Implementing MPC in the new API v0.6

[DarioSlaifsteinSk]

Hi, I'm going around the new API and the JuMP-dev talk and the new API for warmstarting and updating parameters.

In the talk it presents:

for i in 1:N
   optimize!(model)
   warmstart_backend_start_value(model)
   set_parameter_value(x0, value(0))
   set_parameter_value(setpoint, set[i])
end

A couple of doubts/questions:

1. so i'm guessing that the @infinite_parameter(t) is not modified (we always look at the same opt window $\mathcal{D}_t$, right?
2. set_parameter_value also translates to @parameter_functions? For example, when creating a model I usually use:

t = model[:t];
Dt = supports(t);
t0 = supports(t)[1]; Δt=supports(t)[2]-supports(t)[1];
# make new time window
it0 = round(Int,(t0/Δt));
itend = it0+length(supports(t))-1;
measurement = MPPTData[it0:itend];
@parameter_function(model, Ppv == (t) -> measurement[zero_order_time_index(Dt, t)])

could i just updatePpv instead of redefining the whole model?

3. I also use a plant simulator in the middle in case i have plant mismatch. I interface it with a Dict, there wouldn't be any issue right?

great new update, really excited to put it in production.

[pulsipher]

Hi @DarioSlaifsteinSk,

Thanks for reaching out, we are currently making an extensive tutorial and FAQ on MPC to make the new interface more clear, so keep an eye out for that.

In the mean time, I am happy to answer questions.

1. so i'm guessing that the @infinite_parameter(t) is not modified (we always look at the same opt window Dt, right?

Correct, we leave the domain of t alone. Just keep in mind that for @infinite_parameter(model, t in [0, tf]), the current time is treated as t=0 and tf is the final time of the time horizon for the controller and will shift by one time step on each MPC iteration.

2. set_parameter_value also translates to @parameter_functions?

Correct, we can update finite parameters and parameter functions using set_parameter_value. For parameter functions, you need to provide an appropriate function just as you would when using parameter_function. For instance, with a setpoint function you probably would need to update it on each MPC iteration to shift the time input.

3. I also use a plant simulator in the middle in case i have plant mismatch. I interface it with a Dict, there wouldn't be any issue right?

This shouldn't be an issue, if you run into any problems, let me know. Just keep in mind that the persistent interface only supports warmstarts and parameter updates at the moment, other changes will cause the backend to be rebuilt (increasing the computational cost).