Robbins

This problem was introduced by Robbins (1980) as a benchmark involving a third-order state constraint:

\[x_1^{(3)}(t) = u(t), \qquad x_1(t) \ge 0,\]

with control $u(t)$ and cost functional

\[J(x,u) = \int_0^T \big( \alpha x_1(t) + \beta x_1(t)^2 + \gamma u(t)^2 \big) \, dt.\]

The system is represented in first-order form as

\[\begin{aligned} \dot x_1(t) &= x_2(t), \\ \dot x_2(t) &= x_3(t), \\ \dot x_3(t) &= u(t), \end{aligned}\]

with boundary conditions

\[x(0) = (1, -2, 0), \qquad x(T) = (0,0,0),\]

and horizon $T=10$. In our simulations we use parameters $\alpha=3$, $\beta=0$, $\gamma=0.5$.

Qualitative behaviour

Robbins (1980) showed that the exact solution has infinitely many isolated contact points, where the state constraint $x_1(t)=0$ is active. The unconstrained arcs between contacts decrease geometrically, leading to an accumulation point, followed by the trivial singular arc $u=0$, $x=0$.

Numerically, capturing all contact points is challenging because the unconstrained arcs rapidly shrink below the resolution of a given discretisation. Simulations reproduce the initial part of the solution, showing the first few contact points.

References

Robbins, H. M. (1980). Junction phenomena for optimal control with state-variable inequality constraints of third order. Journal of Optimization Theory and Applications, 31, 85–99.
Hermant, A. (2008). Sur l'algorithme de tir pour les problèmes de commande optimale avec contraintes sur l'état (PhD thesis, École Polytechnique X).
Jacobson, D. H., Lele, M. M., & Speyer, J. L. (1971). New necessary conditions of optimality for control problems with state-variable inequality constraints. Journal of Mathematical Analysis and Applications, 35, 255–284.
BOCOP repository: https://github.com/control-toolbox/bocop

Packages

Import all necessary packages and define DataFrames to store information about the problem and resolution results.

using OptimalControlProblems    # to access the Beam model
using OptimalControl            # to import the OptimalControl model
using NLPModelsIpopt            # to solve the model with Ipopt
import DataFrames: DataFrame    # to store data
using NLPModels                 # to retrieve data from the NLP solution
using Plots                     # to plot the trajectories
using Plots.PlotMeasures        # for leftmargin, bottommargin
using JuMP                      # to import the JuMP model
using Ipopt                     # to solve the JuMP model with Ipopt

data_pb = DataFrame(            # to store data about the problem
    Problem=Symbol[],
    Grid_Size=Int[],
    Variables=Int[],
    Constraints=Int[],
)

data_re = DataFrame(            # to store data about the resolutions
    Model=Symbol[],
    Flag=Any[],
    Iterations=Int[],
    Objective=Float64[],
)

Initial guess

The initial guess (or first iterate) can be visualised by running the solver with max_iter=0. Here is the initial guess.

Click to unfold and see the code for plotting the initial guess.

function plot_initial_guess(problem)

    # dimensions
    x_vars = metadata[problem][:state_name]
    u_vars = metadata[problem][:control_name]
    n = length(x_vars) # number of states
    m = length(u_vars) # number of controls

    # import OptimalControl model
    docp = eval(problem)(OptimalControlBackend())
    nlp_oc = nlp_model(docp)

    # solve
    nlp_oc_sol = NLPModelsIpopt.ipopt(nlp_oc; max_iter=0)

    # build an optimal control solution
    ocp_sol = build_ocp_solution(docp, nlp_oc_sol)

    # plot the OptimalControl solution
    plt = plot(
        ocp_sol;
        state_style=(color=1,),
        costate_style=(color=1, legend=:none),
        control_style=(color=1, legend=:none),
        path_style=(color=1, legend=:none),
        dual_style=(color=1, legend=:none),
        size=(816, 220*(n+m)),
        label="OptimalControl",
        leftmargin=20mm,
    )
    for i in 2:n
        plot!(plt[i]; legend=:none)
    end

    # import JuMP model
    nlp_jp = eval(problem)(JuMPBackend())

    # solve
    set_optimizer(nlp_jp, Ipopt.Optimizer)
    set_optimizer_attribute(nlp_jp, "max_iter", 0)
    optimize!(nlp_jp)

    # plot
    t = time_grid(problem, nlp_jp)     # t0, ..., tN = tf
    x = state(problem, nlp_jp)         # function of time
    u = control(problem, nlp_jp)       # function of time
    p = costate(problem, nlp_jp)       # function of time

    for i in 1:n # state
        label = i == 1 ? "JuMP" : :none
        plot!(plt[i], t, t -> x(t)[i]; color=2, linestyle=:dash, label=label)
    end

    for i in 1:n # costate
        plot!(plt[n+i], t, t -> -p(t)[i]; color=2, linestyle=:dash, label=:none)
    end

    for i in 1:m # control
        plot!(plt[2n+i], t, t -> u(t)[i]; color=2, linestyle=:dash, label=:none)
    end

    return plt
end

plot_initial_guess(:robbins)

Solve the problem

OptimalControl model

Import the OptimalControl model and solve it.

# import DOCP model
docp = robbins(OptimalControlBackend())

# get NLP model
nlp_oc = nlp_model(docp)

# solve
nlp_oc_sol = NLPModelsIpopt.ipopt(
    nlp_oc;
    print_level=4,
    tol=1e-8,
    mu_strategy="adaptive",
    sb="yes",
)

Total number of variables............................:     2004
                     variables with only lower bounds:      501
                variables with lower and upper bounds:        0
                     variables with only upper bounds:        0
Total number of equality constraints.................:     1506
Total number of inequality constraints...............:        0
        inequality constraints with only lower bounds:        0
   inequality constraints with lower and upper bounds:        0
        inequality constraints with only upper bounds:        0


Number of Iterations....: 18

                                   (scaled)                 (unscaled)
Objective...............:   1.9466633163125312e+01    1.9466633163125312e+01
Dual infeasibility......:   2.5170381668000052e-14    2.5170381668000052e-14
Constraint violation....:   4.4408920985006262e-16    4.4408920985006262e-16
Variable bound violation:   9.9671948850228893e-09    9.9671948850228893e-09
Complementarity.........:   4.6995235824608513e-09    4.6995235824608513e-09
Overall NLP error.......:   4.6995235824608513e-09    4.6995235824608513e-09


Number of objective function evaluations             = 19
Number of objective gradient evaluations             = 19
Number of equality constraint evaluations            = 19
Number of inequality constraint evaluations          = 0
Number of equality constraint Jacobian evaluations   = 19
Number of inequality constraint Jacobian evaluations = 0
Number of Lagrangian Hessian evaluations             = 18
Total seconds in IPOPT                               = 0.533

EXIT: Optimal Solution Found.

The problem has the following numbers of steps, variables and constraints.

push!(data_pb,
    (
        Problem=:robbins,
        Grid_Size=metadata[:robbins][:N],
        Variables=get_nvar(nlp_oc),
        Constraints=get_ncon(nlp_oc),
    )
)

1×4 DataFrame

Row	Problem	Grid_Size	Variables	Constraints
	Symbol	Int64	Int64	Int64
1	robbins	500	2004	1506

JuMP model

Import the JuMP model and solve it.

# import model
nlp_jp = robbins(JuMPBackend())

# solve
set_optimizer(nlp_jp, Ipopt.Optimizer)
set_optimizer_attribute(nlp_jp, "print_level", 4)
set_optimizer_attribute(nlp_jp, "tol", 1e-8)
set_optimizer_attribute(nlp_jp, "mu_strategy", "adaptive")
set_optimizer_attribute(nlp_jp, "linear_solver", "mumps")
set_optimizer_attribute(nlp_jp, "sb", "yes")
optimize!(nlp_jp)

Total number of variables............................:     2004
                     variables with only lower bounds:      501
                variables with lower and upper bounds:        0
                     variables with only upper bounds:        0
Total number of equality constraints.................:     1506
Total number of inequality constraints...............:        0
        inequality constraints with only lower bounds:        0
   inequality constraints with lower and upper bounds:        0
        inequality constraints with only upper bounds:        0


Number of Iterations....: 18

                                   (scaled)                 (unscaled)
Objective...............:   1.9466633163125316e+01    1.9466633163125316e+01
Dual infeasibility......:   2.5228651541847610e-14    2.5228651541847610e-14
Constraint violation....:   3.6949610038305991e-16    3.6949610038305991e-16
Variable bound violation:   9.9671948850035614e-09    9.9671948850035614e-09
Complementarity.........:   4.6995235897622836e-09    4.6995235897622836e-09
Overall NLP error.......:   4.6995235897622836e-09    4.6995235897622836e-09


Number of objective function evaluations             = 19
Number of objective gradient evaluations             = 19
Number of equality constraint evaluations            = 19
Number of inequality constraint evaluations          = 0
Number of equality constraint Jacobian evaluations   = 1
Number of inequality constraint Jacobian evaluations = 0
Number of Lagrangian Hessian evaluations             = 1
Total seconds in IPOPT                               = 0.048

EXIT: Optimal Solution Found.

Numerical comparisons

Let's get the flag, the number of iterations and the objective value from the resolutions.

# from OptimalControl model
push!(data_re,
    (
        Model=:OptimalControl,
        Flag=nlp_oc_sol.status,
        Iterations=nlp_oc_sol.iter,
        Objective=nlp_oc_sol.objective,
    )
)

# from JuMP model
push!(data_re,
    (
        Model=:JuMP,
        Flag=termination_status(nlp_jp),
        Iterations=barrier_iterations(nlp_jp),
        Objective=objective_value(nlp_jp),
    )
)

2×4 DataFrame

Row	Model	Flag	Iterations	Objective
	Symbol	Any	Int64	Float64
1	OptimalControl	first_order	18	19.4666
2	JuMP	LOCALLY_SOLVED	18	19.4666

We compare the OptimalControl and JuMP solutions in terms of the number of iterations, the $L^2$-norm of the differences in the state, control, and variable, as well as the objective values. Both absolute and relative errors are reported.

Click to unfold and get the code of the numerical comparison.

function L2_norm(T, X)
    # T and X are supposed to be one dimensional
    s = 0.0
    for i in 1:(length(T) - 1)
        s += 0.5 * (X[i]^2 + X[i + 1]^2) * (T[i + 1]-T[i])
    end
    return √(s)
end

function numerical_comparison(problem, docp, nlp_oc_sol, nlp_jp)

    # get relevant data from OptimalControl model
    ocp_sol = build_ocp_solution(docp, nlp_oc_sol) # build an ocp solution
    t_oc = time_grid(ocp_sol)
    x_oc = state(ocp_sol).(t_oc)
    u_oc = control(ocp_sol).(t_oc)
    v_oc = variable(ocp_sol)
    o_oc = objective(ocp_sol)
    i_oc = iterations(ocp_sol)

    # get relevant data from JuMP model
    t_jp = time_grid(problem, nlp_jp)
    x_jp = state(problem, nlp_jp).(t_jp)
    u_jp = control(problem, nlp_jp).(t_jp)
    o_jp = objective(problem, nlp_jp)
    v_jp = variable(problem, nlp_jp)
    i_jp = iterations(problem, nlp_jp)

    x_vars = metadata[problem][:state_name]
    u_vars = metadata[problem][:control_name]
    v_vars = metadata[problem][:variable_name]

    println("┌─ ", string(problem))
    println("│")

    # number of iterations
    println("├─  Number of iterations")
    println("│")
    println("│     OptimalControl : ", i_oc)
    println("│     JuMP           : ", i_jp)
    println("│")

    # state
    for i in eachindex(x_vars)
        xi_oc = [x_oc[k][i] for k in eachindex(t_oc)]
        xi_jp = [x_jp[k][i] for k in eachindex(t_jp)]
        L2_oc = L2_norm(t_oc, xi_oc)
        L2_jp = L2_norm(t_oc, xi_jp)
        L2_ae = L2_norm(t_oc, xi_oc-xi_jp)
        L2_re = L2_ae/(0.5*(L2_oc + L2_jp))

        println("├─  State $(x_vars[i]) (L2 norm)")
        println("│")
        #println("│     OptimalControl : ", L2_oc)
        #println("│     JuMP           : ", L2_jp)
        println("│     Absolute error : ", L2_ae)
        println("│     Relative error : ", L2_re)
        println("│")
    end

    # control
    for i in eachindex(u_vars)
        ui_oc = [u_oc[k][i] for k in eachindex(t_oc)]
        ui_jp = [u_jp[k][i] for k in eachindex(t_jp)]
        L2_oc = L2_norm(t_oc, ui_oc)
        L2_jp = L2_norm(t_oc, ui_jp)
        L2_ae = L2_norm(t_oc, ui_oc-ui_jp)
        L2_re = L2_ae/(0.5*(L2_oc + L2_jp))

        println("├─  Control $(u_vars[i]) (L2 norm)")
        println("│")
        #println("│     OptimalControl : ", L2_oc)
        #println("│     JuMP           : ", L2_jp)
        println("│     Absolute error : ", L2_ae)
        println("│     Relative error : ", L2_re)
        println("│")
    end

    # variable
    if !isnothing(v_vars)
        for i in eachindex(v_vars)
            vi_oc = v_oc[i]
            vi_jp = v_jp[i]
            vi_ae = abs(vi_oc-vi_jp)
            vi_re = vi_ae/(0.5*(abs(vi_oc) + abs(vi_jp)))

            println("├─  Variable $(v_vars[i])")
            println("│")
            #println("│     OptimalControl : ", vi_oc)
            #println("│     JuMP           : ", vi_jp)
            println("│     Absolute error : ", vi_ae)
            println("│     Relative error : ", vi_re)
            println("│")
        end
    end

    # objective
    o_ae = abs(o_oc-o_jp)
    o_re = o_ae/(0.5*(abs(o_oc) + abs(o_jp)))

    println("├─  objective")
    println("│")
    #println("│     OptimalControl : ", o_oc)
    #println("│     JuMP           : ", o_jp)
    println("│     Absolute error : ", o_ae)
    println("│     Relative error : ", o_re)
    println("│")
    println("└─")

    return nothing
end

numerical_comparison(:robbins, docp, nlp_oc_sol, nlp_jp)

┌─ robbins
│
├─  Number of iterations
│
│     OptimalControl : 18
│     JuMP           : 18
│
├─  State x1 (L2 norm)
│
│     Absolute error : 1.9414120258534995e-15
│     Relative error : 4.265858622840488e-15
│
├─  State x2 (L2 norm)
│
│     Absolute error : 2.02022982137481e-15
│     Relative error : 1.6540324217418884e-15
│
├─  State x3 (L2 norm)
│
│     Absolute error : 5.288240514436061e-15
│     Relative error : 2.4086121895793998e-15
│
├─  Control u (L2 norm)
│
│     Absolute error : 6.702629882780922e-14
│     Relative error : 1.1196679276992505e-14
│
├─  objective
│
│     Absolute error : 3.552713678800501e-15
│     Relative error : 1.825027291072722e-16
│
└─

Plot the solutions

Visualise states, costates, and controls from the OptimalControl and JuMP solutions:

# build an ocp solution to use the plot from OptimalControl package
ocp_sol = build_ocp_solution(docp, nlp_oc_sol)

# dimensions
n = state_dimension(ocp_sol)   # or length(metadata[:robbins][:state_name])
m = control_dimension(ocp_sol) # or length(metadata[:robbins][:control_name])

# from OptimalControl solution
plt = plot(
    ocp_sol;
    state_style=(color=1,),
    costate_style=(color=1, legend=:none),
    control_style=(color=1, legend=:none),
    path_style=(color=1, legend=:none),
    dual_style=(color=1, legend=:none),
    size=(816, 240*(n+m)),
    label="OptimalControl",
    leftmargin=20mm,
)
for i in 2:n
    plot!(plt[i]; legend=:none)
end

# from JuMP solution
t = time_grid(:robbins, nlp_jp)     # t0, ..., tN = tf
x = state(:robbins, nlp_jp)         # function of time
u = control(:robbins, nlp_jp)       # function of time
p = costate(:robbins, nlp_jp)       # function of time

for i in 1:n # state
    label = i == 1 ? "JuMP" : :none
    plot!(plt[i], t, t -> x(t)[i]; color=2, linestyle=:dash, label=label)
end

for i in 1:n # costate
    plot!(plt[n+i], t, t -> -p(t)[i]; color=2, linestyle=:dash, label=:none)
end

for i in 1:m # control
    plot!(plt[2n+i], t, t -> u(t)[i]; color=2, linestyle=:dash, label=:none)
end