Abstract
This chapter investigates the possibility of exactly solving three core optimization problems encountered in hybrid system identification: switching linear regression, piecewise affine regression, and bounded-error linear estimation. It formally analyzes their computational complexity, and the main conclusion is that they are \(\mathcal {NP}\)-hard, meaning that, in general, we cannot hope to compute an exact solution in reasonable time. Then, the chapter focuses on more restrictive settings and shows algorithms based on combinatorial enumerations or branch-and-bound methods that can be guaranteed to yield exact (or sufficiently close to exact) solutions in reasonable time for low-dimensional problems, i.e., when the number of regressors is small. Finally, a few numerical results highlight the limitations of such approaches and the need for the heuristics/approximations developed in the next chapters.
This is a preview of subscription content, log in via an institution.
Buying options
Tax calculation will be finalised at checkout
Purchases are for personal use only
Learn about institutional subscriptionsNotes
- 1.
Recall that in a complete procedure for hybrid system identification, at each iteration the data set is reduced to leave only the points that are not well approximated by already estimated submodels, so that N changes in Problem 5.2.
- 2.
One can similarly define the space complexity of algorithms and problems to analyze the memory requirements rather than the running time. Throughout the book, the term “complexity” refers to the time complexity.
- 3.
The normal \(\varvec{g}_{\mathcal {S}_g}\) of the hyperplane \(\{\varvec{x} : \varvec{g}_{\mathcal {S}_g}^\top \varvec{x} + b_{\mathcal {S}_g} = 0\}\) passing through the d points \(\{\varvec{x}_{k_i}\}_{i=1}^d\) of \(\mathbb {R}^d\) can be computed in \(\mathcal {O}(d^3)\) as a unit vector in the null space of \(\begin{bmatrix} \varvec{x}_{k_2}-\varvec{x}_{k_1}&\dots&\varvec{x}_{k_d} - \varvec{x}_{k_1} \end{bmatrix}^\top \), while the offset is given by \(b_{\mathcal {S}_g}= -\varvec{g}_{\mathcal {S}_g}^\top \varvec{x}_{k_i}\) for any of the \(\varvec{x}_{k_i}\)’s.
- 4.
Note that for the particular case of \(d=1\), the inner loop over the labelings of \(\mathcal {S}_g\) could be avoided to reduce the number of cost function evaluations from 2N to N. Assume that the \(x_k\)’s are sorted and fix the label of \(x_k \in \mathcal {S}_g\) to \(q_k=+1\) in the kth outer iteration and the ones of the points \(x_i<x_k\) to \(q_i = -1\). Then, we simply obtain the other labeling with \(q_k=-1\) in the next iteration.
- 5.
For any set of \(d-1\) points \(\varvec{x}_k\), there is a vector \(\varvec{\theta }\) such that \(\varvec{x}_k^\top \varvec{\theta }=0\). Thus, in \(\mathbb {R}^{d+1}\), there is a hyperplane of normal \(\begin{bmatrix}0&\varvec{\theta }^\top \end{bmatrix}^\top \) passing through the origin and the corresponding \(2d-2\) points of \(\mathcal {Z}\), \(\varvec{z}_k\), and \(\varvec{z}_{k+N}\).
References
Lauer, F.: Global optimization for low-dimensional switching linear regression and bounded-error estimation. Automatica 89, 73–82 (2018)
Garey, M.R., Johnson, D.S.: Computers and Intractability: A Guide to the Theory of NP-Completeness. W.H Freeman and Co, New York (1979)
Blondel, V.D., Tsitsiklis, J.N.: A survey of computational complexity results in systems and control. Automatica 36(9), 1249–1274 (2000)
Cook, S.A.: The complexity of theorem-proving procedures. In: Proceedings of the 3rd Annual ACM Symposium on Theory of Computing, Shaker Heights, OH, USA, pp. 151–158 (1971)
Karp, R.M.: Reducibility among combinatorial problems. Complexity of Computer Computations. The IBM Research Symposia Series, pp. 85–103. Plenum Press/Springer, Berlin (1972)
Blum, L., Cucker, F., Shub, M., Smale, S.: Complexity and Real Computation. Springer, Berlin (1998)
Lauer, F.: On the complexity of piecewise affine system identification. Automatica 62, 148–153 (2015)
Lauer, F.: On the complexity of switching linear regression. Automatica 74, 80–83 (2016)
Amaldi, E., Mattavelli, M.: The MIN PFS problem and piecewise linear model estimation. Discret. Appl. Math. 118, 115–143 (2002)
Amaldi, E., Kann, V.: The complexity and approximability of finding maximum feasible subsystems of linear relations. Theor. Comput. Sci. 147(1–2), 181–210 (1995)
Lauer, F.: On the exact minimization of saturated loss functions for robust regression and subspace estimation. Pattern Recognit. Lett. 112, 317–323 (2018)
Breiman, L.: Hinging hyperplanes for regression, classification, and function approximation. IEEE Trans. Inf. Theory 39(3), 999–1013 (1993)
Roll, J., Bemporad, A., Ljung, L.: Identification of piecewise affine systems via mixed-integer programming. Automatica 40(1), 37–50 (2004)
Lau, K., Leung, P., Tse, K.: A mathematical programming approach to clusterwise regression model and its extensions. Eur. J. Op. Res. 116, 640–652 (1999)
Carbonneau, R.A., Caporossi, G., Hansen, P.: Globally optimal clusterwise regression by mixed logical-quadratic programming. Eur. J. Op. Res. 212, 213–222 (2011)
Author information
Authors and Affiliations
Corresponding author
Notes
Notes
This chapter only provided a brief introduction to computational complexity, mainly sufficient to understand the results dedicated to hybrid system identification. For a full understanding of the technical issues involved, many textbooks are available, among which we recommend [2] for a focus on \(\mathcal {NP}\)-completeness. Alternatively, an introduction to these concepts with an automatic control perspective is given in [3]. Two major papers set the foundations of this field: [4] identified the first \(\mathcal {NP}\)-complete problem and [5] established a list of 21 \(\mathcal {NP}\)-complete problems (including the Partition Problem 5.4), from which reductions are still used to prove the \(\mathcal {NP}\)-hardness of many problems. Note that models of computation over the reals have also been considered and we refer to [6] for more details on this topic.
The complexity results (\(\mathcal {NP}\)-hardness and polynomial-time algorithm) for PWA regression were derived in [7], which also formalized the equivalence between linear classifiers and hyperplanes passing through subsets of points. Similar results for switching linear regression and the proof of Theorem 5.1 for the noiseless case are found in [8], while the proof of Theorem 5.2 for the noisy case has not been published elsewhere. The MIN PFS formulation of the general bounded-error problem is due to [9], in which a proof of \(\mathcal {NP}\)-hardness for MIN PFS can be found. For the greedy approach in which the models are estimated one by one, the \(\mathcal {NP}\)-hardness of the bounded-error estimation subproblem, i.e., Problem 5.2, is due to [10], while the polynomial-time algorithm is due to [11].
The global optimization methods based on branch-and-bound for switching regression and bounded-error estimation were derived in [1]. Hinging hyperplane models were introduced in [12] and their global optimization set as a MIQP problem as presented in this chapter was proposed in [13]. Other works on the global optimization of switching linear regression models with a mixed-integer programming approach also appeared in the operation research community; see, e.g., [14, 15].
Rights and permissions
Copyright information
© 2019 Springer Nature Switzerland AG
About this chapter
Cite this chapter
Lauer, F., Bloch, G. (2019). Exact Methods for Hybrid System Identification. In: Hybrid System Identification. Lecture Notes in Control and Information Sciences, vol 478. Springer, Cham. https://doi.org/10.1007/978-3-030-00193-3_5
Download citation
DOI: https://doi.org/10.1007/978-3-030-00193-3_5
Published:
Publisher Name: Springer, Cham
Print ISBN: 978-3-030-00192-6
Online ISBN: 978-3-030-00193-3
eBook Packages: Intelligent Technologies and RoboticsIntelligent Technologies and Robotics (R0)