Survey
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
Mathematical optimization wikipedia , lookup
Computational fluid dynamics wikipedia , lookup
Perturbation theory wikipedia , lookup
Inverse problem wikipedia , lookup
Computational electromagnetics wikipedia , lookup
Multiple-criteria decision analysis wikipedia , lookup
Operational transformation wikipedia , lookup
Perceptual control theory wikipedia , lookup
Hendrik Wade Bode wikipedia , lookup
Distributed control system wikipedia , lookup
Control theory wikipedia , lookup
H∞ Control of System With I/O Delay: A Review of Some Problem-Oriented Methods Leonid Mirkin Faculty of Mechanical Engineering, Technion — IIT, Haifa 32000, Israel Gilead Tadmor Electrical & Computer Engineering, Northeastern University, Boston, MA 02115 [Received April 12, 2001] Systems with input or output delays form the simplest, and yet one of the most widely applied classes of distributed parameter models. This is a review of some problem-oriented H∞ methods for that class, with an emphasis on computational simplicity. Reviewed methods include operator interpolation, game-theoretic state-space treatments in the time domain, a J-spectral factorization approach, and methods exploiting various ideas from sampled-data theory. Some interesting properties of the (sub)optimal solutions are discussed as well. 1. Introduction This article reviews several H∞ methods for rational systems cascaded with input or output delays, see Fig. 1. Those are perhaps the simplest distributed parameter systems, and yet the most widely used as models in industrial applications. Hence the interest in methods developed specifically for that class. In view of the voluminous literature on robust distributed systems design, articles cited here are intended merely as samples, and by no means provide an exhaustive, or even a representative list. Historical perspectives. Solutions of H∞ problems in distributed systems trace back to the mid 1980’s. Some early work [21, 17, 18, 64, 82] appealed to operator interpolation/commutant lifting theorem methods [61, 63]. One major outgrowth, - w u z - P Kτ ? −τ s e y Fig. 1. General LFT setup with single delay 2 L. Mirkin & G. Tadmor with later impact on nonlinear robust control, was the development of the SkewToeplitz methodology [20, 4, 36, 52, 51, 53, 49, 19]. Yet another approach followed early on, was the use of reduction to a finite dimensional (rational) problem [10, 33, 24, 25, 56, 83, 29]. The introduction of time domain/game theoretic state space methods [66, 59, 80, 81, 68, 3] was immediately followed by extensions to a general framework of distributed parameter systems [67, 9, 70, 76, 32, 37]. A feature that facilitates design in systems with input (or output) delays is that their distributed component has very simple dynamics (essentially, an FIR). Indeed, in “classical” and some robust stabilization problems, this feature allows a reduction to a purely rational problem [35, 1, 71], using a Smith predictor. Scope and organization of the paper. The common thread in the several reviewed methods is the utilization of the special structure and time domain representations of systems with I/O delay. The review does not cover the use of tools of broader scope, such as low order approximations, or general factorization techniques, as well as a variety of interesting ideas leading to simplified transform-domain solutions (such as elegant reproducing kernels methods [16], or the simplified computation of Hankel singular values [50], which requires an inner/outer factorization). Some early work meet the set criterion in the context of operator interpolation methods and will be discussed in Section 2. Here, time domain analysis is used to characterize the interpolation conditions for one-block problems. Sections 3 and 4 are devoted to time- and frequency-domain treatments of the “standard problem” in Fig. 1. §3.1 reviews the use of a “lifting” trick, borrowed from work in sampled data control. It transforms the delay problem to a discrete time problem with distributed I/O operators, but with no delays. The game theoretic solution leads to a periodic, time varying compensator. Realizing that periodicity is clearly a technical artifact of “lifting”, immediate followups - reviewed in §3.2 - removed the resulting time variation. Here one considers the interplay between abstract evolution models over the product space M2 = Rn ×L2 [−τ, 0] and the original delay equation, in the associated differential games. A J-spectral factorization based approach is reviewed in §4.1. Here the Smith predictor trick is used to reduce the problem to an equivalent rational problem. The extraction of dead-time controllers from delay-free parameterizations is discussed in §4.2. This approach also has its roots in the sampled-data theory and leads to a simple and intuitively appealing solution. Finally, the controller structure and the achievable cost in the H∞ delay problems are discussed in Section 5. Preliminaries and notation. Given an LTI system G(s) = C(sI − A)−1 B, its 0 conjugate is defined as G∼ = −B − A0 )−1 C 0 . The completion of e−τ s G to [0, τ ] (sI . −τ s −Aτ is defined as follows: π e G = C(e − e−τ s I)(sI − A)−1 B. It can be seen τ −τ s that π τ e G is an entire function whose impulse response has support on [0, τ ] (FIR system). H∞ control of system with i/o delay: review 3 2. Time-domain methods in operator interpolation Arguably the most significant early development in robust H∞ control was the connection made with operator interpolation/commutant lifting problems [61, 63], involving the following general setting. Given are w, m ∈ H∞ , with m(s) inner. One defines a subspace K := H2 m H2 , the orthogonal projection π : H2 7→ K, and an operator T : K 7→ K, via T f = π(wf ), f ∈ K. A function w0 ∈ H∞ is said to interpolate the operator T if it satisfies: (i) T f = π(w0 f ), f ∈ K, and (ii) kT k = kw0 k∞ = kΛw0 k, where Λw0 : H2 7→ H2 is the multiplication operator associated with w0 . In the context of control design, “m” is usually the inner part of the open loop plant, whereas “w” is a frequency shaping weight function [14, 22]. In the interest of simplicity, w can be selected rational, yet non-rational, non-minimum phase elements in a distributed plant, such as delays, cannot be removed from m. The following is a key result Lemma 1 ([61, Proposition 5.1]) Suppose that f ∈ K is a maximal function for the operator T . Then w0 = T f /f ∈ H∞ is an all-pass function, and is the unique solution of the interpolation problem. The main solution steps are thus the characterization of K, π, and T . Early case studies of distributed systems [21, 17, 18, 82, 25] concerned the example of scalar rational systems with a pure input delay. The distributed non-minimum phase part is already inner, m(s) = e−τ s , leading to a particularly simple characterization of K, π and T : assuming an outer rational part, K = H2 e−τ s H2 is isometric L2 [0, τ ], π : L2 [0, ∞) 7→ L2 [0, τ ] is the truncation and T is the restriction to [0, τ ] of convolution with the inverse Laplace transform of w. Thus, T is realized by a causal ordinary differential equation (ODE), its adjoint has an anti causal ODE realization, and Schmidt pairs of T are solutions of a Hamiltonian, two point boundary value problem. Once that observation is made, the step leading to a solution based on transcendental equation characterization of Schmidt pairs, is immediate. The pleasant feature of an inner distributed component fails in the next simplest example of a scalar system with several, commensurate input delays, and neither an inner/outer plant factorization, nor a characterization of K in terms of the plant’s inner part, would then be as simple. The solution outlined in [65, 64] utilizes the fact that the plants outer part has no effect on the definition of K, and on Fredholm’s alternative, whereby K = H2 gH2 = N (Λ0g ), with g being the plant’s transfer function. The main point is that the adjoint multiplication operator, Λ0g , has a system realization and can be characterized in terms of that realization. In a plant with multiple commensurate input delays and an outer rational part, it suffices to consider g as the delay polynomial. Then K is the stable component of the solution space of a difference equation and is isomorphic to initial L2 [0, lτ ] portions of trajectories, A new challenge is to define the right Hilbert space structure in K, to make this isomorphism an isometry, as is needed in order to compute the orthogonal projection π, and adjoint of operators over K, as follows. Let S be the lτ -time units shift along trajectories in K. Identifying trajectories l i f ∈ L2 [0, ∞) ↔ {f i }∞ i=0 ∈ `2 (L2 ([0, τ ], R )), where f (s) = [f (ilτ + s), f ((il + T 1)τ + s), . . . , f (((i + 1)l − 1)τ + s)] , s ∈ [0, τ ], i = 0, 1, . . . , the shift S admits a 4 L. Mirkin & G. Tadmor representation by an l × l matrix E with eigenvalues magnitude smaller than one. Let Q > 0 solve Q − E 0 QE = I. Then, inner products between trajectories in K satisfy hx, yiL2 [0,∞) = ∞ X hxk , y k iL2 [0,τ ] = k=0 ∞ X hE k x0 , E k y 0 iL2 [0,τ ] = hx0 , Qy 0 iL2 [0,τ ] k=0 The Q - weighted L2 ([0, τ ], Rl ) inner product thus defines the isometric representation of K by [0, τ ] trajectories. The projection π : L2 [0, ∞) 7→ K is P∞ k then [πf ]i = E i Q−1 k=0 E 0 f k . The operator T = πΛw |K is the compression of a convolution operator; here too, it admits a causal ODE realization over [0, τ ], albeit subject to certain non-trivial (integral) boundary conditions, its adjoint T 0 has an anti causal ODE realization and the equation (σ 2 I − T 0 T )f = 0 amounts to a two point boundary value problem. Thus singular values are again solutions of a transcendental equation and Schmidt pairs are derived from algebraic conditions on admissible boundary values. Stressed in closing is the key step of utilizing the time domain structure of the delay operator to provide a time domain realization of the interpolation space K as the solution space of a difference equation. This leads to a problem reduction to a solution of a two point boundary value problem. 3. Time-domain methods By the late 1980’s, state space / Riccati equations methods became dominant, beginning with the finite dimensional LTI case [2, 15], continuing with time varying (LTV) / finite time systems, [67, 70, 3] and eventually allowing extensions to distributed systems. Practical impediments in the latter include the need to solve infinite dimensional operator Riccati equations, and compensator realizations in terms of abstract evolution models. Here we review methods utilizing the structure of systems with a pure input lag, to address both issues: operator Riccati equations are reduced to fixed size matrix equations and explicit delay system compensator realizations are provided. Early developments are reported in [3]. The focus of this review is on the subsequent results in [46, 47], followed by [34, 72, 74, 73, 71]. 3.1 A sampled data method: matrix Riccati eqs. and “periodic” solutions Several control and observation problems are considered in [47], and reduced (using standard unitary loop transformations) to a single representative, akin to the “standard problem” in the context of Figure 1. Roughly, the idea here is to break the basic problem into two parts, one of which is just a delay-free estimation of z over [τ, ∞) and another one can be reduced to a Nehari-like problem defined over the finite horizon [0, τ ]. Nagpal and Ravi [47] observed an opportunity in adapting the so-called “lifting” technique similar to that used for some sampled-data control problems [8]. In one guise or another, the various approaches to sampled data design built on the “lifting” trick (already used in in §2) where, in I/O trajectories one identifies f ∈ L2 [0, ∞) ↔ {f k } ∈ `2 (L2 [0, τ ]), with f k (s) = f (kτ + s). This reduces H∞ control of system with i/o delay: review 5 the original (analog) system to a discrete time system with the same state space, say Rn and distributed I/O signals. Two algebraic Riccati equations (AREs) determine a discrete time design, which leads to periodic actuation between samples. Since the system acts in open loop between successive decisions, an added differential Riccati equation (DRE) is also needed, in some variants, to evaluate the I/O gain during these uncontrolled periods. The adaptation of these ideas to the context of the basic delay problem required sophisticated and elaborate computations, and we shall be content with highlights of some important characteristics of [46, 47]. First, as in sampled data control, simple solvability conditions are derived, comprising the AREs for the delay-free problem and a DRE over [0, τ ]. The latter quantifies the extra cost for tolerating delay. In contrast, the controller structure is cumbersome, even in the simplest case. It is derived in terms of the “lifted” representation, and, as mentioned earlier, its “unlifted” form is periodic, time varying. The challenge of “peeling-off” a timeinvariant representation is addressed next. 3.2 A semigroup based reduction: matrix Riccati equations and LTI solutions Inspired by [46, 47], the studies [34, 72, 74, 73, 71] searched for an alternative reduction that would lead to a similar computational component (i.e., the combining AREs and DRE), but that avoids compensator periodicity. Following are the guiding principles, underlying these developments. The first is the use abstract evolution models for the delay system [30, 11, 5, 13, 77, 6, 7, 60, 12, 75]. A natural state in the current setting would be f (t) = (x(t), ut ) ∈ M2 := Rn × L2 [−τ, 0], where ut (s) = u(t + s), s ∈ [−τ, 0], is the relevant control history at the time t. Abstract state space/operator Riccati equations solutions are well known for the linear quadratic (LQ) optimal control problem [78, 31, 57, 58] and have been established also for H∞ problems [67, 9, 70, 76, 32, 37]. As in the non-distributed case, the solution of the operator Riccati equation (ORE) is completely characterized by the quadratic form for the optimal cost in terms of the initial state. In the H∞ case, that is the cost of an associated differential game. Second, as in [46, 47] the explicit solution of the said game is computed in the original setting, based on fixed size matrix AREs/DRE. A positive semidefinite quadratic form for the game’s optimal value is then derived as a function of the initial data f (0) = (x(0), u0 ) ∈ M2 . Returning to the abstract model, that quadratic form provides the desired explicit expression for the solution of the ORE. Third, in complete analogy to finite dimensional results, suboptimal compensators are parameterized by state space equations, in an abstract evolution model. These state space equations are based on perturbations of the infinitesimal generator and of input and output operators from the original model. As it turns out, these perturbed generators are, by themselves, abstract models of delay differential equations. The tools used here are provided by the detailed analysis of the interplay between functional differential equations and their abstract models, in [6, 75]. In summary, the family of solutions described in this section comprise an analysis along two parallel tracks. Along one track, a formal solution of the H∞ 6 L. Mirkin & G. Tadmor problem, including a parameterization of suboptimal compensators, is provided in terms of abstract evolution model counterparts of familiar state space solutions. Along a second track, explicit solution of operator Riccati equations is reduced to matrix computations, via a solution of a differential game in the original delay equation setting. Finally, the abstract evolution model realization of suboptimal compensators is reduced to a delay (neutral) differential equation, using established relations between such equations and their abstract models, and specific details of the computed solution of the operator Riccati equation. In order to provide a flavor of this set of result, the following is an abbreviated version of the main result in [72]. Theorem 1 Consider a system G in minimal LTI realization [A, B, C], cascaded with the input delay e−τ s . Let X, Y > 0 be the stabilizing solutions of the algebraic Riccati equations A0 X + XA − XBB 0 X + C 0 C = 0, AY + Y A0 − Y C 0 CY + BB 0 = 0. (1) I Let γ0 be the infimal achievable induced norm of FK = K (I + KG)−1 [I K] over all stabilizing linear feedback compensators K. Then γ0 > 1, and γ > γ0 if and only if there exists a bounded, positive semidefinite solution of the following differential Riccati equation over [0, τ ] Ż = ZA + A0 Z + 1 0 1 γ 2 −1 ZC CZ + BB 0 , Z(0) = Y (2) 1 such that (γ 2 − 1)I > X 2 Z(τ )X 2 . Given such such γ and the associated Z, a parameterization of stabilizing compensators that guarantee that performance level is in terms of ẋc (t) = Axc (t) − Bv(t − τ ) + Y C 0 (e(t) − Cxc (t)) R0 v(t) = U 0 xc (t) + −τ U 1 (s)v(t + s)ds + b(t) R0 c(t) = V 0 xc (t) + −τ V 1 (s)v(t + s)ds − e(t); b = K 4c where the compensator’s input is the signal “e”, where U 0 , V 0 , U 1 (·) and V 1 (·) are defined in terms of the system matrices and the solutions to the said Riccati equations,as specified in the article, and where K 4 is a stable system with an induced norm smaller than γ 2 − 1. 4. Frequency-domain methods We review two transform-based methods, where state space machinery is restricted to computational purposes. Again, the emphasis is on exploitation of the special structure of the delay block to simplify solutions. 4.1 J-spectral factorization and the Smith predictor A natural way to address delay is via the Smith predictor [62]. The predictor compensates for the delay with a specially constructed internal feedback loop in the H∞ control of system with i/o delay: review 7 controller (see §5.1). Consequently, it reduces some dead-time problems to delayfree forms [54]. This idea, however, had little impact on H∞ methods. While a controller structure, reminiscent of the Smith predictor, can be recovered from some H∞ frequency-domain solutions (e.g., [19]), and the possibility to recast controllers in a generalized Smith predictor form was pointed out by Naimark and Cohen [48], the idea to use a special controller transformation to reduce a problem to an equivalent finite-dimensional one had not been used in the H∞ literature till the work of Meinsma and Zwart [38, 40, 39], where it was applied first to a mixed sensitivity (two-block) problem, and then generalized to the standard (four-block) H∞ problem with a single delay. Here are some details of the developments in [40, 39]. The core step is a Jspectral factorization: given a self-adjoint transfer matrix M , find a (bi-proper, outer, etc) transfer matrix L satisfying M1 M3∼ I 0 M= = L∼ L (3) M3 M2 0 −I Solutions are well understood in the rational case [27]. In dead-time problems the diagonal blocks remain rational, and M3 = e−τ s Mr,3 , with rational Mr,3 . This structure can be exploited in solving (3), as follows. The transfer matrix M3 is (nonuniquely) expanded as e−τ s M2−1 Mr,3 = R − ∆s where R is rational and ∆s ∈ H∞ . In these terms one has M = Ψ ∼ Mr Ψ where Mr (determined the by rational blocks I 0 of M and by R) is a rational and self adjoint, and Ψ = −∆ is invertible in s I H∞ . Thus, a J-spectral factorization is provided by L = Lr Ψ , where Lr is the rational J-spectral factorization of Mr . Necessary to complete the reduction is an additional (stability) test, involving L, which cannot be expressed in terms of rational matrices. That test, however, involves only solvability conditions and does not affect the resulting controller. To see the connection with the Smith predictor, consider the J-spectral factorization controller. As is well known, proper compensators C, providing γ-suboptimal of solutions the corresponding dead-time H∞ problem are those satisfying I −C L−1 Q = 0, for some Q ∈ H∞ , kQk∞ < γ. Substituting I −1 the explicit form of L into this −1 Q expression, C = Cr (I + ∆s Cr ) , where the rational Cr satisfies I −Cr Lr I = 0, i.e., it is the controller corresponding to the “rationalized” version of (3). Thus, the resulting controller has the Smith predictor structure where ∆s serves as a dead-time compensator block (in fact, C above can be presented in the form shown in Fig. 2 on p. 9, where Gr depends on Lr only). Note, that thechoice of ∆s and. R is non unique. In [40] ∆s was chosen in the form ∆s = π τ e−τ s Pa for Pa = M2−1 Mr,3 . When Pa is stable, R can be chosen either as R = 0 (in the spirit of the Morari’s internal model control) or R = Pa (in the spirit of the Smith predictor). In the latter case Mr is equal to the delay-free (τ = 0) version of M , which means that if Pa is stable, then the dead-time problem is solvable iff so is its delay-free counterpart. Typically, however, Pa is unstable. Comparing with the time domain methods in Section 3, the clear advantage is in a simpler and more transparent controller structure. The disadvantage lies in a somewhat less transparent structure of the solvability conditions, which are based 8 L. Mirkin & G. Tadmor upon two H∞ AREs, one of which depends on the delay τ (in [47] only the coupling condition depends on τ ). 4.2 Extraction of dead-time controller from delay-free parameterizations Conventionally, the delay block e−τ s is treated as a part of a plant. This results in an infinite-dimensional plant, and as seen above, typically, a rather involved controller structure. In [41, 42] it is proposed to treat the delay as a part of the controller. This leads to the design of a constrained-structure controller for a finitedimensional plant. The challenge here is to extract admissible (delay) controllers from the well known parameterization of all suboptimal controllers for the rational plant. The idea is borrowed from [45], where the structure of sampled-data controllers in the lifted domain (see §3.1) was studied. It was shown that a controller can be implemented as the cascade of a generalized sampler, a discrete-time controller and a generalized hold iff its lifted “transfer function” is strictly proper. Using this fact, the sampled-data H∞ standard problem was then solved by extracting strictly proper transfer functions from the lifted version of the DGKF [15] parameterization of all continuous-time H∞ controllers. This procedure yields simple formulae both for the digital part of the controller and for the generalized sampler and hold. Moreover, an attractive byproduct of this procedure is that the solvability conditions are naturally separated into the existence conditions for a continuoustime controller and an additional condition imposed by the sampled-data structure. The constraints imposed upon dead-time controllers in the lifted domain are more complicated than in the sampled-data case. For that reason, a direct extension of the technique of [45] to delay systems is much more involved. The extraction approach, however, can be adopted to dead-time systems by translating the deadtime constraints into the frequency domain. This approach was followed in [42] resulting in simple and intuitively appealing solvability conditions and the controller structure for the dead-time H∞ standard problem. Below, the solution procedure in [42] is outlined. The standard (four-block) H∞ problem with a single delay e−τ s in the loop is considered. The solution starts with the parameterization of all γ-suboptimal controllers K for the rational part of the plant. This parameterization is well-known [15] and has the form of an LFT interconnection of a 2 × 2 block transfer matrix G with an arbitrary Qa ∈ H∞ such that kQa k∞ < γ. The next step is then to extract all Qa which bring the LFT in the form e−τ s Kτ for a proper Kτ . This can easily be done by the loop-shifting arguments of [44] yielding the following characterization of all admissible Qa : Qa = ∆22 + e−τ s Qb , where Qb ∈ H∞ is arbitrary and ∆22 is the FIR truncation of the (2, 2) subblock of G−1 . Now, taking into account the norm constraint on Qa one can see that the original (four-block) problem is reduced to the (one-block) problem of the characterization of all Qb ∈ H∞ such that: k∆22 + e−τ s Qb k∞ < γ. (4) According to the Nehari Theorem (see [19, Theorem 15]), (4) is solvable iff γ is larger than the norm of the Hankel operator with symbol eτ s ∆22 that, in turn, H∞ control of system with i/o delay: review u - ∆s Gr 9 y ?y e i - Q Fig. 2. DTC structure of delay controllers Kτ . is equivalent to the L2 [0, τ ] induced norm of ∆22 . Thus, γ > k∆22 kL2 [0,τ ] is the additional solvability condition, as compared with the delay-free case, capturing the effect of delay on achievable performance. Note, that this condition can be expressed either in terms of the matrix exponential of a Hamiltonian matrix or in terms of the solvability of a differential Riccati equation defined over [0, τ ], see [8, 28], where several computational schemes are discussed (as a matter of fact, when expressed in terms of a Riccati equation, these conditions coincide with those obtained by Nagpal and Ravi [47]). Then, for any admissible γ the set of all solutions to (4) can be parameterized in terms of an arbitrary Q ∈ H∞ such that kQk∞ < γ. Finally, the latter is substituted into the original delay-free parameterization yielding, after some elaborate simplification steps, the characterization of all dead-time controllers solving the original problem. 5. Discussion Perhaps the most serious disadvantage of the early H∞ methods for delay systems is the lack of transparency. Design procedures typically involved several intermediate steps, so the controller structures and the achievable cost are not easily recoverable. On the other hand, the methods described in Sections 3 and 4 enable one to compute both the (sub)optimal controllers and the achievable cost explicitly for general problems in terms of the original data. Among other advantages, this enables one to gain deeper insight into the controller structures and the limitations on performance imposed by delay elements. These issues will be discussed below. 5.1 Controller structure One of the most widely used controller configurations to control delay systems is the so-called dead-time compensator (DTC) depicted in Fig. 2. DTC consists of a rational part Gr and a stable internal feedback ∆s containing the delay (the free parameter Q is typically zero). This scheme was proposed by Smith [62], whose idea was to use the internal feedback with ∆s = P22 − e−τ s P22 to convert dead-time problems to their delay-free versions. The rationale behind this approach becomes . apparent when the signal ye = y + ∆s u entering the rational part of the controller 10 L. Mirkin & G. Tadmor in Fig. 2 is considered. Indeed, it is readily seen that ye = e−τ s P21 w + P22 u + P22 − e−τ s P22 u = P21 e−τ s w + P22 u and thus the resulting feedback loop (u ye ) does not contain any delay and is exactly the feedback loop for τ = 0. In other words, ∆s “compensates” the delay in the loop. The scheme of Smith can be applied to stable systems only. Yet it can easily be modified to cope with the general case by choosing ∆s = P̃22 − e−τ s P22 with any P̃22 making ∆s stable. In particular, it is always possible to make ∆s FIR [79]. Some other possible choices are discussed in [40]. Moreover, as shown in [44] the set of all stabilizing dead-time controllers can be parameterized in the DTC form shown in Fig. 2. In the context of H∞ control, the DTC controller structure was derived by Meinsma and Zwart [40] and Mirkin [41]. The version of the H∞ DTC proposed in [41], in which the emphasis is on transparency, is discussed below (for more details see [43]). The set of all γ suboptimal dead-time controllers is parameterized in [41, 42] in the form depicted in Fig. 2, subject to a rational Gr , similar to that for the delay free Q ∈ H∞ such that kQk∞ < γ, and the FIR problem, an arbitrary . ∼ ∆s = π τ e−τ s Pa , where Pa = Fu P, γ −2 P11 . Unlike the Smith predictor and its modifications, the delay in the loop is now not canceled (since e−τ s (P22 − Pa ) 6= 0). Yet it is easily seen that the relationship y = Pa u can be equivalently written as y = P21 w∗ + P22 u, where the “disturbance” w∗ satisfies: w∗ = ∼ 1 γ 2 P11 P11 w∗ + P12 u . Thus, if the w in Fig. 1 were equal to the w∗ above, then the block disturbance ∆s = π τ e−τ s Pa would just compensate the dead time in both feedback and feedforward loops (“predict” y subject to a given w). This agrees well with the result of Palmor and Powers [55], who proposed to transmit measured disturbances into the Smith predictor to improve its disturbance attenuation properties (feedforward Smith predictor). Although the H∞ DTC does not measure the disturbance, it generates it artificially. Taking into account the worst-case nature of the H∞ methodology, one would expect that the disturbance w∗ is generated on the basis of a worst-case scenario. It turns out that this indeed happens. To see this, consider the relationship y = Pa u in the state-space setting. We have: −λ̇ = A0 λ + C10 z ẋ = Ax + B1 w∗ + B2 u and , y = C2 x + D21 w∗ w∗ = γ12 B10 λ where z = C1 x+D12 u. Using the standard arguments from the calculus of variations [26], the w∗R above can roughly be thought of as the maximizing disturbance for the ∞ index J = 0 (z 0 z − γ 2 w0 w)dt subject to any fixed u. H∞ control of system with i/o delay: review 11 Thus, the H∞ DTC attempts to compensate the dead time h assuming that the disturbance w is the worst-case one for the open-loop problem. The fact that w∗ is worst-case for an open-loop problem agrees well with the open-loop nature of the dead-time compensation. 5.2 The “cost of delay” Another important question raising in H∞ control of delay systems concerns the quantification of the effect of delay on achievable performance. Obviously, any delay problem is solvable only if so is its delay-free counterpart. One therefore would expect that the achievable performance level for any dead-time problem can be presented as a sum of the delay-free performance and a term that reflects the cost of delay. The knowledge of the latter may be of value in numerous applications where delay tolerance is important. Unfortunately, in most early treatments, this “cost of delay” cannot be easily recovered from the solvability conditions. This complicates any further analysis considerably. In this respect, [47, 74, 41] offer a clear advantage since the solvability conditions there are explicitly presented in such a two-term form with a transparent “cost of delay” term. In particular, the following result can be summarized: Lemma 2 Consider the system in Fig. 1 with P11 = C1 (sI−A)−1 B1 . The following three conditions are equivalent (below, X, Y > 0 stand for the stabilizing solutions to the matrix Riccati equations associated with the delay-free problem): (a) There exists a stabilizing Kτ so that kF` P, e−τ s Kτ k∞ < γ (b) The problem is solvable for τ = 0 and, in addition, Π(t) defined on [0, τ ] as Π̇ + ΠA + A0 Π + C10 C1 + 0 1 γ 2 ΠB1 B1 Π = 0, Π(τ ) = X, (5) exists and satisfies ρ(YΠ(0)) < γ 2 ; (c) The problem is solvable for τ = 0 and, in addition, kP11 kL2 [0,τ ] < γ and −1 ρ XΣ12 Σ22 < 1 and ρ (Σ22 − XΣ12 )−1 (Σ21 − XΣ11 )Y < γ 2 , where Σ11 Σ21 Σ12 Σ22 . = exp A −C10 C1 1 0 γ 2 B1 B1 0 −A τ . Condition (b) above is from [47] while condition (c) is from [41]. Although the solvability conditions in [74] appear slightly different, they can be brought to a form equivalent to that in [47] using standard manipulations of Hamiltonian matrices. Some observations are in order. First, it appears that in some situations (small enough) delay does not contribute to the final cost, i.e., any performance achievable with τ = 0 can be achieved with “small” τ as well. Indeed, it is known [23] that the delay-free test fails if either (i) ρ(YX) = γ 2 or (ii) X or Y is no longer stabilizing. Assume for simplicity that the triple (C1 , A, B1 ) is observable and controllable. Then the solution Π(t) of (5) is monotonically increasing as t passes from τ to 0. Therefore, if (i) happens prior to (ii) in the delay-free case, then the condition 12 L. Mirkin & G. Tadmor ρ(YΠ(0)) < γ 2 fails for all τ > 0. Yet if the opposite happens for some γ = γ ∗ , then there exists a τ0 > 0 such that for every γ > γ0 condition (b) of Lemma 2 holds true ∀τ < τ0 . Moreover, if the observability or/and controllability of (C1 , A, B1 ) are relaxed, then small enough delay might not affect the achievable performance even when the spectral radius condition fails first. This, for instance, occurs in the case when A is Hurwitz and P11 = 0, see [44]. Another interesting observation is that exactly the same condition appears in the context of sampled-data control when both sampler and hold are the design parameters [69, 45]. In other words, it turns out that the cost of delay is equal to the cost of sampling. This fact is surprising since the constraints imposed by the delay are more restrictive, than those imposed by the sampling. Indeed, as was mentioned in §4.2, a controller K contains the delay e−τ s iff its kernel k(t, s) = 0 whenever t < s + τ . On the other hand, it follows from [45] that a controller is the cascade of a generalized sampler, a discrete-time controller and a generalized hold with the period τ iff its kernel k(t, s) = 0 whenever t < b τs cτ + τ , where b·c stands for the “floor” (round towards −∞) operation. Thus, the sampled-data controller “receives odds” on the intervals b τs c 6 τt − 1 < τs but is not able to take any advantage of it. A possible explanation may be that on this interval the sampled-data controller can act only in an open-loop fashion. Then, the disturbance is “smart enough” to outplay the control. 6. Conclusions In this paper several problem-oriented methods for the H∞ control of continuoustime systems with a single delay in the feedback loop have been reviewed. Reviewed methods include operator interpolation methods (Section 2) and game-theoretic state-space (Section 3) treatments in the time domain and a J-spectral factorization (Subsection 4.1) and extraction (Subsection 4.2) approaches in the frequency domain. Some properties of the resulting solutions and solvability conditions have also been discussed in Section 6. References [1] Z. Artstein. Linear systems with delayed control: A reduction. IEEE Transactions on Automatic Control, AC – 27:869–879, 1982. [2] J. A. Ball and N. Cohen. Sensitivity minimization in an H∞ norm: parameterization of suboptimal solutions. Int. J. Control, 46:785 – 816, 1987. [3] T. Basar and P. Bernhard. H∞ Optimal Control and Related Minmax Design problems: A Dynamic Game Approach. Birkhauser, 1991. [4] H. Bercovici, C. Foiaş, and A. Tannenbaum. On skew Toeplitz operators, I. Operator Theory, Advances and Applications, 29:21 – 45, 1988. [5] C. Bernier and A. Manitius. On semigroups in Rn × lp corresponding to differential equations with delays. Can. J. Math., 30:897 –914, 1977. H∞ control of system with i/o delay: review 13 [6] J. A. Burns, T. L. Herdman, and H. W. Stech. The cauchy problem for linear functional differential equations. In T. L. Herdman, S. M. Rankin, and H. W. Stech, editors, Integral and Differential Equations, pages 139 – 149. Marcel Dekker, 1981. [7] J. A. Burns, T. L. Herdman, and H. W. Stetch. Linear functional differential equations as semigroups on product spaces. SIAM J. Math. Anal., 14:98 – 116, 1983. [8] T. Chen and B.A. Francis. Optimal Sampled-Data Control Systems. Springer-Verlag, 1995. [9] R. C. Curtain. H∞ control for distributed parameters systems: a survey. Proceedings of IEEE Conference on Decision and Control, pages 22–26, 1990. In [10] R. F. Curtain and K. Glover. Robust stabilization of infinite dimensional systems by finite dimensional controllers. System and Control Letters, 7:41 – 47, 1986. [11] R. F. Curtain and A. J. Pritchard. Infinite Dimensional Linear Systems Theory. Springer, 1978. [12] R. F. Curtain and H. J. Zwart. An Introduction to Infinite Dimensional Linear Systems Theory. Springer, 1995. [13] M. C. Delfour. The largest class of hereditary systems defining a c0 – semigroup on the product space. Canadian J. Math., 32:969 – 978, 1980. [14] J.C. Doyle, B.A. Francis, and A.R. Tannenbaum. Feedback control theory. Macmillan, 1992. [15] J.C. Doyle, K. Glover, P.P. Khargonekar, and B.A. Francis. State space solutions to standard H2 and H∞ control problems. IEEE Transactions on Automatic Control, AC – 34:831–847, 1989. [16] H. Dym, T. T. Georgiou, and M. C. Smith. Explicit formulas for optimally robust controllers for delay systems. IEEE Transactions on Automatic Control, AC – 40:656 – 669, 1995. [17] D. S. Flamm. Control of delay systems for minimax sensitivity. PhD thesis, MIT, 1986. [18] D. S. Flamm and S. K. Mitter. h∞ sensitivity minimization for delay systems. System and Control Letters, 9:17 – 24, 1987. [19] C. Foiaş, H. Özbay, and A. Tannenbaum. Robust Control of Infinite Dimensional Systems. Springer, 1996. LNCIS vol. 209. [20] C. Foiaş and A. Tannenbaum. Optimal sensitivity theory for multivariable distributed systems. Int. J. Control, 47:985 – 992, 1988. [21] C. Foiaş, A. Tannenbaum, and G. Zames. Weighted sensitivity minimization for delay systems. IEEE Transactions on Automatic Control, AC – 31:763–766, 1986. [22] B.A. Francis. A Course in H∞ Control Theory. Springer, 1987. [23] P. Gahinet. On the game Riccati equations arising in H∞ control problems. SIAM J. Control Optim., 32(3):635–647, 1994. 14 L. Mirkin & G. Tadmor [24] K. Glover, J. Lam, and J. R. Partington. Balanced realization and Hankel norm approximation of systems involving delays. In Proceedings of IEEE Conference on Decision and Control, pages 1810–1815, 1986. [25] K. Glover and J. R. Partington. Robust stabilization of delay systems by approximation of co-prime factors. System and Control Letters, 14:325–331, 1990. [26] M. Green and D. J. N. Limebeer. Linear Robust Control. Prentice Hall, 1995. [27] M.G. Green, K. Glover, D. Limebeer, and J. Doyle. A j-spectral factorization approach to h∞ control. SIAM J. Control and Optimization, 28(6):1350–1371, 1990. [28] G. Gu, J. Chen, and O. Toker. Computation of L2 [0, h] induced norms. In Proc. 35th IEEE Conf. Decision and Control, pages 4046–4051, Kobe, Japan, 1996. [29] G. Gu, P. P. Khargonekar, E. B. Lee, and P. Misra. Approximation of unstable infinite-dimensional systems. SIAM J. Control and Optim, 30:704–716, 1992. [30] J. K. Hale. Theory of Functional Differential Equations. Springer-Verlag, 1971. [31] A. Ichikawa. Optimal control and filtering of evolution equations with delays in control and observation. Technical Report 53, Control Theory centre, University of warwick, 1977. [32] A. Ichikawa. Differential games and H∞ problems in Hilbert space. In H. Kimura and S. Kodama, editors, Recent Adavances in Mathematical Theory of Systems, Control, Networks and Signal Processing, I, pages 115 – 120. Mita, 1992. [33] K. Ito and R. Tegals. Legendre-tau approximation for functional differential equations. SIAM J. Control and Optim, 24:737 – 759, 1986. [34] A. Kojima and S. Ishijima. Robust controller design for delay systems in the gapmetric. IEEE Trans. Automat. Control, 40(2):370–374, 1995. [35] W. H. Kwon and A. E. Pearson. Feedback stabilization of linear systems with delayed control. IEEE Transactions on Automatic Control, AC – 25:266–269, 1980. [36] T. A. Lypchuk, M. C. Smith, and A. Tannenbaum. Weighted sensitivity minimization: general plants in H∞ and rational weights. Lin. Algebra and its Appl., 109:71 – 90, 1988. [37] C. McMillan and R. Trigiani. Min-max game theory for a class of boundary control problems, analysis and optimization of systems: State and frequesncy domain approaches for infinite dimensional systems, pages 459 – 466. Springer, 1993. LNCIS 185. [38] G. Meinsma and H. Zwart. On H∞ control for dead-time systems. preprint, 1997. [39] G. Meinsma and H. Zwart. The standard H∞ control problem for dead-time systems. In Proc. of the MTNS’98 Symposium, pages 317–320, Padova, Italy, 1998. [40] G. Meinsma and H. Zwart. On H∞ control for dead-time systems. IEEE Trans. Automat. Control, 45(2):272–285, 2000. H∞ control of system with i/o delay: review 15 [41] L. Mirkin. Predictor-based solution to the H∞ control of dead-time systems. In Preprints of 3rd IFAC Symposium on Robust Control Design, Prague, Czech Republic, 2000. [42] L. Mirkin. On the extraction of dead-time controllers from delay-free parametrizations. In Proc. of LTDS’2000, pages 157–162. Ancona, Italy, 2000. [43] L. Mirkin. On the H2 and H∞ dead-time compensation. In Proc. 39th IEEE Conf. Decision and Control, pages 4584–4587, Sidney, Australia, 2000. [44] L. Mirkin and N. Raskin. State-space parametrization of all stabilizing dead-time controllers. In Proc. 38th IEEE Conf. Decision and Control, pages 221–226, Phoenix, AZ, 1999. [45] L. Mirkin and H. Rotstein. On the characterization of sampled-data controllers in the lifted domain. Syst. Control Lett., 29(5):269–277, 1997. [46] K. M. Nagpal and R. Ravi. H∞ Control and estimations problems with delayed measurements: state space solutions. In Proceedings of the American Control Conference, pages 2379–2383, 1994. [47] K. M. Nagpal and R. Ravi. H∞ Control and estimations problems with delayed measurements: state space solutions. SIAM J. Control and Optim, 35:1217 – 1242, 1997. [48] L. Naimark and N. Cohen. H∞ controller design for pure delay and other distributed systems. Int. J. Control, 67(3):355–369, 1997. [49] O. Toker and H. Özbay. Suboptimal robustness in the gap metric for MIMO delay systems. In Proceedings of the American Control Conference, pages 3183–3187, 1994. [50] H. Özbay. A simpler formula for for the singular values of a certain Hankel operator. System and Control Letters, 15:381–390, 1990. [51] H. Özbay, M. C. Smith, and A. Tannenbaum. Controller design for unstable distributed plants. In Proceedings of the American Control Conference, pages 1583– 1588, 1990. [52] H. Özbay and A. Tannenbaum. A skew Toeplitz approach to the H∞ optimal control of multivariable distributed systems. SIAM J. Control and Optim, 28:653–670, 1990. [53] H. Özbay and A. Tannenbaum. On the structure of suboptimal H∞ controllers in the sensitivity minimization for distributed stable plants. Automatica, 27:293 – 305, 1991. [54] Z. J. Palmor. Time-delay compensation — Smith predictor and its modifications. In S. Levine, editor, The Control Handbook, pages 224–237. CRC Press, 1996. [55] Z. J. Palmor and D. W. Powers. Improved dead-time compensator controllers. AIChE Journal, 31(2):215–221, 1985. [56] J. R. Partington, K. Glover, H. J. Zwart, and R. F. Curtain. L∞ Approximations and nuclearity of delay systems. System and Control Letters, 10:59–65, 1988. 16 L. Mirkin & G. Tadmor [57] A. J. Pritchard and D. Salamon. The linear-quadratic problem for retarded systems with delays in the control and observation. IMA Journal of Mathematical Control and Information, pages 335–362, 1985. [58] A. J. Pritchard and D. Salamon. The linear-quadratic problem for infinite dimensional systems with unbounded input and output operators. SIAM J. Control and Optim, 25:121 – 144, 1987. [59] I. Rhee and J.L. Speyer. A game theoretic controller and its relationship to H∞ and linear exponential gausian synthesis. Proc. 28th CDC, Tampa, 1989. [60] D. Salamon. Control And Observation of Neutral Systems. Pitman, 1984. [61] D. Sarason. Generalized interpolation in H∞ . Transaction of Amer. Math. Soc., 127: 179–203, 1967. [62] O. J. M. Smith. Closer control of loops with dead time. Chem. Eng. Progress, 53(5): 217–219, 1957. [63] B. Sz.-Nagy and C. Foiaş. Harmonic Analysis of Operators on Hilbert Space. NorthHolland, Amsterdam, 1970. [64] G. Tadmor. An interpolation problem associated with H∞ optimization in systems with distributed lags. System and Control Letters, 8:313–319, 1987. [65] G. Tadmor. H∞ Interpolation in systems with commensurate input lags. SIAM J. Control and Optim, 27:511–526, 1989. [66] G. Tadmor. H∞ in the time domain: the standard four block problem. In Proceedings of the American Control Conference, 1989. [67] G. Tadmor. Time domain optimal control and worst case linear systems design. In Proceedings of IEEE Conference on Decision and Control, pages 403–406, 1989. [68] G. Tadmor. Worst case design in the time domain: the maximum principle and the standard H∞ problem. Math. Control, Signals and Systems, 3:301–324, 1990. [69] G. Tadmor. H∞ Optimal sampled data control in continuous time systems. Int. J. Control, 56:99–141, 1992. [70] G. Tadmor. The standard H∞ problem and the maximum principle: the general linear case. SIAM J. Control and Optim, 31:831–846, 1993. [71] G. Tadmor. The Nehari problem in systems with distributed input delays is inherently finite dimensional. System and Control Letters, 26:11 – 16, 1995. [72] G. Tadmor. Robust control in the gap: a state space solution in the presence of a single input delay. IEEE Transactions on Automatic Control, AC–42:1330 – 1335, 1997. [73] G. Tadmor. Weighted sensitivity minimization in systems with a single input delay: a state space solution. SIAM J. Control and Optim, 35:1445 – 1469, 1997. [74] G. Tadmor. The standard H∞ problem in systems with a single input delay. IEEE Trans. Automat. Control, 45(3):382–397, 2000. H∞ control of system with i/o delay: review 17 [75] G. Tadmor and J. Turi. Neutral equations and associated semigroups. J. Differential Eqs., 116:59–87, 1995. [76] B. van Keulen. H∞ control for infinite dimensional systems: a state space approach. Birkhäuser, 1993. [77] R. B. Vinter. On the evolution of the state of linear differential equations in M 2 : properties of the generator. J. Inst. Maths. Applics., 21:13 – 23, 1978. [78] R. B. Vinter and R. H. Kwong. The infinite time quadratic control problem for linear systems with state and control delays: an evolution approach. SIAM J. Control and Optim, 19:139 – 153, 1981. [79] K. Watanabe and M. Ito. A process-model control for linear systems with delay. IEEE Trans. Automat. Control, 26(6):1261–1269, 1981. [80] I. Yaesh and U. Shaked. Minimum H∞ norm regulation of linear discrete time systems and its relations to linear quadratic discrete games. Proc. 28th CDC, Tampa, 1989. [81] I. Yaesh and U. Shaked. Minimum H∞ regulation of linear discrete-time systems and its relation to linear quadratic difference games. IEEE Transactions on Automatic Control, AC-35:1061–1064, 1990. [82] K. Zhou and P. P. Khargonekar. On the weighted sensitivity minimization problem for delay systems. System and Control Letters, 8:307–312, 1987. [83] H. J. Zwart, R. F. Curtain, J. R. Partington, and K. Glover. Partial fractions expansions for delay systems. SIAM J. Control and Optim, 32:808–830, 1988.