoo.model.problem import Problem
from pymoo.optimize import minimize
from pymoo.visualization.scatter import Scatter
class MyProblem(Problem):
def __init__(self):
"""
max f1 = X1 <br>
max f2 = 3 X1 + 4 X2 <br>
st X1 <= 20 <br>
X2 <= 40 <br>
5 X1 + 4 X2 <= 200 <br>
"""
super().__init__(n_var=2,
n_obj=2,
n_constr=1,
xl=np.array([0, 0]),
xu=np.array([20, 40]))
def _evaluate(self, x, out, *args, **kwargs):
# define both objectives
f1 = x[:, 0]
f2 = 3 * x[:, 0] + 4 * x[:, 1]
# we have to negate the objectives because by default we assume minimization
f1, f2 = -f1, -f2
# define the constraint as a less or equal to zero constraint
g1 = 5 * x[:, 0] + 4 * x[:, 1] - 200
out["F"] = np.column_stack([f1, f2])
out["G"] = g1
problem = MyProblem()
algorithm = NSGA2()
res = minimize(problem,
algorithm,
('n_gen', 200),
seed=1,
verbose=True)
print(res.X)
print(res.F)
fig, (ax1, ax2) = plt.subplots(nrows=1, ncols=2, figsize=(12, 6))
Scatter(fig=fig, ax=ax1, title="Design Space").add(res.X, color="blue").do()
Scatter(fig=fig, ax=ax2, title="Objective Space").add(res.F, color="red").do()
plt.show()
Pyomo 中的多 objective 优化的任何示例?
我正在尝试最小化 4 个目标(非线性)并且我想使用 pyomo 和 ipopt。也可以访问 Gurobi。
我想看一个非常简单的例子,我们尝试针对决策变量列表(不仅仅是一维,也可能是向量)优化两个或多个 objective(一个最小化和一个最大化) .
我拥有的 Pyomo 书 (https://link.springer.com/content/pdf/10.1007%2F978-3-319-58821-6.pdf) 没有提供任何线索。
据我所知,虽然 Pyomo 支持多 objective 模型的表达,但它还没有自动模型转换来为您生成常见的多 objective 优化公式。
也就是说,您仍然可以自己创建这些公式。查看 epsilon-constraint、1-norm 和 infinity norm 以获得一些想法。
使用 Pyomo,您必须自己实施。我现在正在做。最好的方法是增强的 epsilon 约束方法。它总是高效的,并且总能找到全局帕累托最优。最好的例子在这里:
Effective implementation of the epsilon-constraint method in Multi-Objective Mathematical Programming problems, Mavrotas, G, 2009
.
编辑:我在 pyomo 中编写了上述论文中的示例: 它将首先最大化 f1,然后最大化 f2。然后它将应用正常的 epsilon 约束并绘制低效的 Pareto 前沿,然后它将应用增强的 epsilon 约束,这最终是要采用的方法!
from pyomo.environ import *
import matplotlib.pyplot as plt
# max f1 = X1 <br>
# max f2 = 3 X1 + 4 X2 <br>
# st X1 <= 20 <br>
# X2 <= 40 <br>
# 5 X1 + 4 X2 <= 200 <br>
model = ConcreteModel()
model.X1 = Var(within=NonNegativeReals)
model.X2 = Var(within=NonNegativeReals)
model.C1 = Constraint(expr = model.X1 <= 20)
model.C2 = Constraint(expr = model.X2 <= 40)
model.C3 = Constraint(expr = 5 * model.X1 + 4 * model.X2 <= 200)
model.f1 = Var()
model.f2 = Var()
model.C_f1 = Constraint(expr= model.f1 == model.X1)
model.C_f2 = Constraint(expr= model.f2 == 3 * model.X1 + 4 * model.X2)
model.O_f1 = Objective(expr= model.f1 , sense=maximize)
model.O_f2 = Objective(expr= model.f2 , sense=maximize)
model.O_f2.deactivate()
solver = SolverFactory('cplex')
solver.solve(model);
print( '( X1 , X2 ) = ( ' + str(value(model.X1)) + ' , ' + str(value(model.X2)) + ' )')
print( 'f1 = ' + str(value(model.f1)) )
print( 'f2 = ' + str(value(model.f2)) )
f2_min = value(model.f2)
# ## max f2
model.O_f2.activate()
model.O_f1.deactivate()
solver = SolverFactory('cplex')
solver.solve(model);
print( '( X1 , X2 ) = ( ' + str(value(model.X1)) + ' , ' + str(value(model.X2)) + ' )')
print( 'f1 = ' + str(value(model.f1)) )
print( 'f2 = ' + str(value(model.f2)) )
f2_max = value(model.f2)
# ## apply normal $\epsilon$-Constraint
model.O_f1.activate()
model.O_f2.deactivate()
model.e = Param(initialize=0, mutable=True)
model.C_epsilon = Constraint(expr = model.f2 == model.e)
solver.solve(model);
print('Each iteration will keep f2 lower than some values between f2_min and f2_max, so [' + str(f2_min) + ', ' + str(f2_max) + ']')
n = 4
step = int((f2_max - f2_min) / n)
steps = list(range(int(f2_min),int(f2_max),step)) + [f2_max]
x1_l = []
x2_l = []
for i in steps:
model.e = i
solver.solve(model);
x1_l.append(value(model.X1))
x2_l.append(value(model.X2))
plt.plot(x1_l,x2_l,'o-.');
plt.title('inefficient Pareto-front');
plt.grid(True);
# ## apply augmented $\epsilon$-Constraint
# max f2 + delta*epsilon <br>
# s.t. f2 - s = e
model.del_component(model.O_f1)
model.del_component(model.O_f2)
model.del_component(model.C_epsilon)
model.delta = Param(initialize=0.00001)
model.s = Var(within=NonNegativeReals)
model.O_f1 = Objective(expr = model.f1 + model.delta * model.s, sense=maximize)
model.C_e = Constraint(expr = model.f2 - model.s == model.e)
x1_l = []
x2_l = []
for i in range(160,190,6):
model.e = i
solver.solve(model);
x1_l.append(value(model.X1))
x2_l.append(value(model.X2))
plt.plot(x1_l,x2_l,'o-.');
plt.title('efficient Pareto-front');
plt.grid(True);
免责声明:我是 pymoo 的主要开发人员,Python 中的多 objective 优化框架。
您可能需要考虑 Python 中专注于多 objective 优化的其他框架。例如,在 pymoo 中,上面提到的相当简单的测试问题的定义或多或少是直截了当的。您可以在下面找到它的实现。设计和 objectives space 中的结果如下所示:
pymoo 有据可查,并提供入门指南,演示如何定义您自己的优化问题、获得一组接近最优的解决方案并对其进行分析:https://pymoo.org/getting_started.html
该框架的重点是与多 objective 优化相关的任何事情,包括可视化和决策制定。
import matplotlib.pyplot as plt
import numpy as np
from pymoo.algorithms.nsga2 import NSGA2
from pymoo.model.problem import Problem
from pymoo.optimize import minimize
from pymoo.visualization.scatter import Scatter
class MyProblem(Problem):
def __init__(self):
"""
max f1 = X1 <br>
max f2 = 3 X1 + 4 X2 <br>
st X1 <= 20 <br>
X2 <= 40 <br>
5 X1 + 4 X2 <= 200 <br>
"""
super().__init__(n_var=2,
n_obj=2,
n_constr=1,
xl=np.array([0, 0]),
xu=np.array([20, 40]))
def _evaluate(self, x, out, *args, **kwargs):
# define both objectives
f1 = x[:, 0]
f2 = 3 * x[:, 0] + 4 * x[:, 1]
# we have to negate the objectives because by default we assume minimization
f1, f2 = -f1, -f2
# define the constraint as a less or equal to zero constraint
g1 = 5 * x[:, 0] + 4 * x[:, 1] - 200
out["F"] = np.column_stack([f1, f2])
out["G"] = g1
problem = MyProblem()
algorithm = NSGA2()
res = minimize(problem,
algorithm,
('n_gen', 200),
seed=1,
verbose=True)
print(res.X)
print(res.F)
fig, (ax1, ax2) = plt.subplots(nrows=1, ncols=2, figsize=(12, 6))
Scatter(fig=fig, ax=ax1, title="Design Space").add(res.X, color="blue").do()
Scatter(fig=fig, ax=ax2, title="Objective Space").add(res.F, color="red").do()
plt.show()