Abstract
An optimization approach to linear control systems has recently become very popular. For example, the linear feedback matrix in the classical linear-quadratic regulator problem can be viewed as a variable, and the problem can be reduced to the minimization of the performance indicator for this variable. To this end, one can apply the gradient method and obtain a justification of the convergence. This approach has been successfully applied to a number of problems, including output feedback optimization. The present paper is the first to apply this approach to the peak-to-peak gain minimization problem. A gradient method for finding a static state or output feedback is written out and justified. A number of examples are considered, including the single and double pendulums.
Similar content being viewed by others
Notes
Understood in the sense of the second directional derivative.
REFERENCES
Boyd, S., El Ghaoui, L., Feron, E., et al., Linear Matrix Inequalities in System and Control Theory, Philadelphia: SIAM, 1994.
Abedor, J., Nagpal, K., and Poolla, K., A linear matrix inequality approach to peak-to-peak gain minimization, Int. J. Robust Nonlinear Control, 1996, vol. 6, no. 9–10, pp. 899–927.
Nazin, S.A., Polyak, B.T., and Topunov, M.V., Rejection of bounded exogenous disturbances by the method of invariant ellipsoids, Autom. Remote Control, 2007, vol. 68, no. 3, pp. 467–486.
Khlebnikov, M.V., Polyak, B.T., and Kuntsevich, V.M., Optimization of linear systems subject to bounded exogenous disturbances: the invariant ellipsoid technique, Autom. Remote Control, 2011, vol. 72, no. 11, pp. 2227–2275.
Polyak, B.T., Khlebnikov, M.V., and Shcherbakov, P.S., Upravlenie lineinymi sistemami pri vneshnikh vozmushcheniyakh: tekhnika lineinykh matrichnykh neravenstv (Control of Linear Systems Subject to Exogenous Disturbances: Technique of Linear Matrix Inequalities), Moscow: LENAND, 2014.
Grant, M. and Boyd, S., CVX: Matlab Software for Disciplined Convex Programming, version 2.1. http://cvxr.com/cvx .
Balandin, D.V. and Kogan, M.M., Sintez zakonov upravleniya na osnove lineinykh matrichnykh neravenstv (Synthesis of Control Laws Based on Linear Matrix Inequalities), Moscow: Fizmatlit, 2007.
Kalman, R.E., Contributions to the theory of optimal control, Boletin Soc. Mat. Mexicana, 1960, vol. 5, no. 1, pp. 102–119.
Levine, W. and Athans, M., On the determination of the optimal constant output feedback gains for linear multivariable systems, IEEE Trans. Autom. Control, 1970, vol. 15, no. 1, pp. 44–48.
Mäkilä, P.M. and Toivonen, H.T., Computational methods for parametric LQ problems—a survey, IEEE Trans. Autom. Control, 1987, vol. 32, no. 8, pp. 658–671.
Fazel, M., Ge, R., Kakade, S., and Mesbahi, M., Global convergence of policy gradient methods for the linear quadratic regulator, Proc. 35th Int. Conf. Mach. Learn. (Stockholm, Sweden, July 10–15, 2018), vol. 80, pp. 1467–1476.
Mohammadi, H., Zare, A., Soltanolkotabi, M., and Jovanović, M.R., Global exponential convergence of gradient methods over the nonconvex landscape of the linear quadratic regulator, Proc. 2019 IEEE 58th Conf. Decis. Control (Nice, France, December 11–13, 2019), pp. 7474–7479.
Zhang, K., Hu, B., and Baar T., Policy optimization for \(\mathcal H_2 \) linear control with \(\mathcal H_{\infty } \) robustness guarantee: implicit regularization and global convergence, 2020. https://arxiv.org/abs/1910.09496 .
Bu, J., Mesbahi, A., Fazel, M., and Mesbahi, M., LQR through the lens of first order methods: discrete-time case, 2019. https://arxiv.org/abs/1907.08921 .
Fatkhullin, I. and Polyak, B., Optimizing static linear feedback: gradient method, SIAM J. Control Optim., 2021 (in press). https://arxiv.org/abs/2004.09875 .
Polyak, B.T., Khlebnikov, M.V., and Shcherbakov, P.S., Linear matrix inequalities in control systems with uncertainty, Autom. Remote Control, 2021, vol. 82, no. 1, pp. 1–40.
Nesterov, Y. and Protasov, V.Y., Computing closest stable non-negative matrices, SIAM J. Matrix Anal. Appl., 2020, vol. 41, no. 1, pp. 1–28.
Polyak, B.T., Introduction to Optimization, New York: Optimization Software, 1987.
Horn, R.A. and Johnson, C.R., Matrix Analysis, New York: Cambridge Univ. Press, 1985. Translated under the title: Matrichnyi analiz, Moscow: Mir, 1989.
Lee, C.-H., New results for the bounds of the solution for the continuous Riccati and Lyapunov equations, IEEE Trans. Autom. Control, 1997, vol. 42, no. 1, pp. 118–123.
ACKNOWLEDGMENTS
The authors consider it their pleasant duty to express their gratitude to A.A. Tremba and an anonymous referee for their interest in the paper, critical remarks, and suggestions.
Funding
This work was supported by the Russian Science Foundation, project no. 21-71-30005.
Author information
Authors and Affiliations
Corresponding authors
Additional information
Translated by V. Potapchouck
APPENDIX
Lemma A.1.
Let \(X \) and \(Y \) be solutions to the dual Lyapunov equations with the Hurwitz matrix \(A \) ,
Proof of Lemma A.1. Indeed, by direct computation we have
The next lemma contains some well-known results (see, e.g., [19]), needed in the exposition to follow.
Lemma A.2.
-
1.
For matrices \(A \) and \(B \) of appropriate dimensions one has the relations
$$ \begin {gathered} \begin {aligned} \|AB\|_F&\leqslant \|A\|_F\|B\|, \\[.3em] |\mathrm{tr}\thinspace AB|&\leqslant \|A\|_F\|B\|_F, \\[.3em] \|A\|&\leqslant \|A\|_F, \end {aligned} \\[.3em] AB+{B}^{\top }{A}^{\top }\leqslant \varepsilon A{A}^{\top }+\frac {1}{\varepsilon }{B}^{\top }B\quad \text {for any}\quad \varepsilon >0. \end {gathered}$$ -
2.
For positive semidefinite matrices \(A \) and \(B \) one has the relations
$$ 0\leqslant \lambda _{\min }(A)\lambda _{\max }(B)\leqslant \lambda _{\min }(A)\mathrm{tr}\thinspace B\leqslant \mathrm{tr}\thinspace AB\leqslant \lambda _{\max }(A)\mathrm{tr}\thinspace B\leqslant \mathrm{tr}\thinspace A\mathrm{tr}\thinspace B. $$
Lemma A.3.
For a solution \(P \) of the Lyapunov equation
If, however, \( Q=D{D}^{\top }\) and the pair \((A,D)\) is controllable, then
Proof of Lemma A.3. The estimates (A.1) are well known; see, e.g., [20]. Let us prove the estimate (A.2). The explicit solution of the Lyapunov equation for a Hurwitz matrix has the form
Proof of Lemma 1 .
(a) Equation (6) can be represented in the form
Let us estimate the quantity \(f(\alpha )=\mathrm{tr}\thinspace CP(\alpha ){C}^{\top } \) using Lemma A.3 with the obvious replacements
Now let us show that the function \(f(\alpha )=\mathrm{tr}\thinspace CP(\alpha ){C}^{\top }\) is strictly convex on the interval \((0,2\sigma ) \). In accordance with [5, Lemma 1.2.3], the solution of Eq. (6) is representable in closed form as
Thus, the second derivative of the function \(f(\alpha ) \) is positive and tends to infinity at the endpoints of the interval \( (0,2\sigma )\).
In a similar way, by a straightforward calculation of the fourth derivative, we obtain
(b) Now let us derive a formula for the derivative of the function \(f(\alpha )\). In Eq. (6), the solution \(P \) is a function of \(\alpha \). Let us differentiate this equation; by \(P^{\prime } \) we mean the derivative with respect to \(\alpha \),
(c) In a similar manner, we will obtain an expression for the second derivative \(f(\alpha )\). Differentiating the equation for \(P^{\prime }\) with respect to \(\alpha \), we obtain
Proof of Lemma 3. Consider a sequence of stabilizing controllers \(\{K_j\}\subseteq \mathcal S \) such that \(K_j\to K\in \partial \mathcal S\); i.e., \(\sigma (A+BKC_1)=0 \). This means that for each \(\varepsilon >0 \) there exists a number \(N=N(\varepsilon ) \) such that the inequality
Let \(P_j\) be the solution of the Lyapunov equation (15) associated with the controller \(K_j \),
On the other hand,
Proof of Lemma 4. System (12) closed by the feedback (13) acquires the closed-loop form
Let us calculate the increment in the functional \(f(K) \) by linearizing the corresponding quantities,
Consider the Lyapunov equation (20) dual to (A.4). By Lemma A.1, from Eqs. (A.4) and (20) we have
Proof of Lemma 5. Let us calculate
Linearizing the relevant quantities, we calculate the increment in the functional \(\nabla _Kf(K)[E]\) in direction \(E \),
Thus, denoting \(P^{\prime }=P^{\prime }(K)[E] \) and \(Y^{\prime }=Y^{\prime }(K)[E] \), we have
Further, \(P=P(K)\) is a solution to Eq. (15); let us write it in increments in direction \(E \),
Further, \(Y=Y(K)\) is a solution of the Lyapunov equation (20); let us write it in increments in direction \(E\),
From (22) and (A.5) we have the relation
Corollary A.1.
The action of the Hessian of the function \( f(K)\) on a matrix \( E\in \mathbb R^{p\times l}\) such that \(\|E\|_F=1\) satisfies the estimate
Proof of Corollary A.1. According to (21),
Proof of Lemma 6. According to Corollary A.1, it suffices to estimate from above the quantity
The estimate for \(\alpha \) is established as follows:
Now let us estimate \(\|P\|\) from above,
Finally, let us estimate \(\|P^{\prime }\|_F\) from above. Considering Lemma A.2, note that
Proof of Theorem 3. First of all, Algorithm 1 is well defined at the starting point, since \(K_0 \) is a stabilizing controller by assumption. Further, for sufficiently small \(\gamma _j\), \(f(K) \) is monotone decreasing in the algorithm (motion along the antigradient); i.e., the \(K_j\) remain in the domain \(\mathcal S_0\), and thus we can apply the results in Lemma 6 about the Lipschitz property of the gradient.
Thus, the results on the convergence of the gradient method for unconditional minimization in [18] are applicable. In particular, condition (b) at step 3 in Algorithm 1 will be satisfied after a finite number of partitions, and convergence along the gradient will occur linearly in the gradient method. The proof of Theorem 3 is complete. \(\quad \blacksquare \)
Rights and permissions
About this article
Cite this article
Polyak, B.T., Khlebnikov, M.V. Static Controller Synthesis for Peak-to-Peak Gain Minimization as an Optimization Problem. Autom Remote Control 82, 1530–1553 (2021). https://doi.org/10.1134/S0005117921090034
Received:
Revised:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1134/S0005117921090034