Check how rk45 handles assignment rules

[kirichoi]

No description provided.

[evilnose]

Hi Kiri, may I close this?

[kirichoi]

I think rk45 solver cannot handle assignment rules properly. I don't think this is a priority anymore but should be investigated sometime.

[matthiaskoenig]

If there is an issue it should be picked up by the SBML test suite. I.e. in the continious integration testing the SBML TestSuite should be run with the various integrators, i.e.

• run all deterministic tests with all deterministic integrators
• run all stochastic tests with all stochastic integrators

[hsauro]

I agree, rk45 should be tested against the test suite to see where it stands.

…

On Mon, May 25, 2020 at 11:45 PM Matthias König ***@***.***> wrote: If there is an issue it should be picked up by the SBML test suite. I.e. in the continious integration testing the SBML TestSuite should be run with the various integrators, i.e. - run all deterministic tests with all deterministic integrators - run all stochastic tests with all stochastic integrators — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#445 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAIBSDQOQBFXVKEAVEVK6LDRTNQPVANCNFSM4EQMG75A> .

-- Herbert Sauro, Professor University of Washington, Bioengineering 206-685-2119, www.sys-bio.org [email protected] Books: http://books.analogmachine.org/

[luciansmith]

I checked the rk45 and rk4 integrators against the test suite. The rk45 integrator ran until test 65, and then got into an infinite loop, after failing nearly 100% of the tests before that. I then changed 'epsilon' to 1e-10, which I had found worked for test 1, and re-ran; it then got to test 172, (passing maybe 10 of them, likely the 'soft' tests), at which point it crashed when trying to calculate the factorial of 1315018.

The rk4 integrator did somewhat better--it never crashed/stopped, and successfully passed 1214 out of 1632 of the tests, which is only 74%, but it's a lot higher than 1%...

Given this, it looks like running those integrators against the test suite in general is a losing proposition. We do some spot checking of the integrators in the .rrtest files in Python, so they're at least checked to make sure they run, but more extensive testing seems out of our grasp at the moment.

However, I did implement #853, which should allow us to test integrators against the test suite more easily in the future.

Leaving this issue open for now; it would indeed still be nice to test these integrators.