Performance enhance (Direct Model)

[Thuener]

I'm dealing with a very big model and trying to increase its performance. I saw some mentions of direct_model in the code. Can we use the JuMP direct model with SDDP? Even with binary and Interger variables? What are the implications of using it?

[odow]

Use the undocumented

SDDP.jl/src/user_interface.jl

Line 941 in 73dd799

direct_mode::Bool = false,

model = SDDP.PolicyGraph(; direct_mode = true, optimizer = Gurobi.Optimizer) do sp, node
    # ...
end

What are the implications of using it?

Sometimes faster, sometimes slower. Always lower memory usage. Sometimes it will error. You're usually safe with Gurobi though.

[odow]

Note that

SDDP.jl/src/user_interface.jl

Lines 939 to 941 in 73dd799

	# These arguments are deprecated
	bellman_function = nothing,
	direct_mode::Bool = false,

I realized the reason:

SDDP.jl/src/algorithm.jl

Lines 295 to 306 in 73dd799

	function attempt_numerical_recovery(model::PolicyGraph, node::Node)
	if JuMP.mode(node.subproblem) == JuMP.DIRECT
	@warn(
	"Unable to recover in direct mode! Remove `direct = true` when " *
	"creating the policy graph."
	)
	else
	model.ext[:numerical_issue] = true
	MOI.Utilities.reset_optimizer(node.subproblem)
	optimize!(node.subproblem)
	end
	if !_has_primal_solution(node)

We can't attempt numerical recovery, and it is common enough that the problems I used direct-mode on just led to problems.

So I think I'm going to close this as won't-fix. You can use it if you want, but I don't want to document it for less-experienced users 😄