Quick-and-dirty test of interior point MCP backend

Author: lassepe

Project.toml

@@ -6,12 +6,12 @@ version = "0.1.1"
 [deps]
 BlockArrays = "8e7c35d0-a365-5155-bbbb-fb81a777f24e"
 ChainRulesCore = "d360d2e6-b24c-11e9-a2a3-2a2ae2dbcce4"
-ParametricMCPs = "9b992ff8-05bb-4ea1-b9d2-5ef72d82f7ad"
+MixedComplementarityProblems = "6c9e26cb-9263-41b8-a6c6-f4ca104ccdcd"
 TrajectoryGamesBase = "ac1ac542-73eb-4349-ae1b-660ab3609574"
 
 [compat]
-BlockArrays = "0.16"
+BlockArrays = "0.16, 1"
 ChainRulesCore = "1"
-ParametricMCPs = "0.1.14"
+MixedComplementarityProblems = "0.1.1"
 TrajectoryGamesBase = "0.3.6"
 julia = "1.7"

src/MCPTrajectoryGameSolver.jl

@@ -18,7 +18,7 @@ using TrajectoryGamesBase:
     unflatten_trajectory,
     unstack_trajectory
 
-using ParametricMCPs: ParametricMCPs, SymbolicUtils
+using MixedComplementarityProblems: MixedComplementarityProblems as IPMCPs, SymbolicUtils
 
 using BlockArrays: BlockArrays, mortar, blocks, eachblock
 using ChainRulesCore: ChainRulesCore

src/solve.jl

@@ -17,11 +17,12 @@ function TrajectoryGamesBase.solve_trajectory_game!(
 
     θ = compose_parameter_vector(; initial_state, context, shared_constraint_premultipliers)
 
-    raw_solution = ParametricMCPs.solve(
+    raw_solution = IPMCPs.solve(
+        IPMCPs.InteriorPoint(),
         solver.mcp_problem_representation,
         θ;
-        initial_guess = isnothing(initial_guess) ?
-                        generate_initial_guess(solver, game, initial_state) : initial_guess,
+        # initial_guess = isnothing(initial_guess) ?
+        #                 generate_initial_guess(solver, game, initial_state) : initial_guess,
         parametric_mcp_solve_options...,
     )
 
@@ -33,7 +34,8 @@ Reshapes the raw solution into a `JointStrategy` over `OpenLoopStrategy`s.
 """
 function strategy_from_raw_solution(; raw_solution, game, solver)
     number_of_players = num_players(game)
-    z_iter = Iterators.Stateful(raw_solution.z)
+    # z_iter = Iterators.Stateful(raw_solution.z)
+    z_iter = Iterators.Stateful(raw_solution.x)
 
     substrategies = map(1:number_of_players) do player_index
         private_state_dimension = solver.dimensions.state_blocks[player_index]
@@ -50,24 +52,24 @@ function strategy_from_raw_solution(; raw_solution, game, solver)
     TrajectoryGamesBase.JointStrategy(substrategies, info)
 end
 
-function generate_initial_guess(solver, game, initial_state)
-    ChainRulesCore.ignore_derivatives() do
-        z_initial = zeros(ParametricMCPs.get_problem_size(solver.mcp_problem_representation))
-
-        rollout_strategy =
-            map(solver.dimensions.control_blocks) do control_dimension_player_i
-                (x, t) -> zeros(control_dimension_player_i)
-            end |> TrajectoryGamesBase.JointStrategy
-
-        zero_input_trajectory = TrajectoryGamesBase.rollout(
-            game.dynamics,
-            rollout_strategy,
-            initial_state,
-            solver.dimensions.horizon,
-        )
-
-        copyto!(z_initial, reduce(vcat, flatten_trajetory_per_player(zero_input_trajectory)))
-
-        z_initial
-    end
-end
+# function generate_initial_guess(solver, game, initial_state)
+#     ChainRulesCore.ignore_derivatives() do
+#         z_initial = zeros(ParametricMCPs.get_problem_size(solver.mcp_problem_representation))
+#
+#         rollout_strategy =
+#             map(solver.dimensions.control_blocks) do control_dimension_player_i
+#                 (x, t) -> zeros(control_dimension_player_i)
+#             end |> TrajectoryGamesBase.JointStrategy
+#
+#         zero_input_trajectory = TrajectoryGamesBase.rollout(
+#             game.dynamics,
+#             rollout_strategy,
+#             initial_state,
+#             solver.dimensions.horizon,
+#         )
+#
+#         copyto!(z_initial, reduce(vcat, flatten_trajetory_per_player(zero_input_trajectory)))
+#
+#         z_initial
+#     end
+# end

src/solver_setup.jl

@@ -164,7 +164,7 @@ function Solver(
     ]
     upper_bounds = fill(Inf, length(lower_bounds))
 
-    mcp_problem_representation = ParametricMCPs.ParametricMCP(
+    mcp_problem_representation = IPMCPs.PrimalDualMCP(
         f_symbolic,
         z_symbolic,
         θ_symbolic,

test/Demo.jl

@@ -15,7 +15,7 @@ using BlockArrays: blocks, mortar
 using MCPTrajectoryGameSolver: Solver
 using GLMakie: GLMakie
 using Zygote: Zygote
-using ParametricMCPs: ParametricMCPs
+# using ParametricMCPs: ParametricMCPs
 
 """
 Set up a simple two-player collision-avoidance game:
@@ -77,12 +77,14 @@ function demo_model_predictive_game_play()
         # TODO: potentially allow the user to only warm-start the primals and or add noise
         generate_initial_guess = function (last_strategy, state, time)
             # only warm-start if the last strategy is converged / feasible
-            if !isnothing(last_strategy) &&
-               last_strategy.info.raw_solution.status == ParametricMCPs.PATHSolver.MCP_Solved
-                initial_guess = last_strategy.info.raw_solution.z
-            else
-                nothing
-            end
+            # if !isnothing(last_strategy) &&
+            #    # last_strategy.info.raw_solution.status == ParametricMCPs.PATHSolver.MCP_Solved
+            #    last_strategy.info.raw_solution.status === :solved
+            #     initial_guess = last_strategy.info.raw_solution.z
+            # else
+            #     nothing
+            # end
+            nothing
         end,
     )