Support for events

[AleMorales]

State and time-triggered events are common in dynamical models, so modelling frameworks that generate SciComp IR may want to pass this information along, rather than hook the callbacks after code generation. After all, events are part of a model definition even if they are stored separately. I don't how abstract the event support should be, maybe for now we could allow storing a CallbackSet?

[ChrisRackauckas]

While the goal isn't to re-create Modia.jl at all (if it ever gets released), this is another spot where I think we should look to Modia/Modelica for inspiration, along with Tensorflow.

https://link.springer.com/chapter/10.1007%2F978-3-319-47169-3_15

They treat if statements specially and make people opt-out with a NoEvent().

I have been thinking of doing something like ContinuousEvent(x>y,rho = 3) and one for DiscreteEvent. We could add to this repo a few <:AbstractDiscreteEvents, like PeriodicEvent and TimedEvent (whose implementation in DiffEq would of course just be via the appropriate callbacks in DiffEqCallbacks.jl). They should allow for an initialization as well. Then we can add an @continuous etc. macro which essentially just expands to creating a parameter, creating a ContinuousEvent which changes the parameter (and initializes it appropriately), and if there are any equations in there it would make an ifelse operation. i.e.

@continuous if x > 5
   D*x == x*y
else
   D*x == x - 3
end

would expand to:

Parameter(:if1) # Some gensym name
ContinuousEvent(x>5,if1==1,if1==0,if1==x>5) # condition, upcrossing, downcrossing, initialization
ifelse(if1,D*x == x*y,D*x == x - 3)

It probably needs some bikeshedding to handle more combinations of models well, but you can see where this is going.

[ChrisRackauckas]

Now that #41 is through, it's clear that a ContinuousEvent should be defined via a rootfinding equation with ~. But inequality equations are necessary and I'm not sure about the syntax.

[ChrisRackauckas]

Welp, answered myself.

#43

And that's all we needed. Now all of the conditionals are Operations which specify equality or inequality equations. This means:

@continuous if x > 5
   D*x ~ x*y
else
   D*x ~ x - 3
end

gives

Parameter(:if1) # Some gensym name
ContinuousEvent(x>5,if1~1,if1~0,if1~x>5) # condition, upcrossing, downcrossing, initialization
ifelse(if1,D*x ~ x*y,D*x ~ x - 3)

are all well-defined functions and types on Operation. This can be implemented.

[ufechner7]

Any progress on event support or any documented workaround?

[ChrisRackauckas]

The workaround is just to do it directly in the problem construction, just https://diffeq.sciml.ai/stable/features/callback_functions/ . But we're going to start automating discontinuity handling which will do this.

[ChrisRackauckas]

@YingboMa and I talked about this and we have a plan for what to do.