Вы находитесь на странице: 1из 6

IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 58, NO.

9, SEPTEMBER 2013

2405

[10] M. Pavone and E. Frazzoli, Decentralized policies for geometric pattern formation and path coverage, J. Dyn. Syst., Meas., Control, vol. 129, pp. 633643, 2007. [11] J. Ramirez, New Decentralized Algorithms for Spacecraft Formation Control Based on Cyclic Pursuit Approach, Ph.D. dissertation, MIT, Cambridge, MA, USA, 2010. [12] A. Sinha and D. Ghose, Generalization of linear cyclic pursuit with application to rendezvous of multiple autonomous agents, IEEE Trans. Autom. Control, vol. 51, no. 11, pp. 18191824, Nov. 2006. Fig. 1. Simplest possible adaptive ESC scheme.

Extremum Seeking for Constrained Inputs


Ying Tan, Senior Member, IEEE, Yuping Li, Member, IEEE, and Iven M. Y. Mareels, Fellow, IEEE

AbstractExtremum seeking control (ESC) is an adaptive control scheme that locates an extremum of an input-output map, without any explicit knowledge of this map apart from its existence. As is typical in adaptive control an integrator is used to drive the parameters that are being adapted. Due to this integrator, it is possible that the adapted parameters wander outside their physically relevant domain as the underlying adaptation technique is blind to this constraint. As these constraints may represent realistic operational limits it is important to design ESC to respect them. Two such ESC schemes are proposed. One is based on a constrained optimization approach in which some penalty function is used to adapt the search so as not to violate the constraints. The other technique uses an anti-windup scheme, a widely used mechanism in engineering to prevent windup of integral action in a controller. It is demonstrated that both methods are essentially equivalent. In that for any penalty-function based ESC, there exists an equivalent anti-windup ESC whose phase portrait is a global approximate of the penalty-function ESC. Some simulations are presented to illustrate the main results. Index Terms Extremum seeking control (ESC), input constraints.

I. INTRODUCTION Extremum seeking control (ESC) seeks an optimal input for a generally unknown objective function (or input-to-output). This technical note focuses on adaptive ESC, which has been widely used in various applications [2], [10], [16]. ESC uses a dither signal to probe the input-output map so as to estimate an approximate gradient of this map, and then uses this gradient estimate to drive the input-output map to a desired extremum. As indicated in [16], the simplest possible adaptive ESC scheme when applied to a static plant is shown in Fig. 1. One of the key components of ESC is the integrator. As discussed in [16], together with the dither signal , the integral action can extract the gradient of the unknown static mapping , provided that the mapping is continuous and differentiable so that the gradient is well-dened. Typically, the input variable must live within a particular domain, as restricted by the operational limits of the plant . Given that (in
Manuscript received August 19, 2012; revised August 22, 2012; accepted January 09, 2013. Date of publication March 26, 2013; date of current version August 15, 2013. This work is supported by the Australian Research Council (ARC) under Future Fellow Project: FT0991385. Recommended by Associate Editor L. Zaccarian. The authors are with Department of Electrical and Electronic Engineering, The University of Melbourne, VIC 3010, Australia (e-mail: yingt@unimelb. edu.au; yupingl@unimelb.edu.au; i.mareels@unimelb.edu.au). Color versions of one or more of the gures in this technical note are available online at http://ieeexplore.ieee.org. Digital Object Identier 10.1109/TAC.2013.2254638

Fig. 1) is driven via an integrator it is quite feasible that this input variable will wander outside this acceptable domain, and hence the actual input may get saturated on the boundary of the domain, say as a consequence of actuator limits and will differ from the presumed input. This difference is unfortunate, as with a xed input the output will not correlate with the dither. As a consequence, the integrator will equally saturate, i.e., and get stuck at this inappropriate point on the boundary of the operational domain. This is very similar to integrator wind-up in classic control [3]. In order to solve such a problem in ESC, two approaches have been proposed. From an optimization point of view, the input saturation can be viewed as a constraint. Thus ESC with input saturation can be treated as a constrained optimization problem. A penalty function becomes a natural way to handle input saturation. There is the rich literature on penalty functions ([4], [13], [14]), which may be used to advantage to propose penalty-function-based ESC. The rst result shows that the penalty-function-based ESC converges semi-globally practically asymptotically to the optimal value as long as the optimal input is indeed found within the saturation limits. On the other hand, from a traditional control perspective, anti-windup techniques (see for example [1], [5], [7], [17], [19] and references therein) are of a natural choice to deal with integrator wind-up like behavior described above. As indicated in [6], an anti-windup mechanism may be added in an adaptive ESC loop to avoid this integral windup. In particular, several anti-windup ESC schemes were proposed and used in various applications to handle integral windup [9], [11], [15]. Although these anti-windup compensators work in those applications, they lack a theoretical analysis. Most existing anti-windup mechanisms were developed in the context of linear systems. Generally it is hard to demonstrate how anti-windup mechanisms can work well for general nonlinear (static/dynamic) systems. A simple static anti-windup mechanism [1] is added to ESC to compensate for the saturation effects. It is shown that this anti-windup ESC is very similar to a special case of penalty-function-based ESC, i.e., the trajectories of this anti-windup ESC are close to those of the special penalty-function-based ESC. The convergence of the penalty-functionbased ESC and the closeness of two solutions ensure that the output of this anti-windup ESC converges to the a small neighborhood. The size of which depends on the particular design parameters in the ESC algorithm. In order to show that the anti-windup ESC works, we exploit the link (closeness of solutions) between the anti-windup ESC and the penalty-function-based ESC. It is not surprising that these two methods are closely related to each other as the key ideas of these two methods are the same, though they come from different perspectives. When saturation happens (or input constraints being violated), both the penalty function and the anti-windup component will drive the system back to unsaturated mode. Both methods use some penalty to prevent or overcome saturation. By exploring the similarities between these two methods, we uncover some links between the anti-windup mechanism and the penalty function in constrained optimization. That is, for any penalty-function-based ESC, there exists an equivalent anti-

0018-9286 2013 IEEE

2406

IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 58, NO. 9, SEPTEMBER 2013

windup ESC which approximates the behavior of the penalty-function-based ESC. This link provides new ways to design anti-windup control schemes (not limited to ESC) using the rich literature of penalty functions in constrained optimization, and vice versa. It also provides a mathematical characterization of anti-windup mechanisms used in control engineering. The technical note is organized as follows. Section II provides the problem formulation. The main results are presented in Section III, including penalty-function-based ESC, anti-windup ESC and their links. Simulation results are provided in Section IV, followed by conclusions in Section V. Brief proofs of the main results are provided in the Appendix. II. PROBLEM FORMULATION In this note, the set of real numbers is denoted by . A function is of class if for each xed the function is continuous, zero at zero and strictly increasing and for each the function is strictly decreasing to zero. The notation is the order of magnitude notation. In order to illustrate the main ideas of this technical note, a simple static single-input-single-output (SISO) plant with input saturation is considered. A similar result can be extended to multi-input-multi-output dynamic systems by using multiple ESC loops and singular perturbation techniques.1 (1) is a smooth function and the saturation function where dened as follows: if if if is (2)

This constrained optimization problem can be tackled by using some penalty function methods. The penalty function usually is some measure of violation of the constraints. The measure of violation is nonzero when the constraints are violated and is zero in the region where constraints are not violated. The penalty function approach, instead of solving the original optimization with constraints, generates an auxiliary problem (without constraints) by augmenting the objective function with a penalty function. In the sequel, the following auxiliary objective function is introduced: (6) is some penalty function, satisfying the following where assumption: is absolutely conAssumption 2: The penalty function tinuous. Moreover, the following inequalities hold: (7) Some penalty functions satisfying Assumption 2 (see also [13], [14]) are listed. 1) The logarithmic penalty function: (8) 2) The inverse penalty function:

3) The polynomial penalty function:

where and and are the upper limit and the lower limit of the input respectively. For simplicity, the following assumption is used as in [16, Assumption 3]. This assumption can be relaxed if the derivative of is dened almost everywhere (a.e.), again using a smooth approximation technique. Assumption 1: There exists a unique such that the following holds: (3) Assumption 1 indicates that even though is unknown, it is known that there exists a unique maximum . The control objective of the extremum seeking is to nd the appropriate control input such that the output of the system (1) satises (4)

4) The absolute-value penalty function

where is a positive constant. 5) The Courant-Beltrami penalty function

III. MAIN RESULTS A. Penalty-Function-Based ESC The problem of interests (1) can be treated as some constrained optimization problem (5) where only the output signal is measurable.

1When a dynamic plant is considered, by replacing by and by in Fig. 1, where is a small positive constant. The dynamic system becomes a faster system compared with the slow updating law of . By using besingular perturbation techniques [8], the reduced system in time comes a static system (1). With some appropriate assumptions on the boundary layer system, it can be shown that ESC can work for dynamic systems provided that it works for static systems, see [16] for more details.

where is a positive constant. Remark 1: When the input constraints are violated, the derivative of with respect to is zero. The derivative of will be dominated by that of the penalty function. Assumption 2 shows that when the constraints are violated, if gradient search methods will converge to the (for example, ESC) are used, the output of boundary of feasible region. Once the ESC converges to the boundary is a constant of feasible region in which the penalty function is dominated by the almost everywhere, the derivative of whose extremum can be found by original objective function ESC. is assumed in the conRemark 2: Observe that knowledge of struction of the constrained function. In particular this implies that it is known in advance in which interval of the real line to search for the maximum. The diagram of a penalty-function-based ESC scheme is shown in Fig. 2. The rst result is stated as follows. and satisfying Assumption 1, if Theorem 1: For any satisfying Assumption 2, then the output of the there exists penalty-function-based ESC in Fig. 2 semi-globally practically asymp, uniformly in . totically converges to the optimal value Proof: See Appendix VI-A.

IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 58, NO. 9, SEPTEMBER 2013

2407

Fig. 2. Penalty-function-based ESC scheme.

Fig. 4. Special case of penalty-function-based ESC scheme.

where ordinates,

. Introducing the change of the co, the system takes the following form:

(10)

Fig. 3. Anti-windup ESC scheme.

Remark 3: As discussed in [16], the averaged system coming from averaging technique2 of the ESC approximates the gradient of the . As is not smooth, the gradient of the unknown function is not well-dened, though it is intuitively objective function clear that the averaging technique will smooth out any non-smooth component from the integrator. Smooth approximation techniques are thus employed in the proof to facilitate the convergence analysis. Remark 4: In principle it would sufce to establish the theorem just in , but the theorem is true for all initial conditions for all initial conditions. This is important when generalizing the result to higher dimensions, as only in the 1D case can it be guaranteed that the search trajectory will stay within the desired domain at all times. B. Anti-Windup ESC Instead of treating the input saturation as a constraint in the extremum seeking, anti-windup technique can tackle the windup-like behavior due to the existence of the integrator and input saturation. Anti-windup controller design has been an active research topic for over several decades. Many anti-windup techniques have been proposed in the literature (see for example [1], [5], [7], [17], [19] and references therein). Most anti-windup mechanisms in literature are designed for linear systems. In the context of ESC, the plant of interests is always nonlinear. Usually it is hard to show how anti-windup mechanism can work rigorously for general nonlinear systems. This technical note explores the similarities between anti-windup ESC and penalty-function-based ESC. These similarities are utilized in the convergence analysis of anti-wind ESC. Start with a simple static anti-windup mechanism [9] as shown in Fig. 3, where it is assumed that the output of the saturation of the input is available for measurement. (This is not strictly necessary, an estimate is selected as where is sufcient.) The parameter the positive constant is a design parameter. In the sequel, the system can be represented as (9)
2Averaging is an approximation method for analysis of time-varying systems. , called the As pointed out in [12], an auxiliary time-invariant system averaged system, is used to investigate properties of a time-varying dynamical that depends on a small parameter . system,

is the approxNote that the anti-windup block: imated gradient of a quadratic polynomial penalty function: , when saturation happens. Hence, the way anti-windup works is as if there was a particular penalty function. As with penalty-function-based ESC, an auxiliary objective function is introduced (11) which is a special case of (6) in that a quadratic polynomial penalty function is used. The corresponding penalty-function-based ESC is shown in (4). Theorem 1 informs that this penalty-function-based ESC scheme in Fig. 2 works well. Introduce the following averaged system for the system (10) in and time , it yields co-ordinate

(12) The following proposition shows that the averaged system of the antiwindup ESC in Fig. 3 is a perturbed averaged system of the penaltyfunction-based ESC shown in Fig. 4. The perturbation part can be made arbitrarily small by selecting a sufciently small for a given . , , there exists , Proposition 1: For any such that for any , the following equality holds:

(13) Proof: See Appendix VI-B. The next result follows immediately from Proposition 1 and Theorem 1 and is stated here without proof. be a positive Theorem 2: Assume Assumption 1 holds. Let . For any , there exists a positive pair pair and such that for any and , the solutions for the system (10) satisfy (14)

2408

IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 58, NO. 9, SEPTEMBER 2013

Remark 5: Although the penalty-function-based ESC in Fig. 4 is quite similar to anti-windup ESC in Fig. 3, there is an obvious difference: in the penalty-function-based ESC, is treated as if it were completely unknown and the ESC generates an approximation of is its gradient. In anti-windup ESC, as the second term of known, its gradient is used directly. Moreover, with the penalty function approach the output of the saturated input is not required, whereas with the anti-windup scheme we need to know the saturated input as well as the input, i.e. the output of the saturation is measurable, that is not always practical. Remark 6: Note that since the search size of ESC is in the order is important in the anti-windup of , the selection of ESC, though can be a very large number. As the gradient of in is estimated from the ESC with the coefcient , the . same coefcient is needed for the second term of The role of penalty function in and the anti-windup component in Fig. 3 is very similar: when saturation happens, both the penalty function and the anti-windup component will compensate the effect coming from input saturation. The literature dealing with constrained optimization introduces many different penalty functions as listed above. All of them can be used to design corresponding antiwindup controllers. C. Links Between Penalty-Function-Based ESC and Anti-Windup ESC The similarities between anti-windup ESC (Fig. 3) and penalty-function-based ESC (4) are sufcient motivations to explore the links between these two methods further. Introduce the following function for convenience: (15) where satises Assumption 2. The function the following properties P1 has

Fig. 5. Anti-windup ESC corresponding to the penalty-function-based ESC in Fig. 2.

Assume Assumption 1 holds. Then there exists a positive pair such that for any and , the solutions of the system (18) satisfy (19) Remark 7: Theorem 3 clearly shows a link between the two ESC schemes as presented. That is, for any penalty-function-based ESC scheme, there exists a corresponding anti-windup ESC, which achieves a similar performance. The different choice of the penalty function will lead to different (static) anti-windup mechanism. The rich literature on penalty functions provides many new ways to design various anti-windup mechanisms. Remark 8: It is noted that Property 1 of shows how antiwindup ESC actually works. When the input is far away from the conwill pull the input back to the nearest straints, the anti-windup saturation bound. This property can be used to rigorously characterize the static anti-windup mechanisms. Our future work will explore how to use penalty functions to design various anti-windup mechanism (not limited to ESC) more systematically. Remark 9: Theorem 3 shows that for any penalty-function-based ESC scheme, there exists a corresponding anti-windup mechanism. On satthe other hand, for any integrable nonlinear function isfying Property 1 in Fig. 5, the following function: (20) always satises Assumption 2 and Property 2. This also shows the possibility to get a penalty-function-based ESC from a given static anti-windup ESC. IV. SIMULATION RESULTS AND DISCUSSIONS

(16) P2 For any there exists holds and given satisfying Assumption 2, such that for any , the following

(17) Both properties comes from the denition of and Assumption 2, thus the proof is omitted. , it leads With the introduction of the anti-windup function to an anti-windup ESC coming from the penalty-function-based ESC. Fig. 5 shows the structure of such an anti-windup ESC coming from Fig. 2. This leads to the following closed-loop system:

This section provides simulation results to demonstrate the effectiveness of the two proposed ESC schemes. A very simple example is used to illustrate the main messages of this note (21) , . Let and . Assumption with , and . From Fig. 6, it is 1 holds. Select clear that when there is no input saturation, the ESC works well. With input saturation, the output of the system does not converge to its global , but rather sticks at the boundary of the acceptable maximum . domain as input signal is saturated at A. Penalty-Function-Based ESC and Anti-Windup ESC In this case study, the penalty function to cope with the saturation is set as (22) where is a xed positive number, with which ises Assumption 2. in (22) sat-

(18)

With the help of Property 2, the following result is a direct outcome of Theorem 1. be a function satisfying Assumption 2, Theorem 3: Let as in (15). Let be a positive pair and .

IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 58, NO. 9, SEPTEMBER 2013

2409

Fig. 8. Anti-windup designed from barrier function implemented with PI control. Fig. 6. Performance of the standard ESC with/without input saturation.

penalty function and anti-windup mechanism work along very similar lines. In particular, it has been established that for any penalty-function-based ESC, there is a corresponding (static) anti-windup ESC. This link shows the possibility to generate new anti-windup mechanism in control engineering and new penalty function in constrained optimization in a systematic way. APPENDIX A. Proof of Theorem 1
Fig. 7. Performance of the penalty-function-based ESC and its corresponding anti-windup ESC.

The corresponding compensator in the anti-windup ESC is . In this case, the parameter is selected as 0.01. Fig. 7 compares the closed-loop response with anti-windup ESC (the thick solid line) and the penalty-function-based ESC (the thick dashed line). In both cases, the output converges to a small neighborhood of the maximum. B. Anti-Windup Compensation for a PID Controller satIt is intuitively clear that the anti-windup function isfying Property P1 can be used to compensate the effect of windup in other integral control actions (other than ESC), provided that the closed-loop stability is achieved by the nominal feedback controller when the controller output is within the saturation limits. In order to illustrate this idea, we use an example from [18] in which an anti-windup PID control was proposed. The plant model , is is the Laplace transform of the output and is the where Laplace transform of the input. The nominal stabilizing PI controller . We implement an anti-windup compenis sator derived from an logarithmic barrier function-based ESC, i.e. . The and . As shown saturation bounds are set as in Fig. 8, the output of the plant with anti-windup mechanism, PI controller and saturated input outperforms the PI controller without anti-windup mechanism. The performance of this simple anti-windup mechanism is also comparable to the well-tuned anti-windup PI controller in [18]. V. CONCLUSION Two ESC schemes are proposed to deal with input saturation. In penalty-function-based ESC, one adds some penalty function to the objective function to penalize the violation of input constraints. Antiwindup ESC introduces an anti-windup component to the ESC to prevent wind-up like behavior. It is interesting to observe that both

The proof consists of three steps. satises Step 1 shows the piece-wise continuous function Assumption 1 a.e. Proof: It is easy to show that there exists a unique maximum of and . On the other hand, the derivative of with respect to its rst argument is or The result holds by applying Assumption 2. Step 2 nds a smooth approximation of . is constructed as An approximation of (23)

where and are smooth functions and is some constant within saturation bound. which ensures that and satisfy the following conditions: Two smooth functions

This construction shows that satises Assumption 1 for any given and . Moreover, this construction ensures that and are only different in two small intervals with size . Note the smoothness of and , we can conclude that for all in some compact set. Step 3 shows convergence.

2410

IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 58, NO. 9, SEPTEMBER 2013

The closed-loop system shown in Fig. 2 in the new coordinated is

(24) for which the averaged system in coordinate in time is

The rst term inside the integral is in a standard ESC form. The semiglobal practical convergence is ensured as shown in [16, Corollary 1]). The second term can be made arbitrarily small by selecting sufciently small . Therefore, we can show that the trajectories in (24) will converge to 0 semi-globally practically asymptotically, uniformly in . B. Proof of Proposition 1 A simple calculation leads to

[8] H. K. Khalil, Nonlinear Systems, 3rd ed. Upper Saddle River, NJ: Prentice-Hall, 2002. [9] N. J. Killingsworth and M. Krsti, PID tuning using extremum seeking: Online, model-free performance optimization, IEEE Control Syste. Mag., vol. 2, pp. 7079, 2006. [10] M. Krsti and H. H. Wang, Stability of extremum seeking feedback for general nonlinear dynamic systems, Automatica, vol. 36, pp. 595601, 2000. [11] P. Li and Y. Li, Efcient operation of air-side economizer using extremum seeking control, J. Dyn. Syst., Meas., Control, vol. 132, no. 3, pp. 031009031018, 2010. [12] D. Nesic and P. M. Dower, A note on input to state stability and averaging of systems with inputs, IEEE Trans. Autom. Control, vol. 46, no. 11, pp. 17601765, Nov. 2001. [13] J. Nocedal and S. J. Wright, Numerical Optimization. New York: Springer, 1999, 0-387-98793-2. [14] J. A. Snyman, Practical Mathematical Optimization. New York: Springer Science + Business Media, Inc., 2005. [15] E. Schuster, M. L. Walker, D. A. Humphreys, and M. Krsti, Plasma vertical stabilization with actuation constraints in the DIII-D tokamak, Automatica, vol. 41, pp. 11731179, 2005. [16] Y. Tan, D. Nei, and I. Mareels, On non-local stability properties of extremum seeking control, Automatica, vol. 42, pp. 889903, 2006. [17] S. Tarbouriech and M. Turner, Anti-windup design: An overview of some recent advances and open problems, IET Control Theory Appl., vol. 3, no. 1, pp. 119, 2009. [18] D. Xue, Y. Chen, and D. P. Atherton, Linear Feedback Control: Analysis and Design with MATLAB (Advances in Design and Control). Philadelphia, PA: Society for Industrial and Applied Mathematics, 2009. [19] L. Zaccarian and A. R. Teel, Modern Anti-Windup Synthesis: Control Augmentation for Actuator Saturation. Princeton, NJ: Princeton Univ. Press, 2011.

Stability Analysis of Second-Order Sliding Mode Control Systems With Input-Delay Using Poincar Map
(25) Let , applying Taylor series expansion for function at and performing integration with respect to , it follows that:
AbstractThis note discusses the stability of dynamical systems with input-delay under the second-order sliding mode control algorithm. Poincar Map is constructed to analyze the switching dynamics and to derive the stability conditions. Different parameter setting options are given for ensuring stability. Simulation examples are presented to verify the theoretical results. Index TermsHigh-order sliding mode (HOSM), sliding mode control (SMC).

Xiangjun Li, Xinghuo Yu, and Qing-Long Han

(26) The proof is completed by substituting (26) into (25).

REFERENCES
[1] K. J. mstrm and T. Hgglund, PID Controllers. Theory, Design and Tuning. Research Triangle Park, NC: Instruments Society of America, 1995. [2] K. B. Ariyur and M. Krsti, Real-Time Optimization by Extremum Seeking Control. Hoboken, NJ: Wiley-Interscience, 2003. [3] C. H. Edwards and D. E. Penney, Calculus With Analytic Geometry, 5th ed. Upper Saddle River, NJ: Prentice Hall, 1998. [4] R. Fletcher, An ideal penalty function for constrained optimization, IMA J. Appl. Math., vol. 15, no. 3, pp. 319342, 1975. [5] S. Galeania and A. R. Teel, On a performance-robustness trade-off intrinsic to the natural anti-windup problem, Automatica, vol. 42, pp. 18491861, 2006. [6] B. I. Godoy, J. H. Braslavsky, and J. C. Agero, A simulation study on model predictive control and extremum seeking control for heap bioleaching processes, in Proc. 17th World Congress Int. Fed. Autom. Control, Seoul, Korea, Jul. 611, 2008, pp. 93689373. [7] G. Grimm, A. R. Teel, and L. Zaccarian, Robust linear anti-windup synthesis for recovery of unconstrained performance, Int. J. Robust Nonlin. Control, vol. 14, no. 1315, pp. 11331168, 2004.

I. INTRODUCTION Sliding mode control (SMC) systems are known to be simple for design and robust in parameter variations and disturbances. However,
Manuscript received March 28, 2012; revised August 29, 2012; accepted February 28, 2013. Date of publication April 03, 2013; date of current version August 15, 2013. This work was supported in part by the Australian Research Council Discovery Project DP130104765. Recommended by Associate Editor X. Chen. X. Li and X. Yu are with RMIT University, Melbourne, VIC 3001, Australia (e-mail: xiangjun.li@rmit.edu.au; x.yu@rmit.edu.au). Q.-L. Han is with the Centre for Intelligent and Networked Systems, and the School of Information and Communication Technology, Central Queensland University, Rockhampton QLD 4702, Australia (e-mail: q.han@cqu.edu.au). Color versions of one or more of the gures in this technical note are available online at http://ieeexplore.ieee.org. Digital Object Identier 10.1109/TAC.2013.2256673

0018-9286 2013 IEEE

Вам также может понравиться