Search

This chapter formulates seven lessons for preventing harm in artificial intelligence (AI) systems based on insights from the field of system safety for software-based automation in safety-critical domains. New applications of AI across societal domains and public organizations and infrastructures come with new hazards, which lead to new forms of harm, both grave and pernicious. The chapter addresses the lack of consensus for diagnosing and eliminating new AI system hazards. For decades, the field of system safety has dealt with accidents and harm in safety-critical systems governed by varying degrees of software-based automation and decision-making. This field embraces the core assumption of systems and control that AI systems cannot be safeguarded by technical design choices on the model or algorithm alone, instead requiring an end-to-end hazard analysis and design frame that includes the context of use, impacted stakeholders, and the formal and informal institutional environment in which the system operates. Safety and other values are then inherently socio-technical and emergent system properties that require design and control measures to instantiate these across the technical, social, and institutional components of a system. This chapter honors system safety pioneer Nancy Leveson, by situating her core lessons for today’s AI system safety challenges. For every lesson, concrete tools are offered for rethinking and reorganizing the safety management of AI systems, both in design and governance. This history tells us that effective AI safety management requires transdisciplinary approaches and a shared language that allows involvement of all levels of society.

Biological systems display a variety of complex dynamic behaviors, ranging from periodic orbits to chaos. Regular rhythmic behavior, for instance, is associated with locomotion, while chaotic behavior is observed in neural interactions. Both these cases can be mathematically expressed as the interaction of a collection of coupled bodies or oscillators that are actuated to behave with a desired pattern. In animal locomotion, this desired pattern is the periodic body motion (gait) that interacts with the environment to generate thrust for motion. By contrast, the observed behavior of a network of neurons is possibly chaotic and flexible. This research focuses on the design and analysis of these two types of behaviors in biologically-inspired systems. A fundamental problem in animal locomotion is determining a gait that optimizes an essential performance while satisfying a desired velocity constraint. In this study, a functional model is developed for a general class of three dimensional locomotors with full (six) degrees of freedom, in addition to arbitrary finite degrees of freedom for body shape deformation. An optimal turning gait problem is then formulated for a periodic body movement that minimizes a quadratic cost function while achieving a steady turning motion with prescribed average linear and angular velocities. The problem is shown to reduce equivalently to two separate and simpler minimization problems that are both solvable for globally optimal solutions. Optimal gait theory can also be utilized in order to determine analytical justifications for observed behavior in biological systems. In this study, a simple body-fluid fish model is developed, and steady swimming at various speeds is analyzed using optimal gait theory. The results show that the gait that minimizes bending moment over tail movements and stiffness matches data from observed swimming of saithe. Furthermore, muscle tension is reduced when undulation frequency matches the resonance frequency, which maximizes the ratio of tail-tip velocity to bending moment. The final task is to design the interconnections in a network of Andronov-Hopf oscillators in order to generate desired chaotic behavior. Due to the structure of the oscillators, it is possible to generate chaos by using weak linear coupling to destabilize the phase difference between the oscillators. To this end, a set of sufficient conditions are determined to guarantee the instability of a desired periodic solution through phase destabilization. Subsequently, a condition is found to guarantee the absence of any stable harmonic orbit. Finally, additional properties are considered, where small variations in a parameter can lead to chaotic behavior. With additional research, these results can be expanded to the design of a chaotic neural controller to generate adaptive locomotion for a mechanical rectifier.

The complexity of systems that are relevant to engineering today has grown tremendously. The control techniques based on frequency analysis that were perfectly adequate for simple systems tend to be difficult to use for more complex systems. An important challenge arising for these complex systems is the need to obtain sets satisfying given properties. Once the intended end-use of the sets as well as the properties it should satisfy are clarified, a specific family of sets, called template, is chosen to formulate the search of the appropriate set as a problem that is numerically tractable. In this thesis, we introduce Set Programming as an interface between the requirements that sets should satisfy and the numerical algorithms used to compute sets satisfying given requirements. Several templates are studied in this thesis including the classical templates (polyhedra, zonotopes and ellipsoids) but also more elaborate ones that provide a richer family of sets but are more complicated to implement both theoretically and algorithmically. These includes polysets, piecewise semi-ellipsoids, and piecewise polysets. We study two questions in detail about set programs that both examine different aspects of duality: 1) Conic duality: we analyse what the infeasibility of a set program for a specific template means for the set programs for other templates, e.g., is it infeasible only for this template or for all of them ? 2) Geometric duality: we discuss the choices of represention of convex sets, either in the primal or dual space, depending on the class of set programs that should be solved. Interestingly, this choice seems to be mostly template-independent as it mostly relies on geometric arguments. Finally, we apply our results to several applications ranging from cruise control to energy and information theory. (FSA - Sciences de l'ingénieur) -- UCL, 2020

In tokamak experimental reactors, the magnetic control system is one of the main plasma control systems that is required, together with the density control, since the very beginning, even before first operations. Indeed, the magnetic control drives the current in the external poloidal circuits in order to first achieve the breakdown conditions and, after plasma formation, to track the desired plasma current, shape and position. Furthermore, when the plasma poloidal cross-section is vertically elongated, the magnetic control takes also care of the vertical stabilization of the plasma column, and therefore it is an essential system for operation. This chapter introduces a reference architecture for plasma magnetic control in tokamaks. Given the proposed architecture, the techniques to design all the required control algorithms is also presented. Experimental results obtained on the JET and EAST tokamaks and simulations for machines currently under construction are shown to prove the effectiveness of the proposed architecture and control algorithms.

Smart defense systems using missiles that can fine-tune their velocity profiles have significant technological superiority over their conventional counterparts. This tuning is possible, in part, due to the deployment of advanced sensing, actuation, and computation capabilities and sophisticated guidance, navigation, and control algorithms. The capability to alter velocity during operation helps sustain optimum performance for different flight conditions. In addition, it makes it possible to slow down while turning and then speed up along a straight path, rendering the maneuvers more efficient. This ability to modify velocity (known as throttleability) is also known to increase a missile's no-escape zone, which is the maximum range that the missile can outrun its target [1]. As presented in "Summary," this article discusses the advanced control technologies needed to obtain throttleability.

Tensor decompositions are becoming increasingly popular in various scientific fields, including signal processing, higher-order statistics, and chemometrics. This is due to their useful properties, namely, uniqueness and thus interpretability of the factors, ability to work with high-dimensional data directly, and ability to work with multiple data sets simultaneously (data fusion), among others. Tensor methods are slowly entering the systems and control world but their potential has not yet been fully utilized. We present an overview of re-occuring problems in systems and control where the powerful tensor techniques can be applied successfully, with the hope to increase their impact on the community.

Extreme UltraViolet (EUV) lithography is a new technology for production of integrated circuits. In EUV lithographic machines, optical elements are heated by absorption of exposure energy. Heating induces thermoelastic deformations of optical elements and consequently, it creates wavefront aberrations. These Thermally Induced Wavefront Aberrations (TIWA) can significantly degrade the resolution of EUV lithographic machines. One of the ways to correct TIWA and consequently, to improve the resolution of EUV lithographic machines, is to use the Adaptive Optics (AO) technique and predictive control algorithms. However, the predictive control of TIWA is a challenging problem, mainly because the dynamical behavior of TIWA is described by thermoelastic Partial Differential Equations (PDEs). By discretizing the thermoelastic equations using the finite difference or finite element methods, large-scale state-space models can be obtained. A striking feature of these state-space models is that they have sparse (multi) banded matrices. Consequently, these state-space models can be interpreted as large-scale networks of local subsystems. All this implies that the problem of correcting TIWA can be placed in a much more general context of identifying, estimating and controlling large-scale interconnected (distributed) systems. However, currently used estimation and control algorithms are not computationally feasible for large-scale systems. For this reason, this thesis focuses on the development of computationally efficient identification, estimation and control algorithms for large-scale interconnected systems. In this thesis, we prove that the inverses of (finite-time) Gramians of large-scale interconnected systems, obtained by discretizing PDEs, belong to a class of off-diagonally decaying matrices. Consequently, these inverses can be approximated by sparse (multi) banded matrices with O(N) complexity, where N is the number of local subsystems. To compute the approximate inverses, the Chebyshev approximation method and the Newton iteration are used. On the basis of these theoretical results, we develop novel estimation and identification algorithms for large-scale interconnected systems. The computational and memory complexity of these methods scale with O(N). This is the main advantage over the traditional estimation and identification methods, which complexity is O(N^3). Furthermore, in this thesis we develop and experimentally verify AO control algorithms for predictive correction of TIWA. The experimental results give a good first indication for a possible implementation of the developed predictive control algorithms in EUV lithographic machines. Finally, we develop and experimentally test an iterative learning control algorithm for high performance correction of wavefront aberrations.

Extreme UltraViolet (EUV) lithography is a new technology for production of integrated circuits. In EUV lithographic machines, optical elements are heated by absorption of exposure energy. Heating induces thermoelastic deformations of optical elements and consequently, it creates wavefront aberrations. These Thermally Induced Wavefront Aberrations (TIWA) can significantly degrade the resolution of EUV lithographic machines. One of the ways to correct TIWA and consequently, to improve the resolution of EUV lithographic machines, is to use the Adaptive Optics (AO) technique and predictive control algorithms. However, the predictive control of TIWA is a challenging problem, mainly because the dynamical behavior of TIWA is described by thermoelastic Partial Differential Equations (PDEs). By discretizing the thermoelastic equations using the finite difference or finite element methods, large-scale state-space models can be obtained. A striking feature of these state-space models is that they have sparse (multi) banded matrices. Consequently, these state-space models can be interpreted as large-scale networks of local subsystems. All this implies that the problem of correcting TIWA can be placed in a much more general context of identifying, estimating and controlling large-scale interconnected (distributed) systems. However, currently used estimation and control algorithms are not computationally feasible for large-scale systems. For this reason, this thesis focuses on the development of computationally efficient identification, estimation and control algorithms for large-scale interconnected systems. In this thesis, we prove that the inverses of (finite-time) Gramians of large-scale interconnected systems, obtained by discretizing PDEs, belong to a class of off-diagonally decaying matrices. Consequently, these inverses can be approximated by sparse (multi) banded matrices with O(N) complexity, where N is the number of local subsystems. To compute the approximate inverses, the Chebyshev approximation method and the Newton iteration are used. On the basis of these theoretical results, we develop novel estimation and identification algorithms for large-scale interconnected systems. The computational and memory complexity of these methods scale with O(N). This is the main advantage over the traditional estimation and identification methods, which complexity is O(N^3). Furthermore, in this thesis we develop and experimentally verify AO control algorithms for predictive correction of TIWA. The experimental results give a good first indication for a possible implementation of the developed predictive control algorithms in EUV lithographic machines. Finally, we develop and experimentally test an iterative learning control algorithm for high performance correction of wavefront aberrations.

This text forms part of material taught during a course in advanced reservoir simulation at Delft University of Technology over the past 10 years. The contents have also been presented at various short courses for industrial and academic researchers interested in background knowledge needed to perform research in the area of closed-loop reservoir management, also known as smart fields, related to e.g. model-based production optimization, data assimilation (or history matching), model reduction, or upscaling techniques. Each of these topics has connections to system-theoretical concepts. The introductory part of the course, i.e. the systems description of flow through porous media, forms the topic of this brief monograph. The main objective is to present the classic reservoir simulation equations in a notation that facilitates the use of concepts from the systems-and-control literature. Although the theory is limited to the relatively simple situation of horizontal two-phase (oil-water) flow, it covers several typical aspects of porous-media flow. The first chapter gives a brief review of the basic equations to represent single-phase and two-phase flow. It discusses the governing partial-differential equations, their physical interpretation, spatial discretization with finite differences, and the treatment of wells. It contains well-known theory and is primarily meant to form a basis for the next chapter where the equations will be reformulated in terms of systems-and-control notation. The second chapter develops representations in state-space notation of the porous-media flow equations. The systematic use of matrix partitioning to describe the different types of inputs leads to a description in terms of nonlinear ordinary-differential and algebraic equations with (state-dependent) system, input, output and direct-throughput matrices. Other topics include generalized state-space representations, linearization, elimination of prestart from escribed pressures, the tracing of stream lines, lift tables, computational aspects, and the derivation of an energy balance for porous-media flow. The third chapter first treats the analytical solution of linear systems of ordinary differential equations for single-phase flow. Next it moves on to the numerical solution of the two-phase flow equations, covering various aspects like implicit, explicit or mixed (IMPES) time discretizations and associated stability issues, Newton-Raphson iteration, streamline simulation, automatic time-stepping, and other computational aspects. The chapter concludes with simple numerical examples to illustrate these and other aspects such as mobility effects, well-constraint switching, time-stepping statistics, and system-energy accounting. The contents of this text should be of value to students and researchers interested in the application of systems-and-control concepts to oil and gas reservoir simulation and other applications of subsurface flow simulation such as CO2 storage, geothermal energy, or groundwater remediation.

In this thesis, we develop an investigation in the framework of optimization in systems and control, following a perspective starting from a very common approach (that of Linear Matrix Inequalities or LMIs), which limitations draw us to consider a seldom used approach (the derivative-free optimization methods). This investigation is motivated by the fact that this framework is currently largely dominated by LMIs and convex optimization approaches even when the underlying problems are non-convex (in their original or less approximated formulations), in which case other optimization alternatives than convexity-based methods may often be more adequate. This work has been built with the intent of keeping a general perspective, where the developments are illustrated with some specific problems. Main motivations for this thesis are the following: - investigate the current main trend for optimization in systems and control, - outline some limitations of convex optimization approaches to solve non-convex problems, - put forward why derivative-free methods should be much more considered in systems and control, - solve (hard) problems of optimization in systems and control (i.e. a useful, and sometimes optimal, solution is obtained in practice). A key interest of this line of investigation is that, while relying on fundamental and necessarily theoretical concepts from academic research, it pays particular attention to practical issues of implementation (i.e. the results could be transferred for use in an industrial context). (FSA 3) -- UCL, 2012

In these notes the formulation of models is aimed at obtaining a description of the dynamic behaviour of processes under transient conditions. This implies that we will formulate the equations of motion of the process variables that describe the evolution of the process as a function of time. Our models will formulate the process dynamics in a form as required for the understanding of process operations such as startup and shutdown, or for studying the transitions from one operating condition to another one as, e.g., required by grade changes in a production plant or by changes in the composition of the feedstock. Process dynamic models also are of great importance for providing control engineers with qualitative and quantitative descriptions of the transient behaviour of processes that are to be used in model based control system design.

In these notes the formulation of models is aimed at obtaining a description of the dynamic behaviour of processes under transient conditions. This implies that we will formulate the equations of motion of the process variables that describe the evolution of the process as a function of time. Our models will formulate the process dynamics in a form as required for the understanding of process operations such as startup and shutdown, or for studying the transitions from one operating condition to another one as, e.g., required by grade changes in a production plant or by changes in the composition of the feedstock. Process dynamic models also are of great importance for providing control engineers with qualitative and quantitative descriptions of the transient behaviour of processes that are to be used in model based control system design.

The coming years there is a need to increase production from petroleum reservoirs, and there is an enormous potential to do so by increasing the recovery factor. This is possible by making better use of recent technological developments, such as horizontal wells, downhole valves and sensors. However, actually making better use of these improved capabilities is difficult because of many open problems in reservoir management and production operations processes. Consequently, there is significant scope to increase the recovery factor of oil and gas fields by tailoring tools from the systems and control community to efficiently perform dynamic optimization of wells (e.g. number, locations) and their production settings (e.g. bottom-hole pressures, flow rates, valve settings) based on uncertain reservoir models, in the sense that they lead to good decisions while requiring limited time from the user. This thesis aims at developing these tools, and the main contributions are as follows. Many production setting optimization problems can be written as optimal control problems that are linear in the control. If the only constraints are upper and lower bounds on the control, these problems can be expected to have pure bang-bang optimal solutions. The adjoint method to derive gradients of a cost function with respect to production settings can be combined with robust optimization to efficiently compute settings that are robust against uncertainty in reservoir models. The gradients used in production setting optimization can be used to efficiently compute directions in which to iteratively improve upon an initial well configuration by surrounding the to-be-placed wells by pseudo wells (i.e. wells that operate at a negligible rate). The controllability and observability properties of single-phase flow reservoir model are analyzed. It is shown that pressures near wells in which we can control the flow rate or bottom-hole pressure are controllable, whereas pressures near wells in which we can measure the flow rate or bottom-hole pressure are observable. Finally, a new method of regularization in history matching is presented, based on this controllability and observability analysis.

The coming years there is a need to increase production from petroleum reservoirs, and there is an enormous potential to do so by increasing the recovery factor. This is possible by making better use of recent technological developments, such as horizontal wells, downhole valves and sensors. However, actually making better use of these improved capabilities is difficult because of many open problems in reservoir management and production operations processes. Consequently, there is significant scope to increase the recovery factor of oil and gas fields by tailoring tools from the systems and control community to efficiently perform dynamic optimization of wells (e.g. number, locations) and their production settings (e.g. bottom-hole pressures, flow rates, valve settings) based on uncertain reservoir models, in the sense that they lead to good decisions while requiring limited time from the user. This thesis aims at developing these tools, and the main contributions are as follows. Many production setting optimization problems can be written as optimal control problems that are linear in the control. If the only constraints are upper and lower bounds on the control, these problems can be expected to have pure bang-bang optimal solutions. The adjoint method to derive gradients of a cost function with respect to production settings can be combined with robust optimization to efficiently compute settings that are robust against uncertainty in reservoir models. The gradients used in production setting optimization can be used to efficiently compute directions in which to iteratively improve upon an initial well configuration by surrounding the to-be-placed wells by pseudo wells (i.e. wells that operate at a negligible rate). The controllability and observability properties of single-phase flow reservoir model are analyzed. It is shown that pressures near wells in which we can control the flow rate or bottom-hole pressure are controllable, whereas pressures near wells in which we can measure the flow rate or bottom-hole pressure are observable. Finally, a new method of regularization in history matching is presented, based on this controllability and observability analysis.