Diagnosing problems in models

For structural problems pymob should offer constructive error messages, before problems occurr. Still during parameter inference and optimization many problems can ocurr, the below is an unsorted list of typical problems and ways to diagnose and fix such problems.

NaN likelihoods with JaxSolver

The pymob.solvers.diffrax.JaxSolver is highly efficient, but can be difficult to use.

• In the pymob.simulation.SimulationBase.dispatch_constructor() set throw_exception=True to make the solver fail if it runs into nans. This will make the solver fail loudly if there are problems. The default is that the solver returns infinities if there are problems. Numpyro can handle infinities in the probability functions consider the parameters impossible which can lead to severely biased estimates.

• First make sure this problem is not related to the choice and the parameterization of the solver. In the pymob.simulation.SimulationBase.dispatch_constructor(), try increasing max_steps,

• In dynamical systems parameters are very sensitive to the time scale. Make sure your parameters match the timescale

• Make prior predictive checks

Further reference