Record stats on fitting solution run?

[TorkelE]

E.g. DifferentialEquations record various data of the simulation (like the number f function evaluations): While not directly needed, it can be useful for debugging purposes, or analysis of performance, and similar.

Stuff which it could be useful to know:

• Number of function evaluations (of the cost function).
• Number of occurrences of various warnings/simulations problems (e.g. maxiters reached, instability detected, and similar).

Another advantage of keeping this is that, rather than notifying the user of problems by displaying DiffEq warning messages, one could provide a summary at the end (auto display of diffeq warnings could still be an option). E.g. if the number of warnings of a certain type exceeds some number (e.g. 0), then at the end a summary is printed:

Warning: A large number of simulations encountered problems
Number of model simulations 1000
Number of maxiters reached: 96

[sebapersson]

I think this is a good idea.

Things like number of function evaluations and gradient evaluations should typically be on the domain of the optimization algorithm, however not every algorithms records this, and more details like how the ODE solver failed can be really useful for getting info on whether or not it is worthwhile to switch solver.

How about recording number of cost function, gradient and Hessian evaluations, along with number of warnings for each? (for really fun models the cost might evaluate, but for example the gradient can fail).

Would probably also need a state reset function every time a new optimization problem is built.

[TorkelE]

Yes, something like that sounds good. A little bit unsure about the state reset function? Do You mean for resetting these statistics or something else? Wouldn't these statistics be stored in the optimisation output structure (and whatever thing keeps track of the optimisation during the optimisation process)?

[sebapersson]

You are correct (somehow I imagined they should be in the PEtabODEProblem), but in the optmisation ouput structure makes far more sense!