added: support box constraint in LinMPC and NonLinMPC#379
Merged
Conversation
Codecov Report✅ All modified and coverable lines are covered by tests. Additional details and impacted files@@ Coverage Diff @@
## main #379 +/- ##
==========================================
+ Coverage 98.58% 98.60% +0.02%
==========================================
Files 27 27
Lines 5512 5595 +83
==========================================
+ Hits 5434 5517 +83
Misses 78 78 ☔ View full report in Codecov by Harness. 🚀 New features to boost your workflow:
|
Benchmark Results (Julia v1)Time benchmarks
Memory benchmarks
|
Member
Author
|
Wow these terrible results haha I did something wrong for sure 😆 |
Member
Author
|
Okay I ran some of the benchmark locally and many case studies did not converge anymore, that's why. There's presumably a nasty bug hidden somewhere. |
I just modify them on `setconstraint!` calls
Member
Author
|
Okay I have this MRE that converge to the best solution (at least a very good solution) on using ModelPredictiveControl, ControlSystemsBase, JuMP, Ipopt, BenchmarkTools
function f!(ẋ, x, u, _ , p)
g, L, K, m = p
θ, ω = x[1], x[2]
τ = u[1]
ẋ[1] = ω
ẋ[2] = -g/L*sin(θ) - K/m*ω + τ/m/L^2
end
h!(y, x, _ , _ ) = (y[1] = 180/π*x[1])
p = [9.8, 0.4, 1.2, 0.3]
nu = 1; nx = 2; ny = 1; Ts = 0.1
pendulum_model = NonLinModel(f!, h!, Ts, nu, nx, ny; p)
pendulum_p = p
h2!(y, x, _ , _ ) = (y[1] = 180/π*x[1]; y[2]=x[2])
nu, nx, ny = 1, 2, 2
pendulum_model2 = NonLinModel(f!, h2!, Ts, nu, nx, ny; p)
pendulum_p2 = p
Hp, Hc, Mwt, Nwt, Cwt = 20, 2, [0.5], [2.5], Inf
umin, umax = [-1.5], [+1.5]
model2, p = pendulum_model2, pendulum_p2
plant2 = deepcopy(model2)
plant2.p[3] = 1.25*p[3] # plant-model mismatch
σQ = [0.1, 1.0]; σR=[5.0]; nint_u=[1]; σQint_u=[0.1]
estim2 = UnscentedKalmanFilter(model2; σQ, σR, nint_u, σQint_u, i_ym=[1])
function JE(UE, ŶE, _ , p, _)
Ts = p
τ, ω = @views UE[1:end-1], ŶE[2:2:end-1]
return Ts*dot(τ, ω)
end
p = Ts; Mwt2 = [Mwt; 0.0]; Ewt = 3.5e3
x_0 = [0, 0]; x̂_0 = [0, 0, 0]; ry = [180; 0]
function gc!(LHS, Ue, Ŷe, _, p, ϵ)
Pmax = p
i_τ, i_ω = 1, 2
for i in eachindex(LHS)
τ, ω = Ue[i_τ], Ŷe[i_ω]
P = τ*ω
LHS[i] = P - Pmax - ϵ
i_τ += 1
i_ω += 2
end
return nothing
end
Cwt, Pmax, nc = 1e5, 3, Hp+1
x_0 = [0, 0]; x̂_0 = [0, 0, 0]; ry = [180; 0]
N = 35
optim = JuMP.Model(optimizer_with_attributes(Ipopt.Optimizer,"sb"=>"yes"), add_bridges=false)
transcription = MultipleShooting()
nmpc2_ipopt_ms = NonLinMPC(estim2;
Hp, Hc, Nwt=Nwt, Mwt=[0.5, 0], Cwt, gc!, nc, p=Pmax, optim, transcription
)
nmpc2_ipopt_ms = setconstraint!(nmpc2_ipopt_ms; umin, umax)
JuMP.unset_time_limit_sec(nmpc2_ipopt_ms.optim)
samples, evals, seconds = 100, 1, 15*60
# @btime(
# sim!($nmpc2_ipopt_ms, $N, $ry; plant=$plant2, x_0=$x_0, x̂_0=$x̂_0, progress=false),
# samples=samples, evals=evals, seconds=seconds, setup=GC.gc()
# )
unset_silent(nmpc2_ipopt_ms.optim)
initstate!(nmpc2_ipopt_ms, [0], [0])
setstate!(nmpc2_ipopt_ms, x̂_0)
preparestate!(nmpc2_ipopt_ms, [0])
u = moveinput!(nmpc2_ipopt_ms, ry)We can see that the objective increases throughtout the iterrations: The results on I'm not sure what's happening. I will continue my investigations. Any clue @cvanaret on why the iterations could suddenly increase like this ? edit : There is only one box constraint here, the slack edit 2: It converges to the good solution if I set |
|
@franckgaga I'll try to run Ipopt on your instance to look at what's happening.
|
Member
Author
|
Okay thanks for the tips ! Some progress here: changing the weight I will run the benchmark with this new edit : to answer your question, yes |
This is necessary if there is a change on the state operating point, for the hard terminal constraints and `MultipleShooting` transcription.
Member
Author
|
Ok it's only the custom nonlinear constraint on the pendulum that results in worst runtime benchmarks. But this case is finicky, there are clearly multiple local minima near each other since I slightly change the slack weight For these specific case studies:
I ran them locally on Win 11 and this PR slightly improve the runtime. Not sure why it's different on the CI, it may be the Linux server, but this is clearly also related to the multiple local minima. There is also the DAQP solver results who is very slightly worse. This is a bit counterintuitive since the solver support box constraints natively. I will open an issue on DAQP side to investigate this with Daniel. My conclusion: I will merge this PR. This is just the way it should be: if it's a box constraint, define it as a box constraint. And if the solver does not support them natively, use |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Following discussion at #378, it is a good idea in terms of performances to treat constraints on the decision variable as box constraints, instead of linear inequality constraints, at least for NLP and interior point methods.
In this PR, these constraints:
SingleShootingare treated as box constraints. It applies to both$\mathbf{c_{(\bullet)}}$ it is no longer a box constraint and the associated bound is treated as a linear inequality constraint, like before.
LinMPCandNonLinMPC. Note that if there is a nonzero value in the associated softness parametersThe default QP solver for
LinMPCisOSQP.jl, and it does not support box constraint natively. It is still possible to define box constraints with the bridge mechanism ofJuMP.jl. The new defaultoptimargument forLinMPCisoptim=JuMP.Model(OSQP.MathOptInterfaceOSQP.Optimizer, add_bridges=true). It should not affect the performances since JuMP will automatically convert them as linear equality constraints, as it was the case before this PR. It's possible that the performances withDAQP.jlwill be improved however, since it supports box constraints natively.Warning
Constructing
LinMPCwithoptim=JuMP.Model(OSQP.MathOptInterfaceOSQP.Optimizer, add_bridges=false)will now throw an error. Please useadd_bridges=true(it should not affect performances).Let's see the benchmarks.