Add sample_backward_noise_terms_with_state (#742)

src/algorithm.jl

@@ -659,8 +659,11 @@ function solve_all_children(
             continue
         end
         child_node = model[child.term]
-        for noise in
-            sample_backward_noise_terms(backward_sampling_scheme, child_node)
+        for noise in sample_backward_noise_terms_with_state(
+            backward_sampling_scheme,
+            child_node,
+            outgoing_state,
+        )
             if length(scenario_path) == length_scenario_path
                 push!(scenario_path, (child.term, noise.term))
             else

src/plugins/headers.jl

@@ -115,10 +115,28 @@ abstract type AbstractBackwardSamplingScheme end
     )::Vector{Noise}
 
 Returns a `Vector{Noise}` of noises sampled from `node.noise_terms` using
-`backward_sampling_scheme`
+`backward_sampling_scheme`.
 """
 function sample_backward_noise_terms end
 
+"""
+    sample_backward_noise_terms_with_state(
+        sampler::AbstractBackwardSamplingScheme,
+        node::Node,
+        state::Dict{Symbol,Float64},
+    )::Vector{Noise}
+
+Returns a `Vector{Noise}` of noises sampled conditionally on the `state` using
+`sampler`.
+"""
+function sample_backward_noise_terms_with_state(
+    sampler::AbstractBackwardSamplingScheme,
+    node::Node,
+    ::Dict{Symbol,Float64},
+)
+    return sample_backward_noise_terms(sampler, node)
+end
+
 # =========================== duality_handlers =========================== #
 
 """

test/plugins/backward_sampling_schemes.jl

@@ -7,6 +7,7 @@ module TestBackwardPassSamplingSchemes
 
 using SDDP
 using Test
+import HiGHS
 
 function runtests()
     for name in names(@__MODULE__; all = true)
@@ -89,6 +90,65 @@ function test_MonteCarloSampler_100()
     return
 end
 
+mutable struct WithStateSampler <: SDDP.AbstractBackwardSamplingScheme
+    number_of_samples::Int
+end
+
+function test_WithStateSampler()
+    function sample_backward_noise_terms_with_state(
+        sampler::WithStateSampler,
+        node::SDDP.Node,
+        state::Dict{Symbol,Float64},
+    )
+        if state[:x] / node.index == 1.0
+            return [
+                SDDP.Noise((ϵ = 3.0,), 1 / sampler.number_of_samples) for
+                i in 1:sampler.number_of_samples
+            ]
+        elseif state[:x] / node.index == 3.0
+            return [
+                SDDP.Noise((ϵ = 1.0,), 1 / sampler.number_of_samples) for
+                i in 1:sampler.number_of_samples
+            ]
+        end
+    end
+    model = SDDP.LinearPolicyGraph(
+        stages = 5,
+        lower_bound = 0.0,
+        direct_mode = false,
+        optimizer = HiGHS.Optimizer,
+    ) do node, stage
+        @variable(node, x, SDDP.State, initial_value = 0.0)
+        @variable(node, ϵ)
+        SDDP.parameterize(node, stage * [1, 3], [0.9, 0.1]) do ω
+            return JuMP.fix(ϵ, ω)
+        end
+        @constraint(node, x.out == ϵ)
+    end
+    forward_trajectory = SDDP.forward_pass(
+        model,
+        SDDP.Options(model, Dict(:x => 1.0)),
+        SDDP.DefaultForwardPass(),
+    )
+    for node_index in 1:length(forward_trajectory.scenario_path)
+        state = forward_trajectory.sampled_states[node_index]
+        terms = sample_backward_noise_terms_with_state(
+            WithStateSampler(100),
+            model[node_index],
+            state,
+        )
+        for term in terms
+            @test term.probability == 0.01
+            if state[:x] / node_index == 1.0
+                @test term.term.ϵ == 3.0
+            elseif state[:x] / node_index == 3.0
+                @test term.term.ϵ == 1.0
+            end
+        end
+    end
+    return
+end
+
 end  # module
 
 TestBackwardPassSamplingSchemes.runtests()