previous day
all days

View: session overviewtalk overview

09:00-09:45 Session 5: Plenary 2: Structured Reinforcement Learning

Alexandre Proutiere

Professor in the Department of Automatic Control, KTH Royal Institute of Technology

 Abstract: Real-world Reinforcement Learning (RL) problems often concern dynamical systems with large state and action spaces. For such systems, it is essential to  identify and exploit any structural property of the system dynamics and reward function. Leveraging a known structure in turn may help minimizing the exploration of sub-optimal (state, action) pairs, and speed up the learning process. Modern RL algorithms actually implicitly impose some structure either in the model parameters (transition probabilities and reward function) or directly in the Q-function. Despite the successes of these recent algorithms, our understanding of structured RL problems remains limited. In this talk, we provide an overview of both theoretical and practical methods developed recently to tackle large-scale structured RL problems, and list important remaining challenges in the field.


Location: Q1
10:00-12:00 Session 6A: Automotive technology
Location: Q1
Control for Autonomous Heavy Duty Vehicles and Buses

ABSTRACT. The development of autonomous vehicles redefines exiting and creates many new control challenges. Also, there are significant differences between passenger cars and heavy duty vehicles. This talk gives a status update on current challenges from an industry perspective.

With no human driver the importance of estimation of vehicle and environment parameters is increased. How do we replicate the sanity check function of the driver? Most current work on autonomous vehicles is focused on nominal weather and driving surface conditions, but in the real world rain, snow, and ice have to be handled too. The need for creating a safe vehicle when operating with a human safety driver leads to restrictions being defined for actuators. These restrictions in turn affect the possible control performance. How do we guard against problems arising from the interaction between system components in practice?

System-level simulation for electric vehicles: an active safety case

ABSTRACT. Physical system modeling at an appropriate fidelity level is paramount to successful model-based design. During this session, MathWorks will illustrate how battery cooling, electric powertrain and vehicle dynamics can be represented in an active safety context. Furthermore, design exploration and model calibration based upon optimization and statistical techniques will be showcased to increase confidence in physical models.

Predictive velocity control in a hilly terrain over long look-ahead horizons

ABSTRACT. This paper presents a computational efficient velocity control of vehicles driving in a possibly hilly terrain and over long look-ahead horizons that may stretch to hundreds of kilometers. The controller decouples gear scheduling into an offline optimization problem, from the remaining optimization problem that governs two real-valued states. One of the states, the travel time, is adjoined to the objective by applying the necessary optimality conditions, which results into an online optimization problem that has kinetic energy as the single state. Finally an inner approximation is proposed for the online problem to obtain a quadratic program that can be solved efficiently. The efficiency of the proposed controller is shown for different horizon lengths.

Optimal speed control of a heavy-duty vehicle in the presence of traffc lights

ABSTRACT. The fuel consumption of heavy-duty vehicles in urban driving is strongly dependent on the acceleration and braking of the vehicles. In intersections with traffic lights, large amount of fuel can be saved by adapting the velocity to the phases of the lights. In this paper, a heavy-duty vehicle obtains information about the future signals of traffic lights within a specific horizon. In order to minimize the fuel consumption, the driving scenario is formulated as an optimal control problem. The optimal control is found by applying a model predictive controller, solving at each iteration a quadratic program. In such problem formulation, the constraints imposed by the traffic lights are formulated using a linear approximation of time. Since the fuel-optimal velocity can deviate strongly from how vehicles normally drive, constraints on the allowed velocity are imposed. Simulations are performed in order to investigate how the horizon length of the information from the traffic lights influences the fuel consumption. Compared to a benchmark vehicle without knowledge of future light signals, the proposed controller using a control horizon of 1000 m saves 32 % of energy while reducing the trip time by 5 %. Increasing the control horizon further does not improve the results.

Design and formal verification of a safestop supervisor for an automated vehicle

ABSTRACT. Autonomous vehicles apply pertinent planning and control algorithms under different driving conditions. The mode switch between these planner algorithms should also beautonomous. On top of these nominal planners, a safe fallback routine is needed to bring the vehicle to stop at a safe position if nominal operational conditions are violated, such as if the vehicle experiences a system failure. This paper develops a supervisor to manage all requirements for mode switching between nominal planners, and additional requirements for switching to a safe stop trajectory planner that acts as the fallback routine. The supervisor is developed via a model-based approach and its abstraction is formally verified by model checking. The supervisor is implemented for the Research Concept Vehicle (RCV), an experimental research and demonstration vehicle developed at KTH Royal Institute of Technology. While testing, the RCV with the implemented supervisor was able to autonomously drive in a safe manner between two parking lots and the safe stop trajectory planner could successfully bring the RCV to a safe stop during GPS sensor failure.

Path Planning using convex optimization for automated driving in urban settings

ABSTRACT. Driving in urban environments often presents difficult situations that require expert maneuvering of a vehicle. These situations become more difficult when considering large dimension vehicles, such as busses. In this work we present a framework for path planning that tackles the difficult maneuvering task of busses in urban environments. The path planning is formulated as a convex optimization problem in a road-aligned frame. This type of frame introduces a highly non-linear distortion on the vehicle bodies, motivating the development of a novel type of approximation for obstacle avoidance constraints. Furthermore we propose new optimization goals, that unlike other path planning approaches, explicitly distinguish between drivable, sweepable and obstacle regions. The result is a path planner that is able to maneuver in complex urban environments, taking full advantage of the bus' unusual characteristics.

10:00-12:00 Session 6B: Machine learning
Location: Q2
Linearly constrained Gaussian processes

ABSTRACT. We consider a modification of the covariance function in Gaussian processes to correctly account for known linear operator constraints. By modeling the target function as a transformation of an underlying function, the constraints are explicitly incorporated in the model such that they are guaranteed to be fulfilled by any sample drawn or prediction made. We also propose a constructive procedure for designing the transformation operator and illustrate the result on both simulated and real-data examples.

Gradient compression for communication-limited convex optimization

ABSTRACT. Data-rich applications in machine-learning and control have motivated an intense research on large-scale optimization. Novel algorithms have been proposed and shown to have order-optimal convergence rates in terms of iteration counts. However, in actual implementations, their performance is severely degraded by the cost of exchanging large gradient vectors between computing nodes. Several lossy gradient compression heuristics have recently been proposed to reduce communications, but few theoretical results exist that quantify how they impact algorithm convergence.

This paper establishes and strengthens the convergence guarantees for gradient descent under a family of gradient compression techniques. For convex optimization problems, we derive admissible step sizes and quantify both the number of iterations and the number of bits that need to be exchanged to reach a target accuracy. Finally, we validate the performance of different gradient compression techniques in simulations. The numerical results highlight the properties of different gradient compression algorithms and confirm that fast convergence with limited information exchange is indeed possible.

Data-driven impulse response regularization via deep learning

ABSTRACT. We consider the problem of impulse response estimation for stable linear single-input single-output systems. It is a well-studied problem where flexible non-parametric models recently offered a leap in performance compared to the classical finite-dimensional model structures. Inspired by this development and the success of deep learning we propose a new flexible data-driven model. Our experiments indicate that the new model is capable of exploiting even more of the hidden patterns that are present in the input-output data as compared to the non-parametric models.

Inverse filtering for linear Gaussian state-space models

ABSTRACT. This paper considers inverse filtering problems for linear Gaussian state-space systems. We consider three problems of increasing generality where, given the posterior (sample path of mean and covariance), the aim is to reconstruct the measurements and certain unknown sensor parameters, such as the observation likelihood. The paper is motivated by applications where one wishes to calibrate a Bayesian estimator based on remote observations of the posterior estimates, e.g., determine how accurate an adversary’s sensors are. We evaluate the robustness of the proposed inverse filtering algorithms with respect to noise (e.g., quantization errors) in numerical simulations.

Improving the particle filter in high dimensions using conjugate artificial process noise

ABSTRACT. The particle filter is one of the most successful methods for state inference and identification of general non-linear and non-Gaussian models. However, standard particle filters suffer from degeneracy of the particle weights, in particular for high-dimensional problems. We propose a method for improving the performance of the particle filter for certain challenging state space models, with implications for high-dimensional inference. First we approximate the model by adding artificial process noise in an additional state update, then we design a proposal that combines the standard and the locally optimal proposal. This results in a bias-variance trade-off, where adding more noise reduces the variance of the estimate but increases the model bias. The performance of the proposed method is evaluated on a linear Gaussian state space model and on the non-linear Lorenz'96 model. For both models we observe a significant improvement in performance over the standard particle filter.

Distributed multi-agent Gaussian regression via finite-dimensional approximations

ABSTRACT. We consider the problem of distributedly estimating Gaussian processes in multi-agent frameworks. Each agent collects few measurements and aims to collaboratively reconstruct a common estimate based on all data. Agents are assumed with limited computational and communication capabilities and to gather M noisy measurements in total on input locations independently drawn from a known common probability density. The optimal solution would require agents to exchange all the M input locations and measurements and then invert an M × M matrix, a non-scalable task. Differently, one may consider suboptimal approaches using the first E orthonormal eigenfunctions obtained from the Karhunen-Loève (KL) expansion of the chosen kernel, where typically E << M . The benefits are that the computation and communication complexities scale with E and not with M , and computing the required statistics can be performed via standard average consensus algorithms. In this specific framework we develop new distributed strategies relying on Stein’s unbiased risk estimate (SURE) paradigms for tuning the regularization parameters of these estimation schemes, that are applicable to generic basis functions (thus not necessarily kernel eigenfunctions) and that can again be implemented via average consensus. The proposed distributed tuning strategies are finally tested on both synthetic and real field data.

10:00-12:00 Session 6C: Power systems
Location: Q31
Wind estimation in Skeldar V-200

ABSTRACT. For all aerial vehicles, knowing the wind conditions in which you are flying is of great importance. This information can be used to optimize fuel consumption, to satisfy safety restrictions, for post-flight analysis etc. In this paper, we present a method, used in the Skeldar V-200 helicopter UAV, for estimating the wind speed and direction based on sensor measurements from a pitot-static tube and a GPS receiver.

On the coherence of large-scale networks with distributed PI and PD control

ABSTRACT. We consider distributed control of double-integrator networks, where agents are subject to stochastic disturbances. We study performance of such networks in terms of coherence, defined through an H2 norm metric that represents the variance of nodal state fluctuations. Specifically, we address known performance limitations of the standard consensus protocol, which cause this variance to scale unboundedly with network size for a large class of networks. We propose distributed proportional integral (PI) and proportional derivative (PD) controllers that relax these limitations and achieve bounded variance, in cases where agents can access an absolute measurement of one of their states. This case applies to, for example, frequency control of power networks and vehicular formation control with limited sensing. We discuss optimal tuning of the controllers with respect to network coherence and demonstrate our results in simulations.

Optimal line current limitation of CPLs fed through Input LC filters

ABSTRACT. The achievable bandwidth of power control of constant power loads (CPLs) fed through input LC filters is known to be limited by the input filter resonance frequency. This contribution explicitly shows how this bandwidth restriction also limits dynamic performance of line current limitation, functionality regulated for traction systems by European standards. Fulfillment of such requirements may hence restrict design of the CPL input filter. It is also shown that simple first-order controllers are sufficient for close to optimal current limitation performance, which is verified through simulations.

Fundamental performance limitations utilizing HVDC to damp interarea modes

ABSTRACT. This paper considers power oscillation damping (POD) using active power modulation of high-voltage dc (HVDC) transmissions. An analytical study of how the proximity between interarea modal frequencies in two interconnected asynchronous grids puts a fundamental limit to the achievable performance is presented. It is shown that the ratio between modal frequencies is the sole factor determining the achievable nominal performance. To illustrate the inherent limitations, simulations using a simple proportional controller tuned to optimize performance in terms of POD are done. The influence of limited system information and unmodeled dynamics is shown.

Adaption of 2-D look-up tables applied to OCV-curves for aged battery cells

ABSTRACT. In online automotive applications it is common to use look-up tables, or maps, to model nonlinearities in component models that are to be valid over large operating ranges. If the component characteristics change with ageing or wear, these look-up tables must be updated online. For 2-D look-up tables, the existing methods in the literature only adapt the observable parameters in the look-up table, which means that parameters in operating points that have not been visited for a long time may be far from their true values (see e.g. (Hockerdal et al., 2011)).

In this work, correlations between different operating points are used to also update non-observable parameters of the look-up table. The method is applied to Open Circuit Voltage (OCV) curves for aged battery cells. From laboratory experimental data it is demonstrated that the proposed method can significantly reduce the average deviation from an aged OCV-curve compared to keeping the OCV curve from the beginning of the cell’s life, both for observable and non-observable parameters.

The interconnection of quadratic droop voltage controllers is a Lotka-Volterra system: implications for stability analysis

ABSTRACT. This paper studies the stability of voltage dynamics for a power network in which nodal voltages are controlled by means of quadratic droop controllers with nonlinear AC reactive power as inputs. We show that the voltage dynamics is a Lotka-Volterra system, which is a class of nonlinear positive systems. We study the stability of the closed-loop system by proving a uniform ultimate boundedness result and investigating conditions under which the network is cooperative. We then restrict to study the stability of voltage dynamics under a decoupling assumption (i.e., zero relative angles). We analyze the existence and uniqueness of the equilibrium in the interior of the positive orthant for the system and prove an asymptotic stability result.

10:00-12:00 Session 6D: Networked systems II
Location: Q34
Network centrality measures for efficient driver node placement

ABSTRACT. This paper investigates the problem of driver node placement in complex networks. We identify two different node characteristics that are important if we want to limit the control energy required to steer the network state. The first characteristic is related to the network impact of a node, i.e. to what extent it influences the other nodes, while the other characteristic instead is related to the ability to control a node indirectly from other nodes. These two properties can be combined into node rankings that can be used for driver node placement. By doing so, conflicting control energy requirements, like minimizing the average energy needed to steer the state in any direction and the energy needed for the worst direction, can be simultaneously taken into account. Algebraically, such placement exploit the non-normality of the adjacency matrix of the graph representing the network.

LQG control and scheduling co-design for wireless sensor and actuator networks

ABSTRACT. This paper studies a co-design problem of control, scheduling and routing over a multi-hop sensor and actuator network subject to energy-saving consideration. Sensors are observing multiple independent linear systems and transmit their data to actuators in which controllers are co-located. We formulate an optimization problem, minimizing a linear combination of the averaged linear quadratic Gaussian control performance and the averaged transmission energy consumption. Optimal solutions are derived and their performance is illustrated in a numerical example. Algorithms to recongure routing between sensors and actuators in case of link disconnection are also provided.

Measurement scheduling in networked control systems

ABSTRACT. We discussed a new reachability problem for networked control system where a master node -the controller- broadcasts commands to a set of slave nodes, which must take turn to relay back state measurements. This problem finds applications in some robotics and intelligent transportation systems setups. Constraints on communication demand a joint design of the controller and the measurement schedule. We proved that the problem is formally equivalent to the Pinwheel Problem from scheduling theory, and building on this result we provide conditions for schedulability and reachability.

Fast and resource-efficient control of wireless cyber-physical systems

ABSTRACT. Cyber-physical systems (CPS) integrate the physical world with computing and communication. This poses new challenges for control design and analysis, especially if communication is carried out over wireless channels. We propose to address these challenges by a tight integration of communication and control system, and by leveraging recent machine learning techniques such as deep reinforcement learning (DRL).

On the scale free control of mass chains

ABSTRACT. We give a method for designing a mechanical impedance to suppress the propagation of disturbances along a chain of masses. The key feature of our method is that it is scale free. This means that it can be used to give a single, fixed, design, with provable performance guarantees in mass chains of any length. We illustrate the approach by designing a bidirectional control law in a vehicle platoon in a manner that is independent of the number of vehicles in the platoon.

Secure patching of an output-feedback controller for a class of nonlinear systems under adversarial attack

ABSTRACT. We consider a class of nonlinear systems for which an observer-based output-feedback controller is updated at discrete time instances. However, the received update or patch can be compromised by the attacker to drive the system to instability. In this paper, we provide a checkable condition to ensure that the received patch has not been tampered with. Moreover, we guarantee that the application of the tamper-free patch ensures global asymptotic stability of the control system by choosing the update time instances appropriately. The secure patch update protocol is illustrated on an example involving the output-feedback synchronization of two neuron population models, where the observer gains are updated at discrete time instances.

13:00-14:00 Session 7: Interactive session II
Location: Q33 & Q36
Multi-criteria optimization of PD Controllers for plants including integral action

ABSTRACT. A simple design method for robust PD controllers is presented for systems including integral action. The design method is based on a multi-criteria optimal control formulation, which is easily solved by a few lines of MATLAB code. Most criteria are based on Hinf measures, but since the focus is on reference signal tracking and not load disturbance compensation, the settling time is also included as a relevant performance measure. Since a PD controller is equivalent to a lead filter, the optimal PD controller is compared with ordinary textbook design rules for lead filters. More specifically, it is shown that the common recommendation to place the mid-frequency of the lead filter at the desired gain crossover frequency often gives bad servo performance. The suggested optimal solution, still including robustness and control activity adjustments, is on the other hand a simple and flexible design method for arbitrary plants including integral action.

Secure binary hypothesis testing: performance metric and sensor deployment strategy

ABSTRACT. This paper studies how to deploy sensors in the context of detection in adversarial environments. A fusion center is performing a binary hypothesis testing based on measurements from remotely deployed heterogeneous sensors. An attacker may compromise some of the deployed sensors, which send arbitrary measurements to the fusion center. The problems of interest are: 1) to characterize the performance of the system under attack and thus develop a performance metric; 2) to deploy sensors within a cost budget, such that the proposed performance metric is maximized. In this extended abstract, we briefly present a Nash equilibrium pair of detection algorithm and attack strategy, the performance of which is utilized as the metric. We also sketch out two approximately optimal sensor deployment strategies.

Calibrated predictive uncertainty in classification with neural networks

ABSTRACT. Probabilistic classifiers output a probability distribution on target classes rather than just a class prediction. Besides providing a clear separation of prediction and decision making, the main advantage of probabilistic models is their ability to represent uncertainty about their predictions. In safety-critical applications, it is pivotal for a model to possess an adequate sense of uncertainty, which, for probabilistic models, translates into outputting probability distributions that accurately match the empirical frequencies observed from realized outcomes. In this work, we explore different ways to quantify, visualize, and enhance calibration of neural networks in probabilistic classification.

Asymptotic and finite-time almost global attitude tracking: representations free approach

ABSTRACT. In this paper, the attitude tracking problem is considered using the rotation matrices. Due to the inherent topological restriction, it is impossible to achieve global stability with any continuous attitude control system on SO(3). Hence in this work, we propose some control protocols that achieve almost global tracking asymptotically and in finite time, respectively. In these protocols, no world frame is needed and only relative state information are requested. For the closed-loop systems, Filippov solutions and non-smooth analysis techniques are adopted to handle the discontinuities.

A risk-theoretical approach to H2-optimal control under covert attacks

ABSTRACT. We consider the control design problem of optimizing the H2 performance of a closed-loop system despite the presence of a malicious covert attacker. It is assumed that the attacker has incomplete knowledge on the true process we are controlling. To account for this uncertainty, we employ different measures of risk from the so called family of coherent measures of risk. In particular, we compare the closed-loop performance when a nominal value is used, with three different measures of risk: average risk, worst-case scenario and conditional value- at-risk (CVaR). Additionally, applying the approach from a previous work, we derive a convex formulation for the control design problem when CVaR is employed to quantify the risk. A numerical example illustrates the advantages of our novel approach.

Polynomial nonlinear state-space modeling applied to F-16 aircraft benchmark data

ABSTRACT. A polynomial nonlinear state-space (PNLSS) model can capture many nonlinear behaviors and is successfully used in a large range of applications. In this work, some capabilities and shortcomings of the PNLSS method are shown on experimental data from an F-16 aircraft, which are available as a benchmark.

Control refinement for discrete-time descriptor systems: a behavioural approach via simulation relations

ABSTRACT. The analysis of industrial processes, modelled as descriptor systems, is often computationally hard due to the presence of both algebraic couplings and difference equations of high order. In this paper, we introduce a control refinement notion for these descriptor systems that enables analysis and control design over related reduced-order systems. Utilising the behavioural framework, we extend upon the standard hierarchical control refinement for ordinary systems and allow for algebraic couplings inherent to descriptor systems.

POLO: POLicy-based optimization library

ABSTRACT. Thanks to the developments in data storage technologies and wide adoption of Internet-based devices, it has become commonplace to collect, share and store bulks of data. This, in turn, has led to an increase in problem dimensions in data-driven machine-learning applications. To tackle this dimensionality problem, we need new, parallel optimization algorithms that exploit problem structure and make good use of computing resources such as the multiple cores on the CPUs and GPUs as well as different nodes in the cloud. However, algorithm design is a non-trivial process that involves prototyping and testing new ideas, and finally deploying them on different execution engines. In this work, we present POLO --- a C++ library that emphasizes ease-of-use, flexibility and efficiency in algorithm design. It uses multiple inheritance and template programming to decompose algorithms into essential policies and facilitate code reuse. With its clear separation between algorithm and execution policies, it provides researchers with a simple and powerful platform to prototype ideas and test them against different options. POLO enables users to move seamlessly from serial to multi-threaded shared-memory to multi-node distributed-memory executors. We demonstrate how POLO allows users to implement state-of-the-art asynchronous parallel optimization algorithms in just a few lines of code and report experiment results from shared and distributed-memory computing architectures.

Dynamics of mass chains via repeated compositions of Möbius transformations

ABSTRACT. This note studies the disturbance amplification in an interconnected heterogeneous mass chain in which one end is connected to a movable point. Each interconnection is represented by a general mechanical impedance function. Recurrence relations with respect to the number of masses are derived for the scalar transfer functions from the movable point displacement to a given intermass displacement. In particular, each relation takes the form of a Möbius transformation.

Optimal privacy filter for LQG control with a weighted directed information measure

ABSTRACT. We study for a dynamical agent the joint design of a perturbation-based privacy filter and a controller which is situated in a cloud, within a linear quadratic Gaussian (LQG) framework. The major objective of our study is to shed new light on the trade-off between the privacy level and control performance and on the corresponding achievable region. In our formulation, the state of the agent is observed by the privacy filter, and the perturbed state is available at the controller with one-step delay. We measure the control performance by a cost function that penalizes the state deviation and control effort, and measure the privacy level by a loss function that quantifies the leakage of the private information. We formulate the problem as a nonzero-sum stochastic game, in which the controller seeks to maximize the control performance, while the privacy filter seeks to maximize the control performance subject to a constraint on the privacy level. We characterize the structures of the optimal control policy and optimal perturbation policy using a framework based on stochastic kernels. Then, we use dynamic programming, and find a Nash equilibrium of the game. In particular, we show that the optimal closed-loop control policy is a certainty-equivalence policy and the optimal closed-loop perturbation policy is an affine state-feedback policy with a Gaussian white noise whose covariance at each time depends on the estimation error of the system.

A refined algorithm for curve fitting by segmented straight lines

ABSTRACT. We consider least squares approximation of a function of one variable by a continuous, piecewise-linear approximand that has a small number of breakpoints. This problem was notably considered by Bellman who proposed an approximate algorithm based on dynamic programming. Many suboptimal approaches have been suggested, but so far, the only exact methods resort to mixed integer programming with superpolynomial complexity growth.

In this paper, we present an exact and efficient algorithm based on dynamic programming with a hybrid value function. The achieved time-complexity seems to be polynomial.

Self-triggered control under actuator delays

ABSTRACT. In this paper we address the problem of self-triggered control of nonlinear systems under actuator delays. In particular, for globally asymptotically stabilizable systems we exploit the Lipschitz properties of the system’s dynamics, and present a self-triggered strategy that guarantees the stability of the sampled closed-loop system with bounded actuator delays.

Keeping it Real...Not!

ABSTRACT. Complex-valued signals and systems are ubiquitous in signal processing and telecommunications. However, there are a number of important control applications where a complex-valued system representation greatly facilitates analysis and control design. Two of these have been known for some time: regulation of electric machines, and active vibration damping of rotating machinery.

More recently, we showed that a complex-valued representation simplifies the control of electromagnetic fields in particle accelerators. In this contribution we will touch upon basics of complex-coefficient systems, and exemplify how they have helped us in our work on controlling electromagnetic fields.

Economic optimization of paper machines using EMPC

ABSTRACT. In recent years, modern control algorithms have started to move away from classic set-point-based objectives towards directly optimizing targets better aligned with overall process objectives such as the economic performance. One method which has attracted significant interest in recent years is economic model pre- dictive control (EMPC) which we here consider for economic optimization of a paper machine. In EMPC, an economic model is combined with a dynamic process-model to predict the economic outcome for a given sequence of future control inputs. By optimizing over the possible future control actions — subject to constraints such as product quality, input saturations and rate limitations — an economically optimal input trajectory may be determined and applied to the controlled system. To account for disturbances and uncertainty, the opti- mization problem is solved iteratively in a receding horizon fashion where only the first part of the optimal trajectory is applied to the system before mea- surements are taken and the problem is re-solved, yielding an updated optimal input-trajectory. The results presented in this paper are based on a high-fidelity grey-box model identified from real process data. The simulations reveal that EMPC may improve several aspects of the economic performance of a paper machine. First, EMPC may automatically prioritize among an excessive number of inputs to determine which combinations of inputs to use in order to counter disturbances in the most economically efficient manner. Also, since EMPC makes use of dynamic optimization, it may utilize control inputs which traditionally has not been used for set-point tracking, e.g., inputs with zero steady-state gain. Sec- ond, since EMPC is predictive in nature, it may plan ahead for known process changes such as grade-changes. By preparing the process ahead of time for a grade-change, the transition may be made faster with a significant reduction in lost production, and thereby significant improvements in profitability, especially for machines where grade-changes are frequent. Interestingly, we note that the same controller may be used both for grade-changes and during fixed-grade pro- duction, hence simplifying the common control architecture in which different controllers are used for these operating modes. An important property of EMPC to note is that the controller often drives the process to an operating point where constraints are active. This is not specific for EMPC but rather inherent in essentially all forms of economic op- timization. Nevertheless, this may cause problems with robustness since even small exogenous disturbances or unmodelled dynamics may cause constraint vi- olations. We therefore suggest using an adaptive approach where a constraint margin is introduced in the EMPC optimization problem to ensure that the op- erating point is backed off from the actual constraints relevant for production. We conclude that EMPC has the potential to improve the economic per- formance of paper machines significantly by integrating economic optimization directly into the control architecture. In particular, the observed improvements of grade-changes and the ability to automatically make trade-offs between the use of different control signals shows promise with respect to the economic per- formance.

Fast convex optimization for eigenproblems and beyond

ABSTRACT. We introduce an efficient algorithm for finding the dominant generalized eigenvectors of a pair of symmetric matrices. Combining tools from approximation theory and convex optimization, we develop a simple scalable algorithm with strong theoretical performance guarantees. The algorithm retains the simplicity of the well-known power method but enjoys the asymptotic iteration complexity of the powerful Lanczos method. Unlike these classic techniques, our algorithm is designed to decompose the overall problem into a series of subproblems that only need to be solved approximately. The combination of good initializations, fast iterative solvers, and appropriate error control in solving the subproblems lead to a linear running time in the input sizes compared to the superlinear time for the traditional methods.

The improved running time immediately offers acceleration for several applications. We apply the proposed algorithm to canonical correlation analysis, which is a fundamental statistical tool for learning of a low dimensional representation of high dimensional objects, and demonstrate improved performance over current state-of-the art algorithms.

Anomaly detector metrics for sensor data attacks in control systems

ABSTRACT. We investigate worst-case impacts of stealthy full sensor attacks against linear control systems under a general class of anomaly detectors. We show that the worst-case impact scales with a parameter that is solely determined by the number of sensors in the plant, the detector used, and its tuning. Therefore, we obtain a general metric to compare the performance of detectors of this class of detectors under the investigated sensor attack, which is independent of the plant dynamics and applies to all linear systems with the same number of sensors.

Server state estimation using event-based particle filtering

ABSTRACT. Closed-loop control of cloud resources requires there to be measurements readily available from the process in order to utilize the feedback mechanism to form a control law. Sought states might be unfeasible or impossible to measure in real applications; instead they must be estimated. Running the estimators in real time for all measurements however requires a lot of computational overhead. We propose an event-based particle filter approach to capture the internal dynamics of a server with CPU-intensive workload. Preliminary results show some promise as it outperforms analytic estimators derived from stationary equations in service rate estimation for a simulated example. Further we show that for the same example, an event-based sampling strategy outperforms periodic sampling.

Flexible models for smart maintenance

ABSTRACT. Smart maintenance utilizes empirical data and predicts the machine failure. This technique can be found built into modern machines and industrial plants, but maintenance in existing plants is becoming increasingly important; from having resolved failures as preventive, maintenance moves towards to a predictive strategy. Recent advances in manufacturing industry has paved way for a systematical deployment of Cyber-Physical Systems (CPS), within which information from all related perspectives is closely monitored and synchronized between the physical factory floor and the cyber computational space. Within this project, models of physical objects (so-called cyber twins) have been designed and evaluated. Specifically, a recently published system identification method, LAVA, has been tested in two different industrial case studies: (i) heat exchangers at SSAB in Borlänge and (ii) tool condition monitoring at Svenska Fönster AB in Edsbyn. The objective has been to find a general method for improving the predictive maintenance capabilities of different industrial plants, and the initial resulats are promissing.

Using valued Booleans to find simpler counterexamples in random testing of cyber-physical Systems

ABSTRACT. We propose a new logic of valued Booleans for writing properties which are not just true or false but compute how severely they are falsied. The logic is reminiscent of STL or MTL but gives the tester control over what severity means in the particular problem domain. We use this logic to simplify failing test inputs in the context of random testing of cyber-physical systems and show that it improves the quality of counterexamples found. The logic of valued Booleans might also be used as an alternative to the standard robust semantics of STL formulas in optimization-based approaches to falsication.

Distributed cooperative model predictive control for on-road autonomous driving

ABSTRACT. This paper describes a cooperative control method (DCMPC) for autonomous vehicles, in order to perform different traffic maneuvers. The problem is formulated as a distributed optimal control problem for a system of multiple autonomous vehicles and then solved using nonlinear Model Predictive Control (MPC). It is proved that by applying the proposed method to a system of multiple autonomous vehicle, each vehicle will converge to a neighborhood of its desired destination. Simulation results show that the distributed algorithm scales well with increasing number of vehicles.

14:00-14:45 Session 8: Plenary 3: Long-range autonomy and constraint-based coordination of Multi-Robot systems

Magnus Egerstedt

Executive Director for the Institute for Robotics and Intelligent Machines at the Georgia Institute of Technology and a Professor and the Julian T. Hightower Chair in Systems and Controls in the School of Electrical and Computer Engineering

Abstract: By now, we have a fairly good understanding of how to design coordinated control strategies for making teams of mobile robots achieve geometric objectives in a distributed manner, such as assembling shapes or covering areas. But, the mapping from high-level tasks to these objectives is not particularly well understood. In this talk, we investigate this topic in the context of long-range autonomy, i.e., we consider teams of robots, deployed in an environment over a sustained period of time, that can be recruited to perform a number of different tasks in a distributed, safe, and provably correct manner. This development will involve the composition of multiple barrier certificates for encoding the tasks and safety constraints, as well as a detour into ecology as a way of understanding how persistent environmental monitoring, as a special instantiation of the long-range autonomy concept, can be achieved by studying animals with low-energy life-styles, such as the three-toed sloth.

Location: Q1