Scalable study

We want to compare IDF and MDF formulations with respect to the problem dimension for the aerostructure problem. For that, we use the ScalabilityStudy and PostScalabilityStudy classes.

from __future__ import division, unicode_literals

from gemseo.api import configure_logger, create_discipline, create_scenario
from gemseo.problems.aerostructure.aerostructure_design_space import (
from gemseo.problems.scalable.data_driven.api import (


Create the disciplinary datasets

First of all, we create the disciplinary AbstractFullCache datasets based on a DiagonalDOE.

datasets = {}
disciplines = create_discipline(["Aerodynamics", "Structure", "Mission"])
for discipline in disciplines:
    design_space = AerostructureDesignSpace()
    output = next(iter(discipline.get_output_data_names()))
    scenario = create_scenario(
        discipline, "DisciplinaryOpt", output, design_space, scenario_type="DOE"
    scenario.execute({"algo": "DiagonalDOE", "n_samples": 10})
    datasets[] = discipline.cache.export_to_dataset()

Define the design problem

Then, we instantiate a ScalabilityStudy from the definition of the design problem, expressed in terms of objective function (to maximize or minimize), design variables (local and global) and constraints (equality and inequality). We can also specify the coupling variables that we could scale. Note that this information is only required by the scaling stage. Indeed, MDO formulations know perfectly how to automatically recognize the coupling variables. Lastly, we can specify some properties of the scalable methodology such as the fill factor describing the level of dependence between inputs and outputs.

study = create_scalability_study(
    design_variables=["thick_airfoils", "thick_panels", "sweep"],
    coupling_variables=["forces", "displ"],

Add the disciplinary datasets


Add the optimization strategies

Then, we define the different optimization strategies we want to compare: In this case, the strategies are:

  • MDF formulation with the "NLOPT_SLSQP" optimization algorithm and no more than 100 iterations,

  • IDF formulation with the "NLOPT_SLSQP" optimization algorithm and no more than 100 iterations,

Note that in this case, we compare MDO formulations but we could easily compare optimization algorithms.

study.add_optimization_strategy("NLOPT_SLSQP", 100, "MDF")
study.add_optimization_strategy("NLOPT_SLSQP", 100, "IDF")

Add the scaling strategy

After that, we define the different scaling strategies for which we want to compare the optimization strategies. In this case, the strategies are:

  1. All design parameters have a size equal to 1,

  2. All design parameters have a size equal to 10,

  3. All design parameters have a size equal to 20.

To do that, we pass design_size=[1, 10, 20] to the ScalabilityStudy.add_scaling_strategies() method. design_size expects either:

  • a list of integer where the ith component is the size for the ith scaling strategy,

  • an integer changing the fixed size (if None, use the original size).

Note that we could also compare the optimization strategies while

  • varying the size of the different coupling variables (use coupling_size),

  • varying the size of the different equality constraints (use eq_size),

  • varying the size of the different inequality constraints (use ineq_size),

  • varying the size of any variable (use variables),

where the corresponding arguments works in the same way as design_size, except for variables which expects a list of dictionary whose keys are variables names and values are variables sizes. In this way, we can use this argument to fine-tune a scaling strategy to very specific variables, e.g. local variables.

study.add_scaling_strategies(design_size=[1, 10, 20])

Execute the scalable study

Then, we execute the scalability study, i.e. to build and execute a ScalableProblem for each optimization strategy and each scaling strategy, and repeat it 10 times in order to get statistics on the results (because the ScalableDiagonalModel relies on stochastic features.


Look at the dependency matrices

Here are the dependency matrices obtained with the 1st replicate when design_size=10.







Look at optimization histories

Here are the optimization histories obtained with the 1st replicate when design_size=10, where the left side represents the MDF formulation while the right one represents the IDF formulation.

Objective function

../../_images/MDF_2_1_obj_history-1.png ../../_images/IDF_2_1_obj_history-1.png

Design variables

../../_images/MDF_2_1_variables_history-1.png ../../_images/IDF_2_1_variables_history-1.png

Equality constraints

../../_images/MDF_2_1_eq_constraints_history-1.png ../../_images/IDF_2_1_eq_constraints_history-1.png

Inequality constraints

../../_images/MDF_2_1_ineq_constraints_history-1.png ../../_images/IDF_2_1_ineq_constraints_history-1.png

Post-process the results

Lastly, we plot the results. Because of the replicates, the latter are not displayed as one line per optimization strategy w.r.t. scaling strategy, but as one series of boxplots per optimization strategy w.r.t. scaling strategy, where the boxplots represents the variability due to the 10 replicates. In this case, it seems that the MDF formulation is more expensive than the IDF one when the design space dimension increases while they seems to be the same when each design parameter has a size equal to 1.

post = plot_scalability_results("study")
post.labelize_scaling_strategy("Number of design parameters per type.")
    xmargin=3.0, xticks=[1.0, 10.0, 20.0], xticks_labels=["1", "10", "20"], widths=1.0

Total running time of the script: ( 0 minutes 0.000 seconds)

Gallery generated by Sphinx-Gallery