Training with a different forward model

This tutorial was generated using Literate.jl. Download the source as a .jl file. Download the source as a .ipynb file.

using SDDPimport HiGHSimport Testfunction create_air_conditioning_model(; convex::Bool)    return SDDP.LinearPolicyGraph(;        stages = 3,        lower_bound = 0.0,        optimizer = HiGHS.Optimizer,    ) do sp, t        @variable(sp, 0 <= x <= 100, SDDP.State, initial_value = 0)        @variable(sp, 0 <= u_production <= 200)        @variable(sp, u_overtime >= 0)        if !convex            set_integer(x.out)            set_integer(u_production)            set_integer(u_overtime)        end        @constraint(sp, demand, x.in - x.out + u_production + u_overtime == 0)        Ω = [[100.0], [100.0, 300.0], [100.0, 300.0]]        SDDP.parameterize(ω -> set_normalized_rhs(demand, ω), sp, Ω[t])        @stageobjective(sp, 100 * u_production + 300 * u_overtime + 50 * x.out)    endendconvex = create_air_conditioning_model(; convex = true)non_convex = create_air_conditioning_model(; convex = false)SDDP.train(    convex;    forward_pass = SDDP.AlternativeForwardPass(non_convex),    post_iteration_callback = SDDP.AlternativePostIterationCallback(non_convex),    iteration_limit = 10,)Test.@test isapprox(SDDP.calculate_bound(non_convex), 62_500.0, atol = 0.1)Test.@test isapprox(SDDP.calculate_bound(convex), 62_500.0, atol = 0.1)

Test Passed