Loose ends from OptimizationSystem

[ValentinKaisermayer]

• Handling of nested constraints, see Constraints in nested systems are ignored #1785
• Proper inequality constraint handling, see discussion in Constraints and expression handling in OptimizationProblem #1660, see Adds support for inequalities #1799 and adds Inequality JuliaSymbolics/Symbolics.jl#717
  The definition of inequality constraints is decoupled from the definition of the actual bounds. Also one has to specify an equality constraint first even if it ends up as inequality. As much as possible should be specified during the modelling phase.

@variables x y
@parameters a b
loss = (a - x)^2 + b * (y - x^2)^2
cons = [
    x^2 + y^2 ~ 0, 
    y * sin(x) - x ~ 0
]
sys = OptimizationSystem(loss, [x, y], [a, b], name = :sys, constraints = cons)
prob = OptimizationProblem(sys, [x => 0.0, y => 0.0], [a => 1.0, b => 100.0],
                           lcons = [-1.0, -1.0], ucons = [500.0, 500.0])

• Proper handling of variable bounds, see Adds support for inequalities #1799
  Again this is decoupled from the system modelling and is specified during construction of the OptimizationProblem. However, MTK has already a way of specifiying box constraints on variables via e.g., @variables x [bounds=(-1.0, 1.0)].
• Handling of discrete and logical constraints, e.g., @variables x [binary=true]. see Adds binary and integer options to variable metadata #1798 and Add int field to OptimizationProblem SciMLBase.jl#270
• Support for MathOptInterface.AbstractOptimizer that do not support MathOptInterface.NLPBlock. If a problem is linear or quadratic why use an NLP solver? However, I see this more of an issue for OptimizationMOI.jl. This is important for MPCs which are typically QPs. Support for MathOptInterface.AbstractOptimizer that do not support MathOptInterface.NLPBlock Optimization.jl#369
• Convenient accessing solutions. For ODEProblem one can do sol[x].

[ChrisRackauckas]

Thanks, this is a nice list.

[ValentinKaisermayer]

Yeah!