Friday, November 4, 2016

How do I change the simulation tolerances in Tellurium?

November 4, 2016 8:59 am

For very complicated and large models it may be necessary to adjust the simulator tolerances in order to get the correct simulation results. Sometimes the simulator will terminate a simulation because it was unable to proceed due to numerical errors. In many cases this is due to a bad model and the user must investgate the model to determine what the issue might be. If the model is assumed to be correct then the other option is to change the simulator tolerances. The current option state of the simulator is obtained using the getInfo call, for example:

r = roadrunner.RoadRunner ('mymodel.xml')
roadrunner.RoadRunner() { 
'this' : 22F59350
'modelLoaded' : true
'modelName' : __main
'libSBMLVersion' : LibSBML Version: 5.13.0
'jacobianStepSize' : 1e-005
'conservedMoietyAnalysis' : false
'simulateOptions' : 
< roadrunner.SimulateOptions() 
'this' : 1B89AF28, 
'reset' : 0,
'structuredResult' : 0,
'copyResult' : 1,
'steps' : 50,
'start' : 0,
'duration' : 20
'integrator' : 
< roadrunner.Integrator() >
  name: cvode
      relative_tolerance: 0.00001
      absolute_tolerance: 0.0000000001
                   stiff: true
       maximum_bdf_order: 5
     maximum_adams_order: 12
       maximum_num_steps: 20000
       maximum_time_step: 0
       minimum_time_step: 0
       initial_time_step: 0
          multiple_steps: false
      variable_step_size: false


There are a variety of tuning parameters that can be changed in the simulator. Of interest are the relative and absolute tolerances, the maximum number of steps, and the initial time step.

The smaller the relative tolerance the more accurate the solution, however too small a value will result in either excessive runtimes or more likely roundoff errors. A relative tolerance of 1E-4 means that errors are controlled to 0.01%. An optimal value is roughly 1E-6. The absolute tolerance is used when a variable gets so small that the relative tolerance doesn't make much sense to apply. In these situations, absolute error tolerance is used to control the error. A small value for the absolute tolerance is often desirable, such as 1E-12, we do not recommend going below 1E-15 for either tolerance.

To set the tolerances use the statements:

r.integrator.absolute_tolerance = 5e-10
r.integrator.relative_tolerance = 1e-3

Another parameter worth changing if the simulations are not working well is to change the initial time step. This is often set by the integrator to be a relatively large value which means that the integrator will try to reduce this value if there are problems. Sometimes it is better to provide a small initial step size to help the integrator get started, for example, 1E-5.

r.integrator.initial_time_step = 0.00001

The reader is referred to the CVODE documentation for more details.

No comments: