Parameter sharing and abstraction in MTK

Specific Domains Modelling & Simulations
1

I have started using ParametrizedInterpolation from the standard library in my code, and while it works fine, having to add a clock to each component that uses it tarnishes the code a bit. So I wrote this:

""" A step function, where `step_function.u` is `ys[i]` when `t = ts[i]` (and afterwards).

The point is that while `ParametrizedInterpolation` can be used directly, its interface requires an
explicit `clock`. `StepFunction` is simple: add it as a System, and access its values with `sf.u`

@mtkcompile sf = StepFunction([1,2,3], [4,5,3])
res = solve(ODEProblem(sf, [], [1,10]))
println(res.([1,1.5,2,3,4,8], idxs=sf.val))

# output
[4.0, 4.0, 5.0, 3.0, 3.0, 3.0]
 """
@component function StepFunction(ts, ys; name=:unnamed_interpolator)
    @named interp = ParametrizedInterpolation(ExtrapolatedConstantInterpolation,
                                              Float64.(ys), Float64.(ts))
    @named clk = ContinuousClock()
    @variables val(t)
    eqs = [connect(clk.output, interp.input),
           val ~ interp.output.u]
    return System(eqs, t, [val], []; name, systems=[interp, clk])
end

And it works nicely.

Question 1: is there anything wrong with that?

Questions 2: Now setting the ts parameter looks like my_motor.input.interp.ts => [1,2,10]. How can I make it like my_motor.input.ts => [1,2,10]? I’ve read Initialization of Systems · ModelingToolkit.jl, but I’m not so confident. Should I set the interp.ts to missing, and set up an equation like interp.ts ~ ts with an additional ts parameter in StepFunction?

2

You might want to share clocks in order to reduce the total number of clocks that need to be tracked.

3

You could explore using ModelingToolkitInputs.jl. This will allow you to easy change your step function, like

step = Input(sys.x, [4.0, 5.0, 3.0], [1, 2, 3])
sol = solve(sys; inputs=[step])
1 Like
4

Interesting! Have you used it for optimization? Given that it supports remake, it should work out?