Search

Neural networks are regularly employed in adaptive control of nonlinear systems and related methods of reinforcement learning. A common architecture uses a neural network with a single hidden layer (i.e. a shallow network), in which the weights and biases are fixed in advance and only the output layer is trained. While classical results show that there exist neural networks of this type that can approximate arbitrary continuous functions over bounded regions, they are non-constructive, and the networks used in practice have no approximation guarantees. Thus, the approximation properties required for control with neural networks are assumed, rather than proved. In this paper, we aim to fill this gap by showing that for sufficiently smooth functions, ReLU networks with randomly generated weights and biases achieve $L_{\infty}$ error of $O(m^{-1/2})$ with high probability, where $m$ is the number of neurons. It suffices to generate the weights uniformly over a sphere and the biases uniformly over an interval. We show how the result can be used to get approximations of required accuracy in a model reference adaptive control application., Comment: Under Review for Conference on Decision and Control

Bilevel optimization with traffic equilibrium constraints plays an important role in transportation planning and management problems such as traffic control, transport network design, and congestion pricing. In this paper, we consider a double-loop gradient-based algorithm to solve such bilevel problems and provide a non-asymptotic convergence guarantee of $\mathcal{O}(K^{-1})+\mathcal{O}(\lambda^D)$ where $K$, $D$ are respectively the number of upper- and lower-level iterations, and $0<\lambda<1$ is a constant. Compared to existing literature, which either provides asymptotic convergence or makes strong assumptions and requires a complex design of step sizes, we establish convergence for choice of simple constant step sizes and considering fewer assumptions. The analysis techniques in this paper use concepts from the field of robust control and can potentially serve as a guiding framework for analyzing more general bilevel optimization algorithms.

Classical results in neural network approximation theory show how arbitrary continuous functions can be approximated by networks with a single hidden layer, under mild assumptions on the activation function. However, the classical theory does not give a constructive means to generate the network parameters that achieve a desired accuracy. Recent results have demonstrated that for specialized activation functions, such as ReLUs and some classes of analytic functions, high accuracy can be achieved via linear combinations of randomly initialized activations. These recent works utilize specialized integral representations of target functions that depend on the specific activation functions used. This paper defines mollified integral representations, which provide a means to form integral representations of target functions using activations for which no direct integral representation is currently known. The new construction enables approximation guarantees for randomly initialized networks for a variety of widely used activation functions.

Spectrum estimation is a fundamental methodology in the analysis of time-series data, with applications including medicine, speech analysis, and control design. The asymptotic theory of spectrum estimation is well-understood, but the theory is limited when the number of samples is fixed and finite. This paper gives non-asymptotic error bounds for a broad class of spectral estimators, both pointwise (at specific frequencies) and in the worst case over all frequencies. The general method is used to derive error bounds for the classical Blackman-Tukey, Bartlett, and Welch estimators. In particular, these are first non-asymptotic error bounds for Bartlett and Welch estimators., Comment: 15 pages, 3 figures, under review in IEEE Transactions on Signal Processing

This paper defines geometric criteria which are then used to establish sufficient conditions for persistency of excitation with vector functions constructed from single hidden-layer neural networks with step or ReLU activation functions. We show that these conditions hold when employing reference system tracking, as is commonly done in adaptive control. We demonstrate the results numerically on a system with linearly parameterized activations of this type and show that the parameter estimates converge to the true values with the sufficient conditions met.

Langevin algorithms are gradient descent methods augmented with additive noise, and are widely used in Markov Chain Monte Carlo (MCMC) sampling, optimization, and machine learning. In recent years, the non-asymptotic analysis of Langevin algorithms for non-convex learning has been extensively explored. For constrained problems with non-convex losses over a compact convex domain with IID data variables, the projected Langevin algorithm achieves a deviation of $O(T^{-1/4} (\log T)^{1/2})$ from its target distribution [27] in $1$-Wasserstein distance. In this paper, we obtain a deviation of $O(T^{-1/2} \log T)$ in $1$-Wasserstein distance for non-convex losses with $L$-mixing data variables and polyhedral constraints (which are not necessarily bounded). This improves on the previous bound for constrained problems and matches the best-known bound for unconstrained problems., Comment: 51 pages. Accepted by NeurIPS 2022. Corrected some errors during the reviewing process

This paper is motivated by the problem of quantitatively bounding the convergence of adaptive control methods for stochastic systems to a stationary distribution. Such bounds are useful for analyzing statistics of trajectories and determining appropriate step sizes for simulations. To this end, we extend a methodology from (unconstrained) stochastic differential equations (SDEs) which provides contractions in a specially chosen Wasserstein distance. This theory focuses on unconstrained SDEs with fairly restrictive assumptions on the drift terms. Typical adaptive control schemes place constraints on the learned parameters and their update rules violate the drift conditions. To this end, we extend the contraction theory to the case of constrained systems represented by reflected stochastic differential equations and generalize the allowable drifts. We show how the general theory can be used to derive quantitative contraction bounds on a nonlinear stochastic adaptive regulation problem.

Langevin algorithms are gradient descent methods with additive noise. They have been used for decades in Markov chain Monte Carlo (MCMC) sampling, optimization, and learning. Their convergence properties for unconstrained non-convex optimization and learning problems have been studied widely in the last few years. Other work has examined projected Langevin algorithms for sampling from log-concave distributions restricted to convex compact sets. For learning and optimization, log-concave distributions correspond to convex losses. In this paper, we analyze the case of non-convex losses with compact convex constraint sets and IID external data variables. We term the resulting method the projected stochastic gradient Langevin algorithm (PSGLA). We show the algorithm achieves a deviation of $O(T^{-1/4}(\log T)^{1/2})$ from its target distribution in 1-Wasserstein distance. For optimization and learning, we show that the algorithm achieves $\epsilon$-suboptimal solutions, on average, provided that it is run for a time that is polynomial in $\epsilon^{-1}$ and slightly super-exponential in the problem dimension., Comment: 45 pages. Under Review for COLT 2021

Complex networked systems can be modeled and represented as graphs, with nodes representing the agents and the links describing the dynamic coupling between them. The fundamental objective of network identification for dynamic systems is to identify causal influence pathways. However, dynamically related data-streams that originate from different sources are prone to corruption caused by asynchronous time stamps, packet drops, and noise. In this article, we show that identifying causal structure using corrupt measurements results in the inference of spurious links. A necessary and sufficient condition that delineates the effects of corruption on a set of nodes is obtained. Our theory applies to nonlinear systems, and systems with feedback loops. Our results are obtained by the analysis of conditional directed information in dynamic Bayesian networks. We provide consistency results for the conditional directed information estimator that we use by showing almost-sure convergence., Comment: Paper under review. Submitted to IEEE TAC journal

Complex networked systems can be modeled as graphs with nodes representing the agents and links describing the dynamic coupling between them. Previous work on network identification has shown that the network structure of linear time-invariant (LTI) systems can be reconstructed from the joint power spectrum of the data streams. These results assumed that data is perfectly measured. However, real-world data is subject to many corruptions, such as inaccurate time-stamps, noise, and data loss. We show that identifying the structure of linear time-invariant systems using corrupt measurements results in the inference of erroneous links. We provide an exact characterization and prove that such erroneous links are restricted to the neighborhood of the perturbed node. We extend the analysis of LTI systems to the case of Markov random fields with corrupt measurements. We show that data corruption in Markov random fields results in spurious probabilistic relationships in precisely the locations where spurious links arise in LTI systems., Comment: Submitted to IEEE TAC. Under review

This paper presents algorithms for non-zero sum nonlinear constrained dynamic games with full information. Such problems emerge when multiple players with action constraints and differing objectives interact with the same dynamic system. They model a wide range of applications including economics, defense, and energy systems. We show how to exploit the temporal structure in projected gradient and Douglas-Rachford (DR) splitting methods. The resulting algorithms converge locally to open-loop Nash equilibria (OLNE) at linear rates. Furthermore, we extend stagewise Newton method to find a local feedback policy around an OLNE. In the of linear dynamics and polyhedral constraints, we show that this local feedback controller is an approximated feedback Nash equilibrium (FNE). Numerical examples are provided., Comment: 19 pages. Full version of submission to ACC 2020

Recursive least-squares algorithms often use forgetting factors as a heuristic to adapt to non-stationary data streams. The first contribution of this paper rigorously characterizes the effect of forgetting factors for a class of online Newton algorithms. For exp-concave and strongly convex objectives, the algorithms achieve the dynamic regret of $\max\{O(\log T),O(\sqrt{TV})\}$, where $V$ is a bound on the path length of the comparison sequence. In particular, we show how classic recursive least-squares with a forgetting factor achieves this dynamic regret bound. By varying $V$, we obtain a trade-off between static and dynamic regret. In order to obtain more computationally efficient algorithms, our second contribution is a novel gradient descent step size rule for strongly convex functions. Our gradient descent rule recovers the order optimal dynamic regret bounds described above. For smooth problems, we can also obtain static regret of $O(T^{1-\beta})$ and dynamic regret of $O(T^\beta V^*)$, where $\beta \in (0,1)$ and $V^*$ is the path length of the sequence of minimizers. By varying $\beta$, we obtain a trade-off between static and dynamic regret., Comment: Accepted by AAAI 2020

One of the primary challenges of system identification is determining how much data is necessary to adequately fit a model. Non-asymptotic characterizations of the performance of system identification methods provide this knowledge. Such characterizations are available for several algorithms performing open-loop identification. Often times, however, data is collected in closed-loop. Application of open-loop identification methods to closed-loop data can result in biased estimates. One method used by subspace identification techniques to eliminate these biases involves first fitting a long-horizon autoregressive model, then performing model reduction. The asymptotic behavior of such algorithms is well characterized, but the non-asymptotic behavior is not. This work provides a non-asymptotic characterization of one particular variant of these algorithms. More specifically, we provide non-asymptotic upper bounds on the generalization error of the produced model, as well as high probability bounds on the difference between the produced model and the finite horizon Kalman Filter.

Dynamic games arise when multiple agents with differing objectives control a dynamic system. They model a wide variety of applications in economics, defense, energy systems and etc. However, compared to single-agent control problems, the computational methods for dynamic games are relatively limited. As in the single-agent case, only specific dynamic games can be solved exactly, so approximation algorithms are required. In this paper, we show how to extend a recursive Newton's algorithm and the popular differential dynamic programming (DDP) for single-agent optimal control to the case of full-information non-zero sum dynamic games. In the single-agent case, the convergence of DDP is proved by comparison with Newton's method, which converges locally at a quadratic rate. We show that the iterates of Newton's method and DDP are sufficiently close for the DDP to inherit the quadratic convergence rate of Newton's method. We also prove both methods result in an open-loop Nash equilibrium and a local feedback $O(\epsilon^2)$-Nash equilibrium. Numerical examples are provided., Comment: 19 pages. The shortened version was accepted at CDC 2019. arXiv admin note: substantial text overlap with arXiv:1809.08302

In this paper, we present simple algorithms for Dueling Bandits. We prove that the algorithms have regret bounds for time horizon T of order O(T^rho ) with 1/2 <= rho <= 3/4, which importantly do not depend on any preference gap between actions, Delta. Dueling Bandits is an important extension of the Multi-Armed Bandit problem, in which the algorithm must select two actions at a time and only receives binary feedback for the duel outcome. This is analogous to comparisons in which the rater can only provide yes/no or better/worse type responses. We compare our simple algorithms to the current state-of-the-art for Dueling Bandits, ISS and DTS, discussing complexity and regret upper bounds, and conducting experiments on synthetic data that demonstrate their regret performance, which in some cases exceeds state-of-the-art.

We propose algorithms for online principal component analysis (PCA) and variance minimization for adaptive settings. Previous literature has focused on upper bounding the static adversarial regret, whose comparator is the optimal fixed action in hindsight. However, static regret is not an appropriate metric when the underlying environment is changing. Instead, we adopt the adaptive regret metric from the previous literature and propose online adaptive algorithms for PCA and variance minimization, that have sub-linear adaptive regret guarantees. We demonstrate both theoretically and experimentally that the proposed algorithms can adapt to the changing environments., Comment: This paper is accepted by ICML 2019

Dynamic games arise when multiple agents with differing objectives choose control inputs to a dynamic system. Dynamic games model a wide variety of applications in economics, defense, and energy systems. However, compared to single-agent control problems, the computational methods for dynamic games are relatively limited. As in the single-agent case, only very specialized dynamic games can be solved exactly, and so approximation algorithms are required. This paper extends the differential dynamic programming algorithm from single-agent control to the case of non-zero sum full-information dynamic games. The method works by computing quadratic approximations to the dynamic programming equations. The approximation results in static quadratic games which are solved recursively. Convergence is proved by showing that the algorithm iterates sufficiently close to iterates of Newton's method to inherit its convergence properties. A numerical example is provided., Comment: 11 Pages. Full version of ACC Submission

We propose the algorithms for online convex optimization which lead to cumulative squared constraint violations of the form $\sum\limits_{t=1}^T\big([g(x_t)]_+\big)^2=O(T^{1-\beta})$, where $\beta\in(0,1)$. Previous literature has focused on long-term constraints of the form $\sum\limits_{t=1}^Tg(x_t)$. There, strictly feasible solutions can cancel out the effects of violated constraints. In contrast, the new form heavily penalizes large constraint violations and cancellation effects cannot occur. Furthermore, useful bounds on the single step constraint violation $[g(x_t)]_+$ are derived. For convex objectives, our regret bounds generalize existing bounds, and for strongly convex objectives we give improved regret bounds. In numerical experiments, we show that our algorithm closely follows the constraint boundary leading to low cumulative violation., Comment: The NeurIPS version of the stepsize setup for strongly convex case is not correct. Please see this arxiv setup(e.g.,move the H1 to the denominator)

This paper proposes a semidefinite programming based method for estimating moments of a stochastic hybrid system (SHS). For polynomial SHSs -- which consist of polynomial continuous vector fields, reset maps, and transition intensities -- the dynamics of moments evolve according to a system of linear ordinary differential equations. However, it is generally not possible to solve the system exactly since time evolution of a specific moment may depend upon moments of order higher than it. One way to overcome this problem is to employ so-called moment closure methods that give point approximations to moments, but these are limited in that accuracy of the estimations is unknown. We find lower and upper bounds on a moment of interest via a semidefinite program that includes linear constraints obtained from moment dynamics, along with semidefinite constraints that arise from the non-negativity of moment matrices. These bounds are further shown to improve as the size of semidefinite program is increased. The key insight in the method is a reduction from stochastic hybrid systems with multiple discrete modes to a single-mode hybrid system with algebraic constraints. We further extend the scope of the proposed method to a class of non-polynomial SHSs which can be recast to polynomial SHSs via augmentation of additional states. Finally, we illustrate the applicability of results via examples of SHSs drawn from different disciplines.

This paper proposes a methodology to estimate characteristic functions of stochastic differential equations that are defined over polynomials and driven by L\'evy noise. For such systems, the time evolution of the characteristic function is governed by a partial differential equation; consequently, the stationary characteristic function can be obtained by solving an ordinary differential equation (ODE). However, except for a few special cases such as linear systems, the solution to the ODE consists of unknown coefficients. These coefficients are closely related with the stationary moments of the process, and bounds on these can be obtained by utilizing the fact that the characteristic function is positive definite. These bounds can be further used to find bounds on other higher order stationary moments and also estimate the stationary characteristic function itself. The method is finally illustrated via examples.

Stochastic dynamical systems often contain nonlinearities which make it hard to compute probability density functions or statistical moments of these systems. For the moment computations, nonlinearities in the dynamics lead to unclosed moment dynamics; in particular, the time evolution of a moment of a specific order may depend both on moments of order higher than it and on some nonlinear function of other moments. The moment closure techniques are used to find an approximate, close system of equations the moment dynamics. In this work, we extend a moment closure technique based on derivative matching that was originally proposed for polynomial stochastic systems with discrete states to continuous state stochastic systems to continuous state stochastic differential equations, with both polynomial and trigonometric nonlinearities. We validate the technique using two examples of nonlinear stochastic systems.

This paper develops a unified methodology for probabilistic analysis and optimal control design for jump diffusion processes defined by polynomials. For such systems, the evolution of the moments of the state can be described via a system of linear ordinary differential equations. Typically, however, the moments are not closed and an infinite system of equations is required to compute statistical moments exactly. Existing methods for stochastic analysis, known as closure methods, focus on approximating this infinite system of equations with a finite dimensional system. This work develops an alternative approach in which the higher order terms, which are approximated in closure methods, are viewed as inputs to a finite-dimensional linear control system. Under this interpretation, upper and lower bounds of statistical moments can be computed via convex linear optimal control problems with semidefinite constraints. For analysis of steady-state distributions, this optimal control problem reduces to a static semidefinite program. These same optimization problems extend automatically to stochastic optimal control problems. For minimization problems, the methodology leads to guaranteed lower bounds on the true optimal value. Furthermore, we show how an approximate optimal control strategy can be constructed from the solution of the semidefinite program. The results are illustrated using numerous examples.

In the stochastic formulation of chemical kinetics, the stationary moments of the population count of species can be described via a set of linear equations. However, except for some specific cases such as systems with linear reaction propensities, the moment equations are underdetermined as a lower order moment might depend upon a higher order moment. Here, we propose a method to find lower, and upper bounds on stationary moments of molecular counts in a chemical reaction system. The method exploits the fact that statistical moments of any positive-valued random variable must satisfy some constraints. Such constraints can be expressed as nonlinear inequalities on moments in terms of their lower order moments, and solving them in conjugation with the stationary moment equations results in bounds on the moments. Using two examples of biochemical systems, we illustrate that not only one obtains upper and lower bounds on a given stationary moment, but these bounds also improve as one uses more moment equations and utilizes the inequalities for the corresponding higher order moments. Our results provide avenues for development of moment approximations that provide explicit bounds on moment dynamics for systems whose dynamics are otherwise intractable.

This paper presents a method to approximately solve stochastic optimal control problems in which the cost function and the system dynamics are polynomial. For stochastic systems with polynomial dynamics, the moments of the state can be expressed as a, possibly infinite, system of deterministic linear ordinary differential equations. By casting the problem as a deterministic control problem in moment space, semidefinite programming is used to find a lower bound on the optimal solution. The constraints in the semidefinite program are imposed by the ordinary differential equations for moment dynamics and semidefiniteness of the outer product of moments. From the solution to the semidefinite program, an approximate optimal control strategy can be constructed using a least squares method. In the linear quadratic case, the method gives an exact solution to the optimal control problem. In more complex problems, an infinite number of moment differential equations would be required to compute the optimal control law. In this case, we give a procedure to increase the size of the semidefinite program, leading to increasingly accurate approximations to the true optimal control strategy., Comment: Submitted to IEEE Conference on Decision and Control '16

This paper studies the stability of sampled and networked control systems with sampling and communication times governed by probabilistic clocks. The clock models have few restrictions, and can be used to model numerous phenomena such as deterministic sampling, jitter, and transmission times of packet dropping networks. Moreover, the stability theory can be applied to an arbitrary number of clocks with different distributions, operating asynchronously. The paper gives Lyapunov-type sufficient conditions for stochastic stability of nonlinear networked systems. For linear systems, the paper gives necessary and sufficient conditions for exponential mean square stability, based on linear matrix inequalities. In both the linear and nonlinear cases, the Lyapunov inequalities are constructed from a simple linear combination of the classical inequalities from continuous and discrete time. Crucially, the stability theorems only depend on the mean sampling intervals. Thus, they can be applied with only limited statistical information about the clocks. The Lyapunov theorems are then applied to systems with multirate sampling, asynchronous communication, delays, and packet losses., Comment: This paper has been withdrawn due to mathematical errors

This paper presents an explicit solution to a two player distributed LQR problem in which communication between controllers occurs across a communication link with varying delay. We extend known dynamic programming methods to accommodate this varying delay, and show that under suitable assumptions, the optimal control actions are linear in their information, and that the resulting controller has piecewise linear dynamics dictated by the current effective delay regime., Comment: Extended version of IFAC '14 submission

We consider temporal logic verification of (possibly nonlinear) dynamical systems evolving over continuous state spaces. Our approach combines automata-based verification and the use of so-called barrier certificates. Automata-based verification allows the decomposition the verification task into a finite collection of simpler constraints over the continuous state space. The satisfaction of these constraints in turn can be (potentially conservatively) proved by appropriately constructed barrier certificates. As a result, our approach, together with optimization-based search for barrier certificates, allows computational verification of dynamical systems against temporal logic properties while avoiding explicit abstractions of the dynamics as commonly done in literature.

This paper examines stochastic optimal control problems in which the state is perfectly known, but the controller's measure of time is a stochastic process derived from a strictly increasing L\'evy process. We provide dynamic programming results for continuous-time finite-horizon control and specialize these results to solve a noisy-time variant of the linear quadratic regulator problem and a portfolio optimization problem with random trade activity rates. For the linear quadratic case, the optimal controller is linear and can be computed from a generalization of the classical Riccati differential equation., Comment: Submitted to IEEE TAC

This paper gives a new solution to the output feedback H2 problem for quadratically invariant communication delay patterns. A characterization of all stabilizing controllers satisfying the delay constraints is given and the decentralized H2 problem is cast as a convex model matching problem. The main result shows that the model matching problem can be reduced to a finite-dimensional quadratic program. A recursive state-space method for computing the optimal controller based on vectorization is given., Comment: Draft submitted to IEEE Transactions on Automatic Control

This work presents the solution to a class of decentralized linear quadratic state-feedback control problems, in which the plant and controller must satisfy the same combination of delay and sparsity constraints. Using a novel decomposition of the noise history, the control problem is split into independent subproblems that are solved using dynamic programming. The approach presented herein both unifies and generalizes many existing results.

This paper gives a new solution to the output feedback H_2 model matching problem for a large class of delayed information sharing patterns. Existing methods for such problems typically reduce the decentralized problem to a centralized problem of higher state dimension. In contrast, the controller given in this paper is constructed from the solutions to the centralized control and estimation Riccati equations for the original system. The problem is solved by decomposing the controller into two components. One is centralized, but delayed, while the other is decentralized with finite impulse response (FIR). It is then shown that the optimal controller can be constructed through a combination of centralized spectral factorization and quadratic programming., Comment: Preliminary draft of conference paper