MHE for networks of unconstrained linear systems

Updated:

Tags: , , , ,

Moving horizon estimation for unconstrained networks is surprisingly well-performing in a decentralized setting.


Introduction

In a centralized setting, the Kalman filter achieves optimal performance recursively.
In a decentralized setting, recursive filters (Luenberger observers), are greedy and compromise long-term performance.
Surprisingly, in a decentralized setting, the moving horizon estimation (MHE) framework, even for linear unconstrained systems achieves significantly higher performance than Luenberger-based filters.

In [1] a novel MHE framework and the moving finite horizon (MFH) method are proposed, to achieve higher performance in a decentralized setting with a focus on the implementation feasibility for very large-scale networks of systems. See documentation for MHEMovingFiniteHorizon for more information on the implementation of the MFH method.

This example contains extensive numerical simulations to assess the performance of the proposed MHE filter framework and of the MFH method.

To open this example execute the following command in the MATLAB command window

>> open MFH_SimulationResults

Simulation results

The performance of the novel MHE framework and the moving finite horizon (MFH) method are compared against:

The one-step and finite-horizon methods are based on a Lueberger observer and the PMHE1 method is a state-of-the-art distributed MHE solution whose communication, computational, and memory requirements are of the same magnitude as those of the MFH method.

All the functions, scripts, randomly generated models (and their generation parameters) are available in the example folder.

Small network

Let’s start with a rather small network of $N = 20$ systems. This network can be generated randomly using the function

>> rng_seed = 1;
>> generate_random_network(rng_seed);
Global state size: 40
Maximum absolute eigenvalue: 1.17425

The parameters inside generate_random_network can be adjusted to modify various parameters, e.g. the strength of the coupling between systems. This system was generated with coupled process noise, which is not supported by the PMHE1 method. For an example with the comparison with the PMHE1 method jump to Large-scale network and Large-scale network with weak couplings.

We obtain the following topology

Let’s synthesize the centralized and one-step methods with

>> synthesis_luenberger(1);
----------------------------------------------------------------------------------
Computing centralized kalman filter with: epsl = 1e-05 | maxIt = 1000.
Convergence reached with: epsl = 1e-05 | maxIt = 1000.
A total of 26 iterations were run.
----------------------------------------------------------------------------------
Elapsed time is 0.006611 seconds.
Trace centralized: 30.130313
----------------------------------------------------------------------------------
Running one-step algorithm with: epsl = 0.0001 | maxIt = 10000.
Convergence reached with: epsl = 0.0001 | maxIt = 10000.
A total of 43 iterations were run.
----------------------------------------------------------------------------------
Elapsed time is 0.020311 seconds.
Trace OS: 58.957642
Trace OS/C: 1.956755
----------------------------------------------------------------------------------
Running one-step algorithm with: epsl = 0.0001 | maxIt = 1000.
Convergence reached with: epsl = 0.0001 | maxIt = 1000.
A total of 108 iterations were run.
----------------------------------------------------------------------------------
Elapsed time is 0.048737 seconds.
Maximum absolute control CL eigenvalue: 0.972328

Note: This routine also computes a decentralized one-step controller just to guarantee that for unstable systems the state does not grow unbounded and causes numerical problems.

Let’s synthesize the FH method with

>> synthesis_fh(1,1e-4,75,10);
----------------------------------------------------------------------------------
Running finite-horizon algorithm with:
epsl = 0.0001 | W = 75 | maxOLIt = 10 | findWindowSize = false.
Outer-loop initialization.
	Outer-loop initialization iteration: 1.
	(...)
	Outer-loop initialization iteration: 75.
Outer-loop iteration 1.
	Inner-loop iteration 1.
	(...)
	Inner-loop iteration 75.
Outer-loop iteration 1 finished.
Window convergence within 1.77207e-06.
Maximum absolute CL eigenvalue: 0.880973.
Trace: 50.7401
LTI gain convergence within Inf.
Outer-loop iteration 2.
	Inner-loop iteration 1.
	(...)
	Inner-loop iteration 75.
Outer-loop iteration 2 finished.
Window convergence within 1.30285e-07.
Maximum absolute CL eigenvalue: 0.889327.
Trace: 48.3818
LTI gain convergence within 0.0464766.
Outer-loop iteration 3.
	Inner-loop iteration 1.
	(...)
	Inner-loop iteration 75.
Outer-loop iteration 3 finished.
Window convergence within 1.12747e-07.
Maximum absolute CL eigenvalue: 0.888831.
Trace: 48.3003
LTI gain convergence within 0.00168444.
Outer-loop iteration 4.
	Inner-loop iteration 1.
	(...)
	Inner-loop iteration 75.
Outer-loop iteration 4 finished.
Window convergence within 4.34028e-08.
Maximum absolute CL eigenvalue: 0.888993.
Trace: 48.2961
LTI gain convergence within 8.80485e-05.
Convergence reached with: epsl = 0.0001 | W = 75 | maxOLIt = 10
A total of 4 outer loop iterations were run, out of which 100.0% converged within
the specified minimum improvement.
----------------------------------------------------------------------------------
Elapsed time is 9.011818 seconds.
Trace FH: 48.296081
Trace FH/C: 1.602907
Maximum absolute CL eigenvalue: 0.888993

Finally, let’s synthesize the MFH method for various window sizes and with a null covariance initialization

>> for W = 1:10
synthesis_mfh(1,1e-4,W,1e4);
end
----------------------------------------------------------------------------------
Running moving finite horizon algorithm with:
epsl_inf = 0.0001 | W = 1 | maxIt = 10000 .
Convergence reached with: epsl_inf = 0.0001 | W = 1 | maxOLIt = 10000
A total of 43 outer loop iterations were run.
----------------------------------------------------------------------------------
Trace MFH: 58.957642
Trace MFH/C: 1.956755
For W_ss = 1 (tr = 58.9576) elapsed time is 0.15574 seconds.
----------------------------------------------------------------------------------
Running moving finite horizon algorithm with:
epsl_inf = 0.0001 | W = 2 | maxIt = 10000 .
Convergence reached with: epsl_inf = 0.0001 | W = 2 | maxOLIt = 10000
A total of 34 outer loop iterations were run.
----------------------------------------------------------------------------------
Trace MFH: 52.535678
Trace MFH/C: 1.743615
For W_ss = 2 (tr = 52.5357) elapsed time is 0.189517 seconds.
----------------------------------------------------------------------------------
Running moving finite horizon algorithm with:
epsl_inf = 0.0001 | W = 3 | maxIt = 10000 .
Convergence reached with: epsl_inf = 0.0001 | W = 3 | maxOLIt = 10000
A total of 33 outer loop iterations were run.
----------------------------------------------------------------------------------
Trace MFH: 48.760071
Trace MFH/C: 1.618306
For W_ss = 3 (tr = 48.7601) elapsed time is 0.249681 seconds.
----------------------------------------------------------------------------------
Running moving finite horizon algorithm with:
epsl_inf = 0.0001 | W = 4 | maxIt = 10000 .
Convergence reached with: epsl_inf = 0.0001 | W = 4 | maxOLIt = 10000
A total of 37 outer loop iterations were run.
----------------------------------------------------------------------------------
Trace MFH: 46.909798
Trace MFH/C: 1.556897
For W_ss = 4 (tr = 46.9098) elapsed time is 0.632296 seconds.
----------------------------------------------------------------------------------
Running moving finite horizon algorithm with:
epsl_inf = 0.0001 | W = 5 | maxIt = 10000 .
Convergence reached with: epsl_inf = 0.0001 | W = 5 | maxOLIt = 10000
A total of 33 outer loop iterations were run.
----------------------------------------------------------------------------------
Trace MFH: 45.750378
Trace MFH/C: 1.518417
For W_ss = 5 (tr = 45.7504) elapsed time is 1.62892 seconds.
----------------------------------------------------------------------------------
Running moving finite horizon algorithm with:
epsl_inf = 0.0001 | W = 6 | maxIt = 10000 .
Convergence reached with: epsl_inf = 0.0001 | W = 6 | maxOLIt = 10000
A total of 32 outer loop iterations were run.
----------------------------------------------------------------------------------
Trace MFH: 45.149851
Trace MFH/C: 1.498486
For W_ss = 6 (tr = 45.1499) elapsed time is 1.36903 seconds.
----------------------------------------------------------------------------------
Running moving finite horizon algorithm with:
epsl_inf = 0.0001 | W = 7 | maxIt = 10000 .
Convergence reached with: epsl_inf = 0.0001 | W = 7 | maxOLIt = 10000
A total of 34 outer loop iterations were run.
----------------------------------------------------------------------------------
Trace MFH: 44.726017
Trace MFH/C: 1.484419
For W_ss = 7 (tr = 44.726) elapsed time is 1.88515 seconds.
----------------------------------------------------------------------------------
Running moving finite horizon algorithm with:
epsl_inf = 0.0001 | W = 8 | maxIt = 10000 .
Convergence reached with: epsl_inf = 0.0001 | W = 8 | maxOLIt = 10000
A total of 35 outer loop iterations were run.
----------------------------------------------------------------------------------
Trace MFH: 44.027306
Trace MFH/C: 1.461230
For W_ss = 8 (tr = 44.0273) elapsed time is 2.40886 seconds.
----------------------------------------------------------------------------------
Running moving finite horizon algorithm with:
epsl_inf = 0.0001 | W = 9 | maxIt = 10000 .
Convergence reached with: epsl_inf = 0.0001 | W = 9 | maxOLIt = 10000
A total of 30 outer loop iterations were run.
----------------------------------------------------------------------------------
Trace MFH: 43.226535
Trace MFH/C: 1.434653
For W_ss = 9 (tr = 43.2265) elapsed time is 2.93729 seconds.
----------------------------------------------------------------------------------
Running moving finite horizon algorithm with:
epsl_inf = 0.0001 | W = 10 | maxIt = 10000 .
Convergence reached with: epsl_inf = 0.0001 | W = 10 | maxOLIt = 10000
A total of 29 outer loop iterations were run.
----------------------------------------------------------------------------------
Trace MFH: 43.079594
Trace MFH/C: 1.429776
For W_ss = 10 (tr = 43.0796) elapsed time is 4.15198 seconds.

The performance of the different decentralized methods is graphically depicted below

It is interesting to note that since the MFH method with $W_{ss} = 1$ and the OS method are equivalent, theire performance is the same.

We choose $W_{ss} = 5$ henceforth. It is also interesting to see the evolution of the estimation error covariance throughout the last iteration of the MFH algorithm, i.e. $\mathbf{P}(\tau \mid \tau \mid k), \tau = k-W_{ss},\ldots,k$, which is the third output of the MFH synthesis method.

Notice that:

  • Since convergence was reached, the trace at the beginning and end of the window is equal
  • The estimation performance in the middle of the window is very poor
  • But, it allows for a very good estimate at the end of the window
  • Because the intermediate estimates do not compromise future performance
  • This is the principle behind the performance improvement of the MFH method

Before deploying the MFH synthesis over the network, one can assess its stability by making use of the stability condition in [Section II.B,1].

>> model = 1;
>> Wss = 5;
>> stability_mfh(model,Wss);
MFH synthesis is stable.

Which confirms that the MFH synthesis that was obtained is stable.

Let’s assess the performance of the different estimation solutions resorting to Monte Carlo simulations:

>> model = 1;
>> parameters = struct('sim_T',100,'MFH_w_ss',5,'mc_it',1e4);
>> methods = [true; true; true; true; false];
>> simulation_mc(model, methods, parameters);
Trace C: 	30.130833 	(ri: 0)
Trace OS/C: 	1.956730 	(ri: 1.61975e-08)
Trace FH/C: 	1.602878 	(ri: 1.53113e-11)
Trace MFH/C: 	1.521114	(ri: 4.31462e-11)
Monte Carlo simulation 1/10000.
(...)
Monte Carlo simulation 9999/10000.
Monte Carlo simulation 10000/10000.
Finished Monte Carlo simulations.
Processing simulations.
Processing method 1/5.
Processing method 2/5.
Processing method 3/5.
Processing method 4/5.

After 10000 Monte Carlo simulations the evolution estimation error covariance is depicted below

The steady-state trace of the estimation error covariance matrix is depicted below, which was obtained by averaging the 25 last instants of the MC simulations.

  C OS FH MFH ($W_{ss} = 5$)
$\mathrm{tr}(\mathbf{P_{\mathrm{MC}}})$ 30.15 59.21 48.51 46.03
Relative to MFH $-34.49\%$ $+28.63\%$ $+5.39\%$ $–$

Conlusions:

  • The MFH method converged for different window sizes
  • The MFH method seems to have faster synthesis times than the FH method
  • It was confirmed that the MFH gain synthesis is stable
  • The MFH method performs better than state-of-the-art decentralized Luenberger-based filters
  • The principle behind the performance of the MFH is evident for this example
  • The use of the second best performing method (FH) incurs in a performance penalty of $+5.39\%$

Large-scale network

Let’s start increase the dimension of the network and generate a new model with $N = 500$ systems. This network was generated randomly using the function

>> rng_seed = 40;
>> generate_random_network(rng_seed);
Global state size: 1000
Maximum absolute eigenvalue: 1.21061

This is one of the systems briefly used in [1]. We obtain the following topology

Let’s synthesize the centralized and one-step methods with

>> synthesis_luenberger(40);
----------------------------------------------------------------------------------
Computing centralized kalman filter with: epsl = 1e-05 | maxIt = 1000.
Convergence reached with: epsl = 1e-05 | maxIt = 1000.
A total of 42 iterations were run.
----------------------------------------------------------------------------------
Elapsed time is 6.007283 seconds.
Trace centralized: 844.055346
----------------------------------------------------------------------------------
Running one-step algorithm with: epsl = 1e-05 | maxIt = 10000.
Convergence reached with: epsl = 1e-05 | maxIt = 10000.
A total of 138 iterations were run.
----------------------------------------------------------------------------------
Elapsed time is 259.215993 seconds.
Trace OS: 1788.026977
Trace OS/C: 2.118376
----------------------------------------------------------------------------------
Running one-step algorithm with: epsl = 1e-05 | maxIt = 1000.
Convergence reached with: epsl = 1e-05 | maxIt = 1000.
A total of 678 iterations were run.
----------------------------------------------------------------------------------
Elapsed time is 1277.864145 seconds.
Maximum absolute control CL eigenvalue: 0.992713

Synthesize finite-horizon with

>> synthesis_fh(40,1e-2,100,10);
----------------------------------------------------------------------------------
Running finite-horizon algorithm with:
epsl = 0.01 | W = 100 | maxOLIt = 10 | findWindowSize = false.
Outer-loop recomputation.
	Outer-loop recomputation iteration: 1.
	(...)
	Outer-loop recomputation iteration: 100.
Outer-loop iteration 1.
	Inner-loop iteration 1.
	(...)
	Inner-loop iteration 99.
	Inner-loop iteration 100.
Outer-loop iteration 1 finished.
Window convergence within 1.73139e-06.
Trace: 1446.84
LTI gain convergence within Inf.
Outer-loop iteration 2.
	Inner-loop iteration 1.
	(...)
	Inner-loop iteration 99.
	Inner-loop iteration 100.
Outer-loop iteration 2 finished.
Window convergence within 1.31444e-05.
Trace: 1448.36
LTI gain convergence within 0.0578383.
Outer-loop iteration 3.
	Inner-loop iteration 1.
    (...)
	Inner-loop iteration 99.
	Inner-loop iteration 100.
Outer-loop iteration 3 finished.
Window convergence within 2.09779e-05.
Trace: 1434.14
LTI gain convergence within 0.00767651.
Convergence reached with: epsl = 0.01 | W = 100 | maxOLIt = 10
A total of 3 outer loop iterations were run, out of which 100.0% converged within
the specified minimum improvement.
----------------------------------------------------------------------------------
Elapsed time is 13445.180700 seconds.
Trace FH: 1442.817245
Trace FH/C: 1.709387
Maximum absolute CL eigenvalue: 0.954752

Synthesize MFH gain sequences for window sizes $1–7$

>> parfor W = 1:7
    synthesis_mfh(40,1e-4,W,100);
end
----------------------------------------------------------------------------------
Running moving finite horizon algorithm with:
epsl_inf = 0.0001 | W = 1 | maxIt = 100 .
Convergence reached with: epsl_inf = 0.0001 | W = 1 | maxOLIt = 100
A total of 97 outer loop iterations were run.
----------------------------------------------------------------------------------
Trace MFH: 1785.363719
Trace MFH/C: 2.115221
For W_ss = 1 (tr = 1785.36) elapsed time is 809.953 seconds.

----------------------------------------------------------------------------------
Running moving finite horizon algorithm with:
epsl_inf = 0.0001 | W = 2 | maxIt = 100 .
Convergence reached with: epsl_inf = 0.0001 | W = 2 | maxOLIt = 100
A total of 84 outer loop iterations were run.
----------------------------------------------------------------------------------
Trace MFH: 1608.731635
Trace MFH/C: 1.905955
For W_ss = 2 (tr = 1608.73) elapsed time is 1894.05 seconds.

----------------------------------------------------------------------------------
Running moving finite horizon algorithm with:
epsl_inf = 0.0001 | W = 3 | maxIt = 100 .
Convergence reached with: epsl_inf = 0.0001 | W = 3 | maxOLIt = 100
A total of 69 outer loop iterations were run.
----------------------------------------------------------------------------------
Trace MFH: 1486.333674
Trace MFH/C: 1.760943
For W_ss = 3 (tr = 1486.33) elapsed time is 3636.26 seconds.

----------------------------------------------------------------------------------
Running moving finite horizon algorithm with:
epsl_inf = 0.0001 | W = 4 | maxIt = 100 .
Convergence reached with: epsl_inf = 0.0001 | W = 4 | maxOLIt = 100
A total of 51 outer loop iterations were run.
----------------------------------------------------------------------------------
Trace MFH: 1385.031524
Trace MFH/C: 1.640925
For W_ss = 4 (tr = 1385.03) elapsed time is 7772.36 seconds.

----------------------------------------------------------------------------------
Running moving finite horizon algorithm with:
epsl_inf = 0.0001 | W = 5 | maxIt = 100 .
Convergence reached with: epsl_inf = 0.0001 | W = 5 | maxOLIt = 100
A total of 59 outer loop iterations were run.
----------------------------------------------------------------------------------
Trace MFH: 1340.934405
Trace MFH/C: 1.588681
For W_ss = 5 (tr = 1340.93) elapsed time is 12913.6 seconds.

----------------------------------------------------------------------------------
Running moving finite horizon algorithm with:
epsl_inf = 0.0001 | W = 6 | maxIt = 100 .
Convergence reached with: epsl_inf = 0.0001 | W = 6 | maxOLIt = 100
A total of 59 outer loop iterations were run.
----------------------------------------------------------------------------------
Trace MFH: 1311.597588
Trace MFH/C: 1.553924
For W_ss = 6 (tr = 1311.6) elapsed time is 20756.7 seconds.

----------------------------------------------------------------------------------
Running moving finite horizon algorithm with:
epsl_inf = 0.0001 | W = 7 | maxIt = 100 .
Convergence reached with: epsl_inf = 0.0001 | W = 7 | maxOLIt = 100
A total of 57 outer loop iterations were run.
----------------------------------------------------------------------------------
Trace MFH: 1281.632054
Trace MFH/C: 1.518422
For W_ss = 7 (tr = 1281.63) elapsed time is 27268.1 seconds.

It is very important to note that the projected performance that is output by the implementation of the methods may have a lower precision than the displayed significant digits. It depends on the convergence criteria chosen for each method. One can compute the precise projected performance of the synthesis by making use, for instance, of the initial projection performed in simulation_mc. For the MFH synthesis, with $\mathrm{tr}(\mathbf{P}_{\infty}^{\mathrm{Cent}}) = 844.1$, we get

$W_{ss}$   2 3 4 5 6 7
$\text{tr}(\mathbf{P}_{\infty})$ /$\;\;\;\mathrm{tr}(\mathbf{P}_{\infty}^{\mathrm{Cent}})$ $1.911$ $1.764$ $1.647$ $1.590$ $1.555$ $1.520$

The performance of the different decentralized methods is graphically depicted below

We will consider $W_{ss} = 5$ henceforth.

Similarly to the previous synthetic network (Small network), we can observe the evolution of the trace of the covariance over the window of the last iteration of the MFH method. The same conclusions apply.

Before deploying the MFH synthesis over the network, one can assess its stability by making use of the stability condition in [Section II.B,1].

>> model = 40;
>> Wss = 5;
>> stability_mfh(model,Wss);
MFH synthesis is stable.

Which confirms that the MFH synthesis that was obtained is stable.

Lastly, we can check if the PMHE1 method is guaranteed to converge (with a window size equal to the size chosen for the MFH method)

>> farina_et_al_2010_convergence(40,5);
PMHE1 is not guaranteed to converge.

We conclude that the PMHE1 method is not guaranteed to converge. Note that we considered the process noise between coupled systems to be correlated, which PMHE1 does not support. Nevertheless, the convergence analysis only depends on $\mathbf{A}$ and $\mathbf{C}$, so that had no influence. Also note that, according to [3], there is no indication that it does not converge (it is just not guaranteed to converge).

Removing correlations between process noise of coupled systems (model model41), we can run a simulation to see whether the error dynamics of the PMHE1 method are stable

>> model = 41;
>> parameters = struct('sim_T',1,'PMHE1_N',5);
>> methods = [true; true; false; false; true];
>> simulation(model, methods, parameters);
Luenberger filters simulation completed.
PMHE1 filter iteration: 1
Starting parallel pool (parpool) using the 'local' profile ...
Connected to the parallel pool (number of workers: 30).
PMHE1 filter iteration: 2
(...)
PMHE1 filter iteration: 100
PMHE1 filter simulation completed.

We conclude that, even without process noise correlations, the PMHE1 method does not converge for this network.

Let’s assess the performance of the different estimation solutions resorting to Monte Carlo simulations:

>> model = 40;
>> parameters = struct('sim_T',200,'MFH_w_ss',5,'mc_it',1000);
>> methods = [true; true; true; true; false];
>> simulation_mc(model, methods, parameters);
Trace C: 	844.089937 	(ri: 1.34686e-16)
Trace OS/C: 	2.117501 	(ri: 1.83691e-05)
Trace FH/C: 	1.710140 	(ri: 4.34463e-11)
Trace MFH/C: 	1.590347	(ri: 4.31416e-09)
Monte Carlo simulation 1/1000.
(...)
Monte Carlo simulation 999/1000.
Monte Carlo simulation 1000/1000.
Finished Monte Carlo simulations.
Processing simulations.
Processing method 1/5.
Processing method 2/5.
Processing method 3/5.
Processing method 4/5.

After 1000 Monte Carlo simulations the evolution estimation error covariance is depicted below

The steady-state trace of the estimation error covariance matrix is depicted below, which was obtained by averaging the 20 last instants of the MC simulations.

  C OS FH MFH ($W_{ss} = 5$)
$\mathrm{tr}(\mathbf{P_{\mathrm{MC}}}) \times 10^{-3} $ 0.8449 1.797 1.451 1.348
Relative to MFH $-37.34\%$ $+33.29\%$ $+7.59\%$ $–$

In [1] it is argued that the MHE filter synthesize with the MFH method is consistent. We can perform a simulation and look at a component of the estimation error to confirm that.

>> model = 40;
>> parameters = struct('sim_T',1000,'MFH_w_ss',5);
>> methods = [true; true; true; true; false];
>> simulation(model, methods, parameters);
Luenberger filters simulation completed.
MFH filter simulation completed.

The evolution of a randomly chosen component of the error and the corresponding $3\sigma$ bounds is depicted below.

We can confirm that it is zero-mean and the calculated estimation covariance is not over-confident.

Conclusions:

  • The PMHE1 method does not converge
  • The MFH method converges for different window sizes
  • The MFH method supports correlated process noise
  • The MFH method is scalable
  • It was confirmed that the MFH gain synthesis is stable
  • For low window sizes the MFH method is significantly faster to synthesize than the FH method
  • The MFH method performs better than state-of-the-art decentralized Luenberger-based filters
  • The use of the second best performing method (FH) incurs in a performance penalty of $+7.59\%$
  • The MHE framework of the MFH method is consistent

Large-scale network with weak couplings

Let’s increase the dimension of the network even more and generate a new model with $N = 1000$ systems. The parameters for the generation of this network are set so that:

  • The dynamic couplings between systems are weak (so that the PMHE1 method converges)
  • The process noise is uncorrelated between systems (to employ the PMHE1 method)

This is the second synthetic network used in [1]. This network was generated randomly using the function

>> seed = 42;
>> generate_random_network(seed);
Global state size: 2000
Maximum absolute eigenvalue: 1.13125

We obtain the following topology

Lets’s synthesize the centralized and OS methods

>> synthesis_luenberger(42);
----------------------------------------------------------------------------------
Running one-step algorithm with: epsl = 0.0001 | maxIt = 1000.
Convergence reached with: epsl = 0.0001 | maxIt = 1000.
A total of 319 iterations were run.
----------------------------------------------------------------------------------
Elapsed time is 17276.108066 seconds.
Maximum absolute control CL eigenvalue: 1.000605

It was not possible to compute the FH synthesis in a reasonable amount of time (a few days).

Lets synthesize the MFH method with $W_{ss} = 2$

>> synthesis_mfh(42,1e-4,2,1e3);
----------------------------------------------------------------------------------
Running moving finite horizon algorithm with:
epsl_inf = 0.0001 | W = 2 | maxIt = 1000 .
Convergence reached with: epsl_inf = 0.0001 | W = 2 | maxOLIt = 1000
A total of 246 outer loop iterations were run.
----------------------------------------------------------------------------------
Trace MFH: 6722.959282
Trace MFH/C: 2.601179
For W_ss = 2 (tr = 6722.96) elapsed time is 74398.1 seconds.

Before deploying the MFH synthesis over the network, one can assess its stability by making use of the stability condition in [Section II.B,1].

>> model = 42;
>> Wss = 2;
>> stability_mfh(model,Wss);
MFH synthesis is stable.

Which confirms that the MFH synthesis that was obtained is stable.

Lastly, we find that the PMHE1 method is guaranteed to converge (with a window size equal to the size chosen for the MFH method)

>> farina_et_al_2010_convergence(42,2);
PMHE1 is guaranteed to converge.

We can simulate this methods can be simulated using function simulation for $W_{\text{PMHE1}} = 2$ and $W_{\text{PMHE1}} = 5$. The evolution of the error norm is depicted below

The average of the error norm for each of the methods is depicted below

  C OS MFH ($W_{ss} = 2$) PMHE1 ($W_{\text{PMHE1}} = 2$) PMHE1 ($W_{\text{PMHE1}} = 5$)
  49.95 173.9 81.92 132.3 140.1
Relative to MFH $-39.02\%$ $+112.3\%$ $–$ $+61.54\%$ $+70.98\%$

Conclusions:

  • The FH method could not be synthesized in a reasonable amount of time for $N=1000$ systems
  • The MFH method is scalable
  • It was confirmed that the MFH gain synthesis is stable
  • The use of the second best performing method (PMHE1) incurs in a performance penalty of $+61.54\%$

Case of FH convergence issues

Let’s generate another synthetic network with $N = 500$ systems. This network was generated randomly using the function

>> rng_seed = 47;
>> generate_random_network(rng_seed);
Global state size: 1000
Maximum absolute eigenvalue: 1.27932

We obtain the following topology

The centralized and OS gains can be computed similarly to the previous synthetic networks.

>> model = 47;
>> synthesis_luenberger(model);
----------------------------------------------------------------------------------
Computing centralized kalman filter with: epsl = 1e-05 | maxIt = 1000.
Convergence reached with: epsl = 1e-05 | maxIt = 1000.
A total of 130 iterations were run.
----------------------------------------------------------------------------------
Elapsed time is 15.921722 seconds.
Trace centralized: 822.287978
----------------------------------------------------------------------------------
Running one-step algorithm with: epsl = 0.0001 | maxIt = 10000.
Convergence reached with: epsl = 0.0001 | maxIt = 10000.
A total of 185 iterations were run.
----------------------------------------------------------------------------------
Elapsed time is 447.193402 seconds.
Trace OS: 3696.130373
Trace OS/C: 4.494934

Let’s attempt to synthesize the FH algorithm

>> synthesis_fh(47,1e-2,100,20);
----------------------------------------------------------------------------------
Running finite-horizon algorithm with:
epsl = 0.01 | W = 100 | maxOLIt = 20 | findWindowSize = false.
Outer-loop initialization.
	Outer-loop initialization iteration: 1.
	(...)
	Outer-loop initialization iteration: 100.
Outer-loop iteration 1.
	Inner-loop iteration 1.
	(...)
	Inner-loop iteration 100.
Outer-loop iteration 1 finished.
Window convergence within 4.99137e-05.
Maximum absolute CL eigenvalue: 1.02226.
Trace: 2944.17
LTI gain convergence within Inf.
Outer-loop iteration 2.
	Inner-loop iteration 1.
	(...)
	Inner-loop iteration 100.
Outer-loop iteration 2 finished.
Window convergence within 3.35438e-05.
Maximum absolute CL eigenvalue: 0.974221.
Trace: 2750.14
LTI gain convergence within 0.0659041.
Outer-loop iteration 3.
	Inner-loop iteration 1.
	(...)
	Inner-loop iteration 100.
Outer-loop iteration 3 finished.
Window convergence within 1.36555e-06.
Maximum absolute CL eigenvalue: 0.974231.
Trace: 2683.93
LTI gain convergence within 0.0240748.
Outer-loop iteration 4.
	Inner-loop iteration 1.
	(...)
	Inner-loop iteration 100.
Outer-loop iteration 4 finished.
Window convergence within 1.22254e-06.
Maximum absolute CL eigenvalue: 0.974207.
Trace: 2647.45
LTI gain convergence within 0.0135906.
Outer-loop iteration 5.
	Inner-loop iteration 1.
	(...)
	Inner-loop iteration 100.
Outer-loop iteration 5 finished.
Window convergence within 5.80219e-05.
Maximum absolute CL eigenvalue: 0.97423.
Trace: 2788.91
LTI gain convergence within 0.0534327.
Outer-loop iteration 6.
	Inner-loop iteration 1.
	(...)
	Inner-loop iteration 100.
Outer-loop iteration 6 finished.
Window convergence within 1.0339e-05.
Maximum absolute CL eigenvalue: 0.976927.
Trace: 2306.17
LTI gain convergence within 0.173094.
Outer-loop iteration 7.
	Inner-loop iteration 1.
	(...)
	Inner-loop iteration 100.
Outer-loop iteration 7 finished.
Window convergence within 0.000820389.
Maximum absolute CL eigenvalue: 1.04333.
Trace: 1978.28
LTI gain convergence within 0.142179.
Outer-loop iteration 8.
	Inner-loop iteration 1.
	(...)
	Inner-loop iteration 100.
Outer-loop iteration 8 finished.
Window convergence within 0.000166565.
Maximum absolute CL eigenvalue: 1.52682.
Trace: 1757.67
LTI gain convergence within 0.111517.
Outer-loop iteration 9.
	Inner-loop iteration 1.
	(...)
	Inner-loop iteration 100.
Outer-loop iteration 9 finished.
Window convergence within 1.97906e-05.
Maximum absolute CL eigenvalue: 1.01125.
Trace: 1575.71
LTI gain convergence within 0.103522.
Outer-loop iteration 10.
	Inner-loop iteration 1.
	(...)
	Inner-loop iteration 100.
Outer-loop iteration 10 finished.
Window convergence within 4.43926e-05.
Maximum absolute CL eigenvalue: 1.01106.
Trace: 1533.47
LTI gain convergence within 0.0268107.
Outer-loop iteration 11.
	Inner-loop iteration 1.
	(...)
	Inner-loop iteration 100.
Outer-loop iteration 11 finished.
Window convergence within 0.000609564.
Maximum absolute CL eigenvalue: 1.01723.
Trace: 1511.44
LTI gain convergence within 0.0143601.
Outer-loop iteration 12.
	Inner-loop iteration 1.
	(...)
	Inner-loop iteration 100.
Outer-loop iteration 12 finished.
Window convergence within 6.1706e-05.
Maximum absolute CL eigenvalue: 1.01814.
Trace: 1495.83
LTI gain convergence within 0.0103318.
Outer-loop iteration 13.
	Inner-loop iteration 1.
	(...)
	Inner-loop iteration 100.
Outer-loop iteration 13 finished.
Window convergence within 0.0002569.
Maximum absolute CL eigenvalue: 1.0185.
Trace: 1490.23
LTI gain convergence within 0.00374232.
Convergence reached with: epsl = 0.01 | W = 100 | maxOLIt = 20
A total of 13 outer loop iterations were run, out of which 100.0% converged within
the specified minimum improvement.
----------------------------------------------------------------------------------
Elapsed time is 54070.481776 seconds.
Trace FH: 1490.230726
Trace FH/C: 1.812298
Maximum absolute CL eigenvalue: 1.018502

The steady-state $\mathrm{tr}(\mathbf{P}(k))$ converged within the specified minimum improvement, but the gains do not converge towards a steady-state gain, thus, the finite-horizon synthesis fails. In fact, the closed-loop error dynamics matrix is unstable

>> load('./models/model47/synth_FH.mat')
>> max(abs(eig((eye(size(A))-KFH*C)*A)))
ans =
    1.0185

Let’s attempt to synthesize the MFH method

>> for W = 1:5
synthesis_mfh(47,1e-4,W,1e3);
end
----------------------------------------------------------------------------------
Running moving finite horizon algorithm with:
epsl_inf = 0.0001 | W = 1 | maxIt = 1000 .
Convergence reached with: epsl_inf = 0.0001 | W = 1 | maxOLIt = 1000
A total of 185 outer loop iterations were run.
----------------------------------------------------------------------------------
Trace MFH: 3696.130373
Trace MFH/C: 4.494934
For W_ss = 1 (tr = 3696.13) elapsed time is 1191.84 seconds.
----------------------------------------------------------------------------------
Running moving finite horizon algorithm with:
epsl_inf = 0.0001 | W = 2 | maxIt = 1000 .
Convergence reached with: epsl_inf = 0.0001 | W = 2 | maxOLIt = 1000
A total of 192 outer loop iterations were run.
----------------------------------------------------------------------------------
Trace MFH: 2750.737752
Trace MFH/C: 3.345224
For W_ss = 2 (tr = 2750.74) elapsed time is 2805.09 seconds.
----------------------------------------------------------------------------------
Running moving finite horizon algorithm with:
epsl_inf = 0.0001 | W = 3 | maxIt = 1000 .
Convergence reached with: epsl_inf = 0.0001 | W = 3 | maxOLIt = 1000
A total of 199 outer loop iterations were run.
----------------------------------------------------------------------------------
Trace MFH: 2617.781922
Trace MFH/C: 3.183534
For W_ss = 3 (tr = 2617.78) elapsed time is 5026.36 seconds.
----------------------------------------------------------------------------------
Running moving finite horizon algorithm with:
epsl_inf = 0.0001 | W = 4 | maxIt = 1000 .
Convergence reached with: epsl_inf = 0.0001 | W = 4 | maxOLIt = 1000
A total of 194 outer loop iterations were run.
----------------------------------------------------------------------------------
Trace MFH: 2572.792649
Trace MFH/C: 3.128822
For W_ss = 4 (tr = 2572.79) elapsed time is 8524.05 seconds.
----------------------------------------------------------------------------------
Running moving finite horizon algorithm with:
epsl_inf = 0.0001 | W = 5 | maxIt = 1000 .
Convergence reached with: epsl_inf = 0.0001 | W = 5 | maxOLIt = 1000
A total of 196 outer loop iterations were run.
----------------------------------------------------------------------------------
Trace MFH: 2531.702569
Trace MFH/C: 3.078851
For W_ss = 5 (tr = 2531.7) elapsed time is 12877.1 seconds.

Note that the MFH method converges and has stable errror dynamics

>> model = 47;
>> for W = 1:5
stability_mfh(model,W);
end
MFH synthesis is stable.
MFH synthesis is stable.
MFH synthesis is stable.
MFH synthesis is stable.
MFH synthesis is stable.

Conclusions:

  • The FH method sometimes fails to converge to a constant gain
  • In those cases the MFH still coverges

References

[1] Pedroso, L. and Batista, P., 2023. Decentralized moving horizon estimation for large-scale networks of interconnected unconstrained linear systems. IEEE Transactions on Control of Network Systems, doi: 10.1109/TCNS.2023.3244086.

[2] Viegas, D., Batista, P., Oliveira, P. and Silvestre, C., 2018. Discrete-time distributed Kalman filter design for formations of autonomous vehicles. Control Engineering Practice, 75, pp.55-68. doi:10.1016/j.conengprac.2018.03.014.

[3] Farina, M., Ferrari-Trecate, G. and Scattolini, R., 2010. Moving-horizon partition-based state estimation of large-scale systems. Automatica, 46(5), pp.910-918.