Cannot build GenericNonlinearExpr

[aml5600]

Hello! I am new to InfiniteOpt but very interested in solving an optimal control problem. I have an MTK model provided to me, from which I can generate functions to evaluate the system dynamics and constraint equations.

I have variables for the states, controls, and constraints:

@infinite_parameter(opt_model, tn in [0, 1], num_supports = 51, derivative_method=OrthogonalCollocation(4))
InfiniteOpt.@variable(opt_model, x_lb[i] <= x_opt[i = 1:nx] <= x_ub[i], Infinite(tn), start = guess_xs[i])
InfiniteOpt.@variable(opt_model, u_lb[i] <= u_opt[i = 1:nu] <= u_ub[i], Infinite(tn), start = guess_us[i])
InfiniteOpt.@variable(opt_model, c_lb[i] <= c_opt[i = 1:nc] <= c_ub[i], Infinite(tn))

and am trying to evaluate the dynamics function:

dx = f(x_opt, u_opt, p, tn)

but am seeing

ERROR: Cannot build `GenericNonlinearExpr` because a term is complex-valued: `(-1//2 * ((44.56410256410257 x_opt[8](tn) - 48.27777777777778 x_opt[7](tn) - 3645.513174722222) + -1.1025000009488324 * sqrt(x_opt[6](tn)² + x_opt[5](tn)²)^2) * 20)::NLPExpr`
Stacktrace:
 [1] error(s::String)
   @ Base ./error.jl:35
 [2] _throw_if_not_real(x::NLPExpr)
   @ JuMP ~/.julia/packages/JuMP/7rBNn/src/nlp_expr.jl:334
 [3] *(x::NLPExpr, y::GenericNonlinearExpr{GeneralVariableRef})
   @ JuMP ~/.julia/packages/JuMP/7rBNn/src/nlp_expr.jl:391
 [4] macro expansion
   @ ~/.julia/packages/SymbolicUtils/EGhOJ/src/code.jl:468 [inlined]
 [5] macro expansion
   @ ~/.julia/packages/RuntimeGeneratedFunctions/M9ZX8/src/RuntimeGeneratedFunctions.jl:163 [inlined]
 [6] macro expansion
   @ ./none:0 [inlined]
 [7] generated_callfunc
   @ ./none:0 [inlined]
 [8] (::RuntimeGeneratedFunctions.RuntimeGeneratedFunction{…})(::Vector{…}, ::Vector{…}, ::Vector{…}, ::GeneralVariableRef)
   @ RuntimeGeneratedFunctions ~/.julia/packages/RuntimeGeneratedFunctions/M9ZX8/src/RuntimeGeneratedFunctions.jl:150
 [9] top-level scope

Not sure if relevant, but I have a) pointed some MTK functions back to Base: ModelingToolkit.NaNMath.cos(x) = Base.cos(x) and b) registered an interpolation function inside the MTK model.

Any tips on how to debug this? Thanks for the great package!

[pulsipher]

Hi there and welcome.

To better help you, would you please provide a minimum working example? From the code provided is not clear how f is produced and what you intend to do with dx if it were to work.

[aml5600]

The function f is generated by MTK from the model provided to me by:

(f, f_ip), dvs, psym, io_sys = ModelingToolkit.generate_control_function(sys, inputs, split = false);

where sys is an MTK ODESystem. This is the dynamics, so x_dot = f(x, u, p, t) where x is the state variable, u are inputs, and p are scalar parameters.

The plan is to enforce the dynamics as a constraint via:

@constraint(opt_model, [i = 1:nx], (∂((x_opt[i]), tn) - dx[i]) == 0)

I spent yesterday trying to recreate the issue with a MWE but was unable to isolate the issue. Unfortunately, I am unable to paste the full model (it is also spread across various files). I understand that is not helpful... =/

In general, are there typical failure cases where the NLPExpr is unable to be converted to Jump's GenericNonlinearExpr? Looking at the stacktrace, I cannot tell if the function is actually "not real", as the _is_real check is basically filtering out non-jump types?

[pulsipher]

Thanks for clarifying. The problem is likely that the current released version of InfiniteOpt has a legacy interface for nonlinear expressions based on InfiniteOpt.NLPExpr which may have issues with JuMP's new GenericNonlinearExprs. The new version of InfiniteOpt, that will be released shortly, is updated to use GenericNonlinearExprs.

To see if it fixes your problem, please try installing the current development version of InfiniteOpt:

using Pkg
Pkg.add(url = "https://github.com/infiniteopt/InfiniteOpt.jl", rev = "master")

[aml5600]

I'll give it a try, thank you!

[aml5600]

updating to master did the trick, thank you!!

[pulsipher]

Awesome, happy to help