Optimization Studies in Python¶

The optimization study introduced in the preceding lesson used AnyBody’s builtin facilities for optimizing. Sometimes that is not enough, either because the objective functions depends on data that is not easily included in AnyBody, or because a different solver is needed.

In this tutorial we use an external optimizer together with AnyBody. The example is based on the model from the previous lesson but uses an optimizer from the Scipy python library. The same could also have been achived with other optimization frameworks (like NLopt, or languages (like MatLab).

Requirements¶

Before we begin you need to install the Anaconda Python distribution. It is free and comes with most of the necessary packages preinstalled.

We also need one additional Python library (AnyPyTools) which will make it easier to work with AnyBody from Python. AnyPyTools can be easily installed from the command prompt. Open the Anaconda command prompt and type:

conda config --add channels conda-forge
conda install anypytools


Example Python script¶

First we show the full Python program used in this tutorial. Secondly, we will go through and explain the different sections in the file.

  1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50  import math import scipy from anypytools import AnyPyProcess from anypytools.macro_commands import Load, OperationRun, Dump, SetValue def run_model(saddle_height, saddle_pos, silent=False): """Run the AnyBody model and return the metabolism results""" macro = [ Load("BikeModel2D.main.any"), SetValue("Main.BikeParameters.SaddleHeight", saddle_height), SetValue("Main.BikeParameters.SaddlePos", saddle_pos), OperationRun("Main.Study.InverseDynamics"), Dump("Main.Study.Output.Pmet"), Dump("Main.Study.Output.Abscissa.t"), ] app = AnyPyProcess(silent=silent) results = app.start_macro(macro) return results[0] def objfun(designvars): """Calculate the objective function value""" saddle_height = designvars[0] saddle_pos = designvars[1] result = run_model(saddle_height, saddle_pos, silent=True) if "ERROR" in result: raise ValueError("Failed to run model") pmet = scipy.integrate.trapz(result["Pmet"], result["Abscissa.t"]) return float(pmet) def seat_distance_constraint(designvars): """Compute contraint value which must be larger than zero""" return math.sqrt(designvars[0] ** 2 + designvars[1] ** 2) - 0.66 constaints = {"type": "ineq", "fun": seat_distance_constraint} bounds = [(0.61, 0.69), (-0.22, -0.05)] initial_guess = (0.68, -0.15) solution = scipy.optimize.minimize( objfun, initial_guess, constraints=constaints, bounds=bounds, method="SLSQP" ) print(solution) 

A copy of the file can be downloaded here. For now you can place the optimize.py next to your main file BikeModel2D.main.any. If you didn’t complete the model from lesson 2, you can download the finshed model here.

Importing the necessary libraries¶

The first part of the code is the import statements. They include the libraries which is used by the code:

 1 2 3 4 5  import math import scipy from anypytools import AnyPyProcess from anypytools.macro_commands import Load, OperationRun, Dump, SetValue 

Running a model from Python¶

For the external optimizers to work, we need a way to run AnyBody models from Python and record the results of the simulations, so we need to create a function to do this. There are more information on how to do this in the documentaion for AnyPyTools. So here we will just show how the code looks and not discuss the details.

  8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23  def run_model(saddle_height, saddle_pos, silent=False): """Run the AnyBody model and return the metabolism results""" macro = [ Load("BikeModel2D.main.any"), SetValue("Main.BikeParameters.SaddleHeight", saddle_height), SetValue("Main.BikeParameters.SaddlePos", saddle_pos), OperationRun("Main.Study.InverseDynamics"), Dump("Main.Study.Output.Pmet"), Dump("Main.Study.Output.Abscissa.t"), ] app = AnyPyProcess(silent=silent) results = app.start_macro(macro) return results[0] result = run_model(0.66, -0.16) print(result["Main.Study.Output.Pmet"]) 

The function run_model() takes saddle_height and saddle_pos as input and return the Pmet metabolism output from AnyBody.

If you use an interactive Python environment (like IPython) you could try calling the function directly to to test it and investigate the results:

In [4]: result = run_model(0.66, -0.16)
[****************100%******************]  1 of 1 completeTotal time: 0.8 seconds

In [5]: print(result.keys())
odict_keys(['Main.Study.Output.Pmet', 'Main.Study.Output.Abscissa.t'])

In [6]: print(result["Main.Study.Output.Pmet"])
[  17.20903341   73.49291834  209.58490241  379.67002659  559.57715608
736.92126247  901.88875426 1045.75303378 1162.65470516 1248.32088806
1299.79539032 1315.38241529 1294.6947524  1238.68684947 1149.59584772
1030.78784505  886.60667952  722.43408728  547.1840971   368.64175002
198.07134668   53.41928909   25.84379129   30.60376508   23.17442367
24.30809055  139.3209062   292.35610808  469.73382854  649.02576552
821.74094457  977.02863522 1108.05435136 1209.79739513 1278.65973442
1312.31195028 1309.70451022 1271.1212895  1198.17227557 1093.68215448
961.51890402  806.51623776  634.74029158  458.00117565  280.40563034
121.30841553   21.54859903   28.97200722   26.82989147   17.2090334 ]


As we expected the output contains the Main.Study.Output.Pmet value for each timestep in our model.

Defining the objective function¶

The next step is to define the objective function. The objective function should take a list of design values as input and return the objective function value. In Lesson 2 the objective function was the time integral of the metabolism variable. So we will do the same here with Scipy’s scipy.integrate.trapz(): function.

 23 24 25 26 27 28 29 30 31 32 33  def objfun(x): saddle_height = x[0] saddle_pos = x[1] result = run_model(saddle_height, saddle_pos, silent=True) if "ERROR" in result: raise ValueError("Failed to run model") # Integrate Pmet pmet = scipy.integrate.trapz(result["Pmet"], result["Abscissa.t"]) return float(pmet) 

Note

The function also checks the results for errors reported by AnyBody and raises a ValueError exception if that happens. There could be ways of handle error without failing but it is important to handle model failures, otherwise they may go unnoticed or mess with the optimization results.

Again, we can run this function interactively to test it:

In [9]: pmet = objfun([0.66, -0.16])

In [10]: print(pmet)
505.329399532772


Now we get the time integral of the Pmet variable as as single value, and we are now ready to define the optimization process.

Setting up the optimization study¶

We wrap things up by creating a function, similar to what we did in AnyBody, as well as defining the bounds and initial guess for the design variables.

 37 38 39 40 41 42 43 44  def seat_distance_constraint(x): """ Compute contraint value which must be larger than zero""" return (math.sqrt(x[0] ** 2 + x[1] ** 2) - 0.66) constaints = {"type": "ineq", "fun": seat_distance_constraint} bounds = [(0.61, 0.69), (-0.22, -0.05)] initial_guess = (0.68, -0.15) 

The documentation scipy.optimize.minimize() has more information on how to define bounds, contraints, tolerances, etc.

Finally, we call the scipy.optimize.minimize function run the optimizer. In this case we used the SLSQP algorithm.

 46 47 48  solution = scipy.optimize.minimize( objfun, initial_guess, constraints=constaints, bounds=bounds, method="SLSQP" ) 

Let us try to do this interactively and look at the results.

In [11]: solution = scipy.optimize.minimize(
...:     objfun, initial_guess, constraints=constaints, bounds=bounds, method="SLSQP"
...: )

In [12]: print(solution)
fun: 503.57634385063113
jac: array([39.43954086, -6.95677948])
message: 'Optimization terminated successfully.'
nfev: 56
nit: 12
njev: 12
status: 0
success: True
x: array([ 0.65010304, -0.11386853])


And there we have it! We can now take advantage of the many different algorithms and settings available for scipy.optimize.minimize(). We could also use a different package or customize our own algorithm, constraints etc. The possibilities are practically endless.

For more information regarding the AnyPyTools python package follow this link. You can also check out this webcast.