Abstract
In this paper, we consider the use of a linear periodic controller (LPC) for the control of linear time-invariant (LTI) plants in the decentralized setting with an \(H_{\infty }\)-performance criterion in mind. Here we show that if the plant is centrally stabilizable and detectable, the graph associated with the plant is strongly connected, and a technical condition on the relative degree holds, then we can design a decentralized LPC to provide a level of \(H_{\infty }\) performance as close as desired to that provided by a given (stabilizing) LTI centralized controller; this will be the case even if the plant has an unstable decentralized fixed mode (DFM). This means, in particular, that under the listed conditions, we can design a decentralized LPC to provide near \(H_{\infty }\)-optimal centralized performance.
Similar content being viewed by others
Notes
For completeness, we’d like to point out that there are specific application problems such as that of controlling a platoon of vehicles or that of networked control, e.g. see [24], in which \(u_i\) depends not only on \(y_i\) but also on a delayed version of the rest of y, but this is beyond the scope of our setup.
Actually, now we only need the system to be centrally stabilizable and detectable.
The set \(\text{ sp } (A + B K C_2 )\) is the set of eigenvalues of \(A + B K C_2\).
We probe in channel i during \([T_i , T_{i+1} )\).
We probe with elements of \(\sqcap _i [k+1]\) during \([\tilde{T}_i, \; \tilde{T}_{i+1} )\).
One plant output is \(\mathcal {F}(P,K_\mathrm{cen}) r\) and the other plant output is \(\mathcal {F}(P,K_\mathrm{dec} (T)) r\).
References
Aghdam AG, Davison EJ (2006) Structural modification of systems using discretization and generalized sampled-data holds. Automatica 42:1935–1941
Anderson BDO, Moore JB (1981) Time-varying feedback laws for decentralized control. IEEE Trans Autom Control 26:1133–1139
Chen T, Francis BA (1995) Optimal sampled-data control systems. Springer, London
Corfmat JP, Morse AS (1976) Decentralized control of linear multivariable systems. Automatica 12:479–495
Davison EJ, Zhang H (1995) Design of low order compensators for large scale decentralized systems. In: IFAC symposium on large scale systems: theory and applications. London, pp 509–514
Davison EJ, Wang SH (1973) On the stabilization of decentralized control systems. IEEE Trans Autom Control 18:473–478
Doyle JC, Glover K, Khargonekar PP, Francis BA (1989) State-space solutions to standard \({\cal{H}}_2 \) and \({\cal{H}}_{\infty } \) control problems. IEEE Trans Autom Control 34:831–847
Gong Z, Aldeen M (1997) Stabilization of decentralized control systems. Math Syst Estim Control 7:1–16
Khargonekar PP, Ozguler AB (1994) Decentralized control and periodic feedback. IEEE Trans Autom Control 39:877–882
Kobayashi L, Yoshikawa T (1982) Graph-theoretic approach to controllability and localizability of decentralized control systems. IEEE Trans Autom Control 27:1096–1108
Lessard L, Lall S (2011) A state-space solution to the two player decentralized optimal control problem. In: 2011 Allerton conference on communications, control and computing
Miller DE (2006) Near optimal LQR performance for a compact set of plants. IEEE Trans Autom Control 51:1423–1439
Miller DE (2016) A comparison of LQR optimal performance in the decentralized and centralized settings. IEEE Trans Autom Control 61:2308–2311
Miller DE, Chen T (2002) Simultaneous stabilization with near optimal \(H_{\infty }\) performance. IEEE Trans Autom Control 47:1986–1998
Miller DE, Davison EJ (2012) An algebraic characterization of quotient decentralized fixed modes. Automatica 48:1639–1644
Miller DE, Davison EJ (2014) Near optimal LQR control in the decentralized setting. IEEE Trans Autom Control 59:327–340
Miller DE, Qiu L (2005) Near optimal \(H_{\infty }\) tracking for a compact set of plants. In: Proceedings of the IEEE 2005 conference on decision and control, Seville, Spain
Ozguner U, Davison EJ (1985) Sampling and decentralized control systems. In: Proceedings of the American control conference, pp 257–262
Qi X, Salapaka MV, Voulgaris PG, Khammash M (2004) Structured optimal and robust control with multiple criteria: a convex solution. IEEE Trans Autom Control 49:1623–1640
Rotkowitz M, Lall S (2006) A characterization of convex problems in decentralized control. IEEE Trans Autom Control 51:274–286
Ranganathan T, Miller DE (2015) Near optimal \({\cal{H}}_{\infty }\) performance in the decentralized setting. In: 2015 IEEE conference on control applications, Sydney, Australia, pp 1823–1828
Shah P, Parrilo PA (2010) \(H_2\)-optimal decentralized control over posets: a state space solution for state-feedback. In: 49th IEEE conference on decision and control, Atlanta, Georgia, pp 6722–6727
Swigart J, Lall S (2010) An explicit state-space solution for a decentralized two-player optimal linear-quadratic regulator. In: 2010 American control conference, pp 6385–6390
Voulgaris PG (1999) Control under structural constraints: an input–output approach. In: Lecture notes in control and information sciences, pp 287–305
Vale J, Miller DE (2008) Tracking in the presence of a time-varying uncertain gain. In: Proceedings of the 2008 Allerton conference on communication, control, and computing, Allerton, Illinois
Wang SH (1982) Stabilization of decentralized control systems via time-varying controllers. IEEE Trans Autom Control 27:741–749
Willems JL (1989) Time-varying feedback for the stabilization of fixed modes in decentralized control. Automatica 25:127–131
Acknowledgements
This work was supported by the Natural Sciences and Engineering Research Council of Canada via a research grant.
Author information
Authors and Affiliations
Corresponding author
Additional information
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Appendix
Appendix
Proof of Lemma 1
Fix \(\bar{n}\in \mathbf{N}\) and \(\tilde{h}\in (0,1)\). Let \(t_0\in \mathbf{R}\), \(x_0\in \mathbf{R}\), \(h\in (0,\tilde{h})\), \(\bar{u}\in \mathbf{R}\) and \(\phi \in \mathbf{R}\) be arbitrary, and choose u(t) as indicated. If we solve the plant equation (1) and use the Cauchy–Schwarz inequality, then we obtain
Clearly there exists a constant \(\gamma _1>0\) such that
which yields the second bound.
Using the fact that \(\bar{y}_i(t) = v_i^{\mathrm {T}}y_i(t)\) for \(t\in [t_0,t_0+\bar{n}h)\) we have
Before we proceed further, let us get a bound on the term \( \mu _1(t)\). Using the Cauchy–Schwarz inequality, we have
Note that
so we see that \(\bar{C}_2 A^i E = 0\) for \(i=0,1,\ldots , \eta _2-2\), so there exists a constant \(\gamma _2\) so that
which means that there exists a constant \(\gamma _3\) so that
Now we sample \(\bar{y}_i\) in the interval of \([t_0, t_0+\bar{n}h)\) and form \(\bar{\mathcal Y}_i(t_0)\); we see that there exists a constant \(\gamma _4\) and a function \(\mu _2(h)\) so that
with
If we analyse \(\bar{y}_i(t)\) for \(t\in [t_0+\bar{n}h, t_0+2\bar{n}h)\) in a similar fashion and form \({\bar{\mathcal Y}}_i(t_0+\bar{n}h)\), then we see that there exists a constant \(\gamma _5\) and a function \(\mu _3(h)\) so that
with
Using the fact that \(\bar{B}_j = B\bar{w}_j\) and subtracting \({\bar{\mathcal Y}}_i(t_0+\bar{n}h)\) from \({\bar{\mathcal Y}}_i(t_0)\) we obtain
so
We can form a bound on \(\left\| x_0 - x(t_0+\bar{n}h) \right\| \) using (60) and bounds on \(\mu _2(h)\) and \(\mu _3(h)\) using (61) and (62), respectively. Using the fact that \(\Vert H(h)^{-1} \Vert = \frac{ \bar{n}!}{h^{\bar{n}}}\) and \(S^{-1} = {\mathcal O} (1)\), we see that there exists a constant \(\gamma _6\) so that
which provides the first bound. \(\square \)
Proof of Lemma 2
We will model the proof on that of Lemma 2 of [16]. Here our objective is to show the existence of a LPC and we will not attempt to find one of the lowest order.
Since the controller is LPC, it is sufficient to look at what happens on the interval \([kT,(k+1)T)\). In channel i, we partition the state into three sub-states: \(\psi _i^1\), \(\psi _i^2\) and \(\psi _i^3\). We use \(\psi _i^1\) of dimension \({ql_i}\) to store \(\{y_i(kT),y_i(kT+h),\ldots ,y_i(kT+(q-1)h)\}\). The second sub-state \(\psi _i^2\) of dimension \({m_i}\) stores \(\hat{\sqcap }_i[k]\). The last sub-state \(\psi _i^3\) of dimension \(\ell \) only comes into play in channel p: \(\psi _p^3\) stores \(\nu [k]\).
We limit our analysis to the interval of [0, T) as it can easily be extended due to the periodic nature of the controller. Let \(e_i \in \mathbf{R}^q\) denote the ith normal vector. We set
It is clear that for \(j=0,1,\ldots ,q-1\), the vector \( \left[ \begin{array}{c} \psi _i^1[j] \\ y_i(jh) \end{array} \right] \) contains all the elements \(\{y_i(0),y_i(h),\ldots ,y_i(jh)\}\).
Now we turn to the second sub-state. Set
If we initialize \(\psi _i^2[0]\) to be \(\hat{\sqcap }_i [0]\), then we see that
From the formula for \(\hat{\sqcap }_i [1]\) given in the controller description, we see that it is a weighted sum of the elements of \(\{ y_i (0) , \; y_i (h) ,\ldots , \; y_i ((q-1)h) \}\); using the fact that these elements are contained in \( \left[ \begin{array}{c} \psi _i^1[q-1] \\ y_i ( ( q-1)h) \end{array} \right] \), it follows that we can define \(L_i^{21}[q-1]\) and \(M_i^2[q-1]\) so that
We conclude that (16) holds.
Now we turn to the third sub-state. With \(q_p =2\bar{n}(l-l_p)\) we set \((L_i^{31},L_i^{33},M_i^3)[\cdot ] = 0\) for \(i =1,\;2,\ldots ,\;p-1\) and
From the formula given for \(\hat{y}_i (kT)\) in the controller description, we see that \(\hat{y} (0)\) is a weighted sum of the elements of \(\{ y_p (0) , \; y_p (h) , \cdots , \; y_p (q_p h) \} \); using the fact that these elements are contained in \( \left[ \begin{array}{c} \psi _p^1 [q_p ] \\ y_p ( q_p h ) \end{array} \right] \), we can choose \(L_p^{31}[q_p]\) and \(M_p^3[q_p]\) such that
So if we initialize \(\psi _p^3[0]=\nu [0]\), we see that
as required.
At this point, we have defined \(L_i\) and \(M_i\) of \(K_\mathrm{dec}\). It remains to define the time-varying matrices \(Q_i\) and \(R_i\) related to the output. To this end, we define \(\phi _i[j]= \left[ \begin{array}{c} \psi _i[j] \\ y_i(jh) \end{array} \right] \). It is straightforward to verify that \(\phi _i[j]\) contains \(\{\hat{\sqcap }_i[0],y_i(0),\ldots ,y_i(jh)\}\) for \(j=0,1,\ldots , q-1\); for the case of \(i=p\), \(\phi _i[j]\) also contains
-
(i)
For \(j=0,\;1,\ldots ,\;q_p-1\), the control signal \(u_i(jh)\) equals \(\hat{\sqcap }_i[0]\) plus a linear combination of the elements of \(y_i(0)\). Since \(\hat{\sqcap }_i[0]\) and \(y_i(0)\) are contained in \(\phi _i [j]\), we see that \(u_i(jh)\) is a linear combination of the elements of \(\phi _i[j]\) for \(j=0,1,\ldots ,q_p-1\).
-
(ii)
At \(j=q_p\), \(u_i(j) = \hat{\sqcap }_i[0]\) and \(\hat{\sqcap }_i[0]\) is contained in \(\psi _i[j]\), so it is a linear combination of the elements of \(\phi _i(j)\).
-
(iii)
Let \(j \in \{ q_p+1,\,\ldots ,\;q-1 \}\). The control signal \(u_i(jh)\), \(i=1,\ldots ,p-1\), equals \(\hat{\sqcap }_i[0]\), which is contained in \(\psi _i[j]\), so it is a linear combination of the elements of \(\phi _i[j]\). The control signal \(u_p(jh)\) equals \(\hat{\sqcap }_p[0]\) plus a linear combination of the elements of \(\sqcap [1] = H \nu [1] + J \hat{y} (0) \); but \(\nu [1]\) lies in \(\phi _p [j]\) and \(\hat{y} (0)\) is a linear combination of the elements of \(\phi _p [j]\), so \(\sqcap [1]\) is a linear combination of the elements of \(\phi _p[j]\).
Using the fact that \(u(t) = u(jh)\) for \(t\in [jh,(j+1)h)\), \(j\in \mathbf{Z}^+ \), we conclude that for \(j=0,1,\ldots ,q-1\), the control signal \(u_i(jh)\) is a linear combination of the elements of \(\phi _i[j]\), which means that \(Q_i\) and \(R_i\) can be defined so that the controller (5) is identical to that of (9)–(14). \(\square \)
Proof of Lemma 3
Suppose that the controller (9)–(14) is applied to the plant (1), and let \(x_0\in \mathbf{R}^n\), \(\nu [0]\in \mathbf{R}^\ell \), \(\hat{\sqcap }[0]\in \mathbf{R}^m\), \(k\in \mathbf{Z}^+\) and \(T>0\) be arbitrary. With the estimate \(\hat{y}(kT)\) of y(kT) given by (10) we see that
We can use Lemma 1 to derive the estimation error of each individual quantity \([y_i (kT) ]_j \) for \(i\in \{1, 2, \ldots , p-1\}\) and \(j\in \{1, 2, \ldots , l_i\}\). More specifically, extending the error bound provided in (8) to the general case, we have
We now obtain a bound on the first term of the RHS: for \( t \in [kT, kT + 2 \bar{n}( l - l_p) h )\), we have
so it is easy to see that
Using this we can simplify (64):
which yields (21).
Now let us find a conservative bound on the signal u(t). Observe that
Using (66) to bound \(\hat{y}(kT)\) and the fact that \(\nu [k+1] = \nu [k]+{\mathcal {O}}(T)\nu [k]+{\mathcal {O}}(T)\hat{y}(kT)\), this simplifies to
which yields (19).
Now we return to the state x(t). Using (67) in (65), we can obtain a bound on \(\Vert x(t) - x(kT) \Vert \):
for \(t \in [kT, (k+1) T ) ,\) which yields Eq. (18).
We now perform a similar analysis as above to obtain a bound on the quantity \(\hat{\sqcap }[k+1]-\sqcap [k+1]\). We first use Lemma 1 to obtain the estimation error of each element of \(\hat{\sqcap }_i[k+1]_j\) for \(i\in \{1,2,\ldots ,p-1\}\) and \(j\in \{1,2,\ldots ,m_i\}\). More specifically, we use the bound in (8) with a minor change reflecting the fact that we are now probing with \(T^\delta \bar{w}_p\sqcap _i[k+1]_j\) rather than \(T^\delta \bar{w}_i[y_i(kT)]_j\). So (8) becomes (for \(i=j=1\)):
As a result, we end up with
But it is easy to see that
After we use the bound on \(\hat{y}(kT)\) given by (66) to simplify this, we substitute the resulting expression for \(\sqcap [k+1]\) into (69) and use (68) to obtain a bound on \(\max \limits _{\tau \in [kT,(k+1)T]}\left\| x(\tau ) \right\| \), yielding
which yields (20). \(\square \)
Rights and permissions
About this article
Cite this article
Ranganathan, T., Miller, D.E. Near optimal \(\mathcal {H}_{\infty }\) performance in the decentralized setting. Math. Control Signals Syst. 30, 12 (2018). https://doi.org/10.1007/s00498-018-0217-1
Received:
Accepted:
Published:
DOI: https://doi.org/10.1007/s00498-018-0217-1