Datasets:
tensorkelechi
/
arxiv-cstext

Dataset card Data Studio Files
xet
Community
1
text
stringlengths
0
89.3k
Chen et al 25
Ours 32Additive uncertainty u1D464Fig 2 Region of attraction RoA of 25 and our method 32 relative to the maximal RCI set in percent where we vary
the uncertainty in the dynamics matrix A ie ϵA left and the additive disturbance ie σw right The other uncertainty
parameters are fixed to ϵB 01σw 01 and ϵA 01ϵB 01 respectively and the horizon is N510
double integrator example before showing the benefits
of the asynchronous computation scheme on a more com
plex example All examples are implemented in Python
using CVXPY 34 and are solved using MOSEK 35
The examples were run on a machine equipped with an
Intel i97940X 31 GHz CPU and 32 GB of RAM
61 Double Integrator
In the first example we use the same system as in 25
ie uncertain LTI system 1 with discretetime dy
namic matrices
A
1 015
01 1
B
01
11
subject to polytopic constraints and disturbance
8
8
x1
x2
8
8
4u4
σw
σw
w1
w2
σw
σw
with polytopic model uncertainty
Aco
ϵA0
0 0
ϵA0
0 0
Bco
0
ϵB
0
ϵB
where the uncertainty parameters σw ϵA and ϵBare
varied Additionally we use the cost function lz v
zQzvRvwith Q 10I2andR 1 and hori
zonN 510 As the terminal sets we use the max
imal robust control invariant RCI set of dynamics 1
for 25 and the maximal RPI set Definition 4 with
Kfthe LQR controller for 32 We design the auxil
iary disturbance set Wto be equivalent to Wfor both
methods and since all uncertainty sets are hyperrectan
gles we use Remark 2 to modify 32 This setup allows a
direct comparison of 32 to the method proposed in 25
Section 5 and therefore also a comparison to the other
methods compared against in 25 Section 6Table 1
Computation times for RoA computations in double inte
grator example
Horizon Comp Time ms
N Min Median Rel Diff
Chen et al 255 299 501
10 905 1442
Ours 325 376 592 182
10 1133 1651 145
Figure 2 shows the region of attraction RoA ie the
set of initial states for which the MPC method is feasible
for both methods as a fraction of the maximal RCI set for
dynamics 1 In Figure 2 left we vary the parametric
uncertainty ϵA0005 while fixing ϵBσw 01
and in Figure 2 right we vary the additive distur
bance σw0006 while fixing ϵAϵB 01 We
notice that in general the method in 25 is less conser
vative than 32 ie its RoA covers a larger part of
the maximal RCI set This is due to the additional con
straints in 32 needed for recursive feasibility and the
use of an RCI terminal set compared to the RPI ter
minal set in 32 However with increasing horizon N
the RoAs of the two methods become more similar with
the RoA of 32 increasing and the RoA of the method
in 25 decreasing for larger horizons This is because the
method in 25 tightens the RCI terminal set with the dis
turbance overapproximation which becomes more con
servative for larger horizons thus reducing the RoA In
contrast the RoA of 32 increases with the horizon
since the MPC controller is enhancing the terminal RPI
set thus reducing conservativeness with larger horizons
Given that 25 Section 6 shows filterbased SLTMPC
outperforming all other investigated methods we can