www.neuron.yale.edu

I am currently using multiple run fitter to do a rather large parameter search on a single cell and from time to time I get an error like this :

IDA initialization failure, weighted norm of residual=1151.87

I'm curious what causes this? This is under a variable time step context; when you look at the time step is usually quite large, 100ms or so. Are there any tricks to avoiding this error, or a bit of code that will reset the time steps and continue the simulation?

Thanks

An IDA initialization error means that my default attempt to set the states so that f(y,y') = 0 failed to do so within the requested accuracy.
(f is the set of differential-algebraic equations governing the model)
The default attempt consists of a short 1e-9ms fixed implicit step with the hope that the step is short enough so that the differential states do not change by an observable amount, and long enough (i.e. > 0) so that the
algebraic states have setttled into their quasi-steady state by the end of the fixed step, and long enough (i.e. >> 0) so that the round off error does not
screw everything up via subtractive cancellation. See
http://www.neuron.yale.edu/neuron/stati ... init_dteps
for further information and how to select some alternative methods for initialization that may give better results for your model. However it
may be the case that the optimizer selected a parameter value that
just plain makes it impossible to initialize the problem.
Note that you can set the initialization style so that it produces a warning
instead of an error.
Without examining your model in further detail during such an initialization error I can't be more precise about how to avoid them. My overall understanding is that in a DAE system, there are some or many linear combinations of states that have a time constant of 0 and that if the
next smallest time constant is not >> than cvode.dae_init_dteps then
it will be very difficult to find a consistent initial condition such that
all the linear combinations of states with time constant > 0 remain fixed.

If you haven't already done so, you might consider applying reasonable constraints and
log scaling to some parameters (e.g. conductances and time constants) in the hope that,
strictly by chance, this might help the MRF avoid regions where initialization errors occur.

hines had mentioned that I could set it to produce a warning instead of a full stop. Would you be able to set it to run a proc that would return the parameters to a set of more stable values and continue the run? Where abouts would you find these bits of code controlling these errors?