multi-level - 如何在 OpenMDAO 1.x 中使用嵌套问题？

Question

我正在尝试在 OpenMDAO 上实现协作优化和其他多层次架构。我读到 here，这可以通过在问题的子类中定义一个单独的 solve_nonlinear 方法来完成。

问题是在运行问题实例时，未调用定义的 solve_linear。这是代码-

from __future__ import print_function, division
import numpy as np
import time

from openmdao.api import Component,Group, IndepVarComp, ExecComp,\
    Problem, ScipyOptimizer, NLGaussSeidel, ScipyGMRES


class SellarDis1(Component):
    """Component containing Discipline 1."""

    def __init__(self):
        super(SellarDis1, self).__init__()

        self.add_param('z', val=np.zeros(2))
        self.add_param('x', val=0.0)
        self.add_param('y2', val=1.0)

        self.add_output('y1', val=1.0)

    def solve_nonlinear(self, params, unknowns, resids):
        y1 = z1**2 + z2 + x1 - 0.2*y2"""

        z1 = params['z'][0]
        z2 = params['z'][1]
        x1 = params['x']
        y2 = params['y2']

        unknowns['y1'] = z1**2 + z2 + x1 - 0.2*y2

    def linearize(self, params, unknowns, resids):
        J = {}

        J['y1','y2'] = -0.2
        J['y1','z'] = np.array([[2*params['z'][0], 1.0]])
        J['y1','x'] = 1.0

        return J

class SellarDis2(Component):

    def __init__(self):
        super(SellarDis2, self).__init__()

        self.add_param('z', val=np.zeros(2))
        self.add_param('y1', val=1.0)

        self.add_output('y2', val=1.0)

    def solve_nonlinear(self, params, unknowns, resids):

        z1 = params['z'][0]
        z2 = params['z'][1]
        y1 = params['y1']
        y1 = abs(y1)

        unknowns['y2'] = y1**.5 + z1 + z2

    def linearize(self, params, unknowns, resids):
        J = {}

        J['y2', 'y1'] = 0.5*params['y1']**-0.5
        J['y2', 'z'] = np.array([[1.0, 1.0]])

        return J

class Sellar(Group):

    def __init__(self):
        super(Sellar, self).__init__()

        self.add('px', IndepVarComp('x', 1.0), promotes=['*'])
        self.add('pz', IndepVarComp('z', np.array([5.0,2.0])), promotes=['*'])

        self.add('d1', SellarDis1(), promotes=['*'])
        self.add('d2', SellarDis2(), promotes=['*'])

        self.add('obj_cmp', ExecComp('obj = x**2 + z[1] + y1 + exp(-y2)',
                                     z=np.array([0.0, 0.0]), x=0.0, y1=0.0, y2=0.0),
                 promotes=['*'])

        self.add('con_cmp1', ExecComp('con1 = 3.16 - y1'), promotes=['*'])
        self.add('con_cmp2', ExecComp('con2 = y2 - 24.0'), promotes=['*'])

        self.nl_solver = NLGaussSeidel()
        self.nl_solver.options['atol'] = 1.0e-12

        self.ln_solver = ScipyGMRES()

    def solve_nonlinear(self, params=None, unknowns=None, resids=None, metadata=None):

        print("Group's solve_nonlinear was called!!")
        # Discipline Optimizer would be called here?
        super(Sellar, self).solve_nonlinear(params, unknowns, resids)


class ModifiedProblem(Problem):

    def solve_nonlinear(self, params, unknowns, resids):

        print("Problem's solve_nonlinear was called!!")
        # or here ?
        super(ModifiedProblem, self).solve_nonlinear()


top = ModifiedProblem()
top.root = Sellar()

top.driver = ScipyOptimizer()
top.driver.options['optimizer'] = 'SLSQP'

top.driver.add_desvar('z', lower=np.array([-10.0, 0.0]),
                     upper=np.array([10.0, 10.0]))
top.driver.add_desvar('x', lower=0., upper=10.0)
top.driver.add_objective('obj')
top.driver.add_constraint('con1', upper=0.0)
top.driver.add_constraint('con2', upper=0.0)


top.setup(check=False)
top.run()

以上代码的输出是-

Group's solve_nonlinear was called!!
Group's solve_nonlinear was called!!
Group's solve_nonlinear was called!!
Group's solve_nonlinear was called!!
Group's solve_nonlinear was called!!
Group's solve_nonlinear was called!!
Group's solve_nonlinear was called!!
Optimization terminated successfully.    (Exit mode 0)
            Current function value: [ 3.18339395]
            Iterations: 6
            Function evaluations: 6
            Gradient evaluations: 6
Optimization Complete
-----------------------------------

这意味着任何时候都不会调用 Problem 子类中定义的 solve_nonlinear。那么，我应该在 Group 的子类中调用学科优化器吗？

此外，我如何在两个优化问题(系统和学科)之间传递目标变量，特别是将各个学科的优化全局变量返回给系统优化器。

感谢大家。

Answer 1

你是对的，Problem 上的 solve_nonlinear 永远不会被调用，因为 Problem 不是 OpenMDAO 组件并且没有 solve_nonlinear 方法。为了在另一个问题中运行子模型问题，您要做的是将其封装在一个组件实例中。它看起来像这样:

class SubOptimization(Component)

    def __init__(self):
        super(SubOptimization, self).__init__()

        # Inputs to this subprob
        self.add_param('z', val=np.zeros(2))
        self.add_param('x', val=0.0)
        self.add_param('y2', val=1.0)

        # Unknowns for this sub prob
        self.add_output('y1', val=1.0)

        self.problem = prob = Problem()
        prob.root = Group()
        prob.add('px', IndepVarComp('x', 1.0), promotes=['*'])
        prob.add('d1', SellarDis1(), promotes=['*'])

        # TODO - add cons/objs for sub prob

        prob.driver = ScipyOptimizer()
        prob.driver.options['optimizer'] = 'SLSQP'

        prob.driver.add_desvar('x', lower=0., upper=10.0)
        prob.driver.add_objective('obj')
        prob.driver.add_constraint('con1', upper=0.0)
        prob.driver.add_constraint('con2', upper=0.0)

        prob.setup()

        # Must finite difference across optimizer
        self.fd_options['force_fd'] = True

    def solve_nonlinear(self, params, unknowns, resids):

        prob = self.problem

        # Pass values into our problem
        prob['x'] = params['x']
        prob['z'] = params['z']
        prob['y2'] = params['y2']

        # Run problem
        prob.run()

        # Pull values from problem
        unknowns['y1'] = prob['y1']

您可以将此组件放入您的主要问题中(连同一个用于规程 2 的组件，尽管 2 实际上不需要次级优化，因为它没有局部设计变量)并围绕它优化全局设计变量。

一个警告:这不是我尝试过的(我也没有测试上面不完整的代码片段)，但它应该能让您走上正轨。您可能会遇到一个错误，因为这并没有真正经过太多测试。当我有时间的时候，我会为 OpenMDAO 测试整理一个像这样的 CO 测试，这样我们就安全了。