[Utilities] Drop MOI.RawParameters when resetting a CachingOptimizer

[odow]

Closes #1220.

@blegat: An alternative is to not pass any OptimizerAttributes. If you're passing a new empty optimizer, it seems weird that the parameters would follow.

What do we expect to happen in the following at the JuMP level?

model = Model(Cbc.Optimizer)
set_silent(model)
set_optimizer(model, Clp.Optimizer)
MOI.get(model, MOI.Silent())  # true or false?

[blegat]

This is currently a mess. We have a documented option:

MathOptInterface.jl/src/MathOptInterface.jl

Lines 110 to 114 in f71ed83

	then an [`UnsupportedAttribute`](@ref) error is thrown. Unsupported *optimizer*
	attributes are treated differently:
	* If `warn_attributes` is `true`, a warning is displayed, otherwise,
	* the attribute is silently ignored.

which is not implemented as optimizer attributes are in fact not passed in copy_to but rather passed by the CachingOptimizer outside the copy_to function.
Note that we do have a warning for starting values though:

MathOptInterface.jl/src/Utilities/copy.jl

Lines 284 to 291 in f71ed83

	if !MOI.supports(dest, attr, supports_args...)
	@warn(
	"$attr is not supported by $(typeof(dest)). This " *
	"information will be discarded."
	)
	continue
	end
	end

I would probably add this warning option to CachingOptimizer instead of copy_to. Then you check supports before calling set and throws a warning if it's not supported and the user did not give the option to ignore this.

So

model = Model(Cbc.Optimizer)
set_silent(model)
set_optimizer(model, Clp.Optimizer)
MOI.get(model, MOI.Silent())

returns true if Clp supports MOI.Silent and throws a warning otherwise.

model = Model(Cbc.Optimizer, warn_option_name=false) # do we give the option here ?
set_silent(model)
set_optimizer(model, Clp.Optimizer, warn_option_name=false) # or here ?
MOI.get(model, MOI.Silent())

returns true if Clp supports MOI.Silent and ignores it silently otherwise.

One issue with this is that many solvers simply do MOI.supports(::Optimizer, ::MOI.RawParameter) = true without checking the parameter string but I'm in favor of fixing the solvers supports function.

[odow]

Passing RawParameter is also fraught though, even if it supports the parameter name. Because one solver might have print_level that takes an integer, and another might have print_level that takes a string or some solver-constant.

[joaquimg]

I know it would be very breaking, but what if RawParameter takes name and value?

[blegat]

We could do the following:

• If a RawParameter is set when no optimizer is attached: throw a warning or an error
• If a RawParameter is set when an optimizer is attached: set it to the optimizer but not to the cache.

The issues is that if some solver define custom optimizer attributes, we have the same issue. So I don't thing that RawParameter is a well-defined special case. Maybe we could have a subtype AbstractSolverSpecificOptimizerAttribute or a MOI.is_solver_specific trait

[odow]

We have

MathOptInterface.jl/src/Utilities/cachingoptimizer.jl

Lines 646 to 657 in 1315e31

	# Force users to specify whether the attribute should be queried from the
	# model_cache or the optimizer. Maybe we could consider a small whitelist of
	# attributes to handle automatically.
	# These are expert methods to get or set attributes directly in the model_cache
	# or optimizer.
	struct AttributeFromModelCache{T <: MOI.AnyAttribute}
	attr::T
	end
	struct AttributeFromOptimizer{T <: MOI.AnyAttribute}

which we haven't really utilized yet.

[odow]

Discussed on the developer call. Consensus was that we should only copy MOI.Silent and MOI.TimeLimitSec, if they are set.

Edit: I resolved to copy all copyable optimizer attributes that are supported, excluding RawParameters.

[blegat]

So if I undersand correctly, this is a workaround for solvers not implementing supports correctly for RawParameter? Otherwise, this would we caughr by the second condition

[blegat]

Or is it because RawParameter with the same name might have different meaning?

[blegat]

This should be clarified in a comment

[odow]

because RawParameter with the same name might have different meaning?

This.