13 results
Now showing1 - 10 of 13
- Some of the metrics are blocked by yourconsent settings
Item type:Publication, Access status: with files , On dissipativity of cross-entropy loss in training ResNets — A turnpike towards architecture searchThe training of ResNets and neural ODEs can be formulated and analyzed from the perspective of optimal control. This paper proposes a dissipative formulation of the training of ResNets and neural ODEs for classification problems. Specifically, we consider a variant of the cross-entropy (label smoothing) as a loss function and as a regularization in the stage cost. Based on our dissipative formulation of the training, we prove that the training OCPs for ResNets and neural ODEs alike exhibit the turnpike phenomenon. We illustrate this finding with numerical results for the two spirals and MNIST datasets. Crucially, our training formulation ensures that the transformation of the data from input to output is achieved in the first layers. In the following layers, which constitute the turnpike, the data remains at an equilibrium state and therefore these layers do not contribute to the transformation learned. In principle, these layers can be pruned after training, resulting in a network with only the necessary number of layers thus simplifying tuning of hyperparameters.Publicationtype: Journal Article33 17 - Some of the metrics are blocked by yourconsent settings
Item type:Publication, Access status: without files , On the relationship between stochastic turnpike and dissipativity notionsIn this article, we introduce and study different dissipativity notions and different turnpike properties for discrete-time stochastic nonlinear optimal control problems. The proposed stochastic dissipativity notions extend the classic notion of Jan C. Willems to LR random variables and to probability measures. Our stochastic turnpike properties range from a formulation for random variables via turnpike phenomena in probability and in probability measures to the turnpike property for the moments. Moreover, we investigate how different metrics (such as Wasserstein or Lévy-Prokhorov) can be leveraged in the analysis. Our results are built upon stationarity concepts in distribution and in random variables and on the formulation of the stochastic optimal control problem as a finite-horizon Markov decision process. We investigate how the proposed dissipativity notions connect to the various stochastic turnpike properties and we work out the link between different forms of dissipativity.Publicationtype: Journal Article26 - Some of the metrics are blocked by yourconsent settings
Item type:Publication, Access status: with files , On the turnpike to design of deep neural networks : explicit depth boundsIt is well-known that the training of Deep Neural Networks (DNN) can be formalized in the language of optimal control. In this context, this paper leverages classical turnpike properties of optimal control problems to attempt a quantifiable answer to the question of how many layers should be considered in a DNN. The underlying assumption is that the number of neurons per layer—i.e., the width of the DNN—is kept constant. Pursuing a different route than the classical analysis of approximation properties of sigmoidal functions, we prove explicit bounds on the required depths of DNNs based on asymptotic reachability assumptions and a dissipativity-inducing choice of the regularization terms in the training problem. Numerical results obtained for the two spiral task data set for classification indicate that the proposed constructive estimates can provide non-conservative depth bounds.Publicationtype: Journal Article31 30 - Some of the metrics are blocked by yourconsent settings
Item type:Publication, Access status: with files , Optimal control of port-Hamiltonian systems: energy, entropy, and exergy(Elsevier, 2024-10-26); ; ; ; We consider irreversible and coupled reversible–irreversible nonlinear port-Hamiltonian systems and the respective sets of thermodynamic equilibria. In particular, we are concerned with optimal state transitions and output stabilization on finite-time horizons. We analyze a class of optimal control problems, where the performance functional can be interpreted as a linear combination of energy supply, entropy generation, or exergy supply. Our results establish the integral turnpike property towards the set of thermodynamic equilibria providing a rigorous connection of optimal system trajectories to optimal steady states. Throughout the paper, we illustrate our findings by means of two examples: a network of heat exchangers and a gas-piston system.Publicationtype: Journal Article45 33 - Some of the metrics are blocked by yourconsent settings
Item type:Publication, Access status: with files , Manifold turnpikes, trims, and symmetries(Springer, 2022-12); Scopus© Citations 7 21 47 - Some of the metrics are blocked by yourconsent settings
Item type:Publication, Access status: without files , Turnpike properties in optimal control: An overview of discrete-time and continuous-time resultsThe turnpike property refers to the phenomenon that in many optimal control problems, the solutions for different initial conditions and varying horizons approach a neighborhood of a specific steady state, then stay in this neighborhood for the major part of the time horizon, until they may finally depart. While early observations of the phenomenon can be traced back to works of Ramsey and von Neumann on problems in economics in 1928 and 1938, the turnpike property received continuous interest in economics since the 1960s and recent interest in systems and control. The present chapter provides an introductory overview of discrete-time and continuous-time results in finite and infinite-dimensions. We comment on dissipativity-based approaches and infinite-horizon results, which enable the exploitation of turnpike properties for the numerical solution of problems with long and infinite horizons. After drawing upon numerical examples, the chapter concludes with an outlook on time-varying, discounted, and open problems.Publicationtype: Book Part29 - Some of the metrics are blocked by yourconsent settings
Item type:Publication, Access status: without files , Dissipativity in infinite-horizon optimal control : Willems' 1971 paper revisitedJan Willems introduced the system-theoretic notion of dissipativity in his foundational two-part paper which appeared in the Archive of Rational Mechanics and Analysis in 1972. Even earlier, in a likewise pivotal 1971 IEEE Transactions on Automatic Control paper, he investigated infinite-horizon least-squares optimal control and the algebraic Riccati equation from a dissipativity point of view. This note revisits infinite-horizon optimal control leveraging strict dissipativity. We discuss the interplay between dissipativity and stability properties in continuous-time infinite-horizon problems without assuming linear dynamics or quadratic cost functions. Finally, we compare our recent results from Faulwasser and Kellett (2021) to the original findings of Willems (1971).Publicationtype: Conference Paper13 - Some of the metrics are blocked by yourconsent settings
Item type:Publication, Access status: without files , On continuous-time infinite horizon optimal control—Dissipativity, stability, and transversality(2021-12); This paper analyzes the interplay between dissipativity and stability properties in continuous-time infinite-horizon Optimal Control Problems (OCPs). We establish several relations between these properties, which culminate in a set of equivalence conditions. Moreover, we investigate convergence and stability of the infinite-horizon optimal adjoint trajectories. The workhorse for our investigations is a notion of strict dissipativity in OCPs, which has been coined in the context of economic model predictive control. With respect to the link between stability and dissipativity, the present paper can be seen as an extension of the seminal work on least squares optimal control by Jan C. Willems (1971). Furthermore, we show that strict dissipativity provides a conclusive answer to the question of adjoint transversality conditions in infinite-horizon optimal control which has been raised by Hubert Halkin (1974). Put differently, we establish conditions under which the adjoints converge to their optimal steady-state value. We draw upon several examples to illustrate our findings. Moreover, we discuss the relation of our findings to results available in the literature.Publicationtype: Journal Article9 - Some of the metrics are blocked by yourconsent settings
Item type:Publication, Access status: without files , Control of port-Hamiltonian systems with minimal energy supply(Elsevier, 2021-11-01); ; ; ; We investigate optimal control of linear port-Hamiltonian systems with control constraints, in which one aims to perform a state transition with minimal energy supply. Decomposing the state space into dissipative and non-dissipative (i.e. conservative) subspaces, we show that the set of reachable states is bounded w.r.t. the dissipative subspace. We prove that the optimal control problem exhibits the turnpike property with respect to the non-dissipative subspace, i.e., for varying initial conditions and time horizons optimal state trajectories evolve close to the conservative subspace most of the time. We analyze the corresponding steady-state optimization problem and prove that all optimal steady states lie in the non-dissipative subspace. We conclude this paper by illustrating these results by a numerical example from mechanics.Publicationtype: Journal Article6 - Some of the metrics are blocked by yourconsent settings
Item type:Publication, Access status: without files , Towards necessary and sufficient stability conditions for NMPCThe analysis of convergence and stability conditions for Non-linear Model Predictive Control (NMPC) schemes has seen significant progress during the last decades. Numerous results in the literature state sufficient conditions for convergence and/or stability of NMPC. Yet, necessary and sufficient conditions are—to the best of the author's knowledge—not known and, given the variety of choices for MPC design, seemingly appear to be difficult. This paper analyzes the convergence of NMPC of continuous-time systems in time-invariant settings. We introduce the receding-horizon Hamiltonian—i.e., the value of the optimal control Hamiltonian along the sequence of OCPs—as a novel tool for stability/convergence analysis. Based on the assumption of strict dissipativity, we prove necessary and sufficient conditions for asymptotic convergence of the closed loop to the optimal steady state. Numerical results obtained for the Van de Vusse reactor illustrate the proposed conditions.Publicationtype: Conference Paper18