document constraint bounds

[mlubin]

Hi there. I just read the tutorial and didn't find a mention of how to set lower and upper bounds on constraints. (I was able to figure this out by looking at the code.) I recommend discussing this in the intro tutorial.

[binkob]

@mlubin: Could you share which specific .jl file you are referring to? I am interested

[jbcaillau]

@sshin23 @frapac hi there, on a related matter: when declaring (lower and upper) bounds on ExaModels variables, is it equivalent (= are the generated models eventually identical) to

• declare the variable then, later, the box constraint
• declare the variable with its bounds?

[sshin23]

@jbcaillau how to handle constraint (via variable bound or separate constraint) is up to the solver. Ipopt and MadNLP can be more efficient when you declare them as variable bounds. Generally, declaring them as variable bounds is likely to give better performance as it can provide the information in a more structured way.

[jbcaillau]

@jbcaillau how to handle constraint (via variable bound or separate constraint) is up to the solver. Ipopt and MadNLP can be more efficient when you declare them as variable bounds. Generally, declaring them as variable bounds is likely to give better performance as it can provide the information in a more structured way.

thanks @sshin23 yes, indeed. another choice could be, at the modeller level, to merge box constraints (declared further) with variable declarations. i get from your answer that this is not the case for ExaModels.

[sshin23]

Thanks, @jbcaillau. Could you suggest an example of such a modeling syntax? Can't you do this with the current syntax?

e.g., x = variable(c; lvar = -1, uvar = 1)?

[jbcaillau]

@sshin23 sure. following use case: imagine the user declares

x = variable(c)
# then later
constraint(c, x; lcon = -1, ucon = 1)

i assume it is always a good idea (*) to merge the two declarations into one according to what you mentioned:

x = variable(c; lvar = -1, uvar = 1)

of course, this can also be made at the solver level.

(*) there might indeed be cases when the user doesn't want such a merging (?)

[sshin23]

@jbcaillau I see. I'd imagine you want to reset some bound values after solving the problem?

Would something like this work?

variable!(x, model; lvar = -1, uvar = 1)

[jbcaillau]

@jbcaillau I see. I'd imagine you want to reset some bound values after solving the problem?

for instance that, yes

Would something like this work?

variable!(x, model; lvar = -1, uvar = 1)

👍🏽 sure!

[jbcaillau]

@sshin23 more precisely, imagine that the user declares

x = ExaModels.variable(c, n, 1:N+1; start=xs) # n >= 2
# then later
ExaModels.constraint(c, x[2, j] for j ∈ 1:N+1; lcon=0, ucon=vmax)

the later being a constraint only on a part of the array x. could be nice to merge the two things.

[sshin23]

Is there any reason you'd prefer constraint over variable?

You can achieve the same thing with

x = ExaModels.variable(c, n, 1:N+1; 
    uvar = [j == 2 ? vmax : Inf for i=1:n, j=1:N+1],  
    lvar = [j == 2 ? 0 : -Inf for i=1:n, j=1:N+1]
)

[jbcaillau]

thanks @sshin23 ; not a matter a preference but of sequence of actions. indeed, would be nice to be able to update bounds associated with a variable, e.g. with your syntax above:

x = ExaModels.variable(c, n, 1:N+1)
# then later
ExaModels.variable!(c, x; uvar = [j == 2 ? vmax : Inf for i=1:n, j=1:N+1],  lvar = [j == 2 ? 0 : -Inf for i=1:n, j=1:N+1])