Files in this item



application/pdfWENDT-DISSERTATION-2017.pdf (32MB)
(no description provided)PDF


Title:Optimal nonlinear control and estimation using global domain linearization
Author(s):Wendt, Luke Adam
Director of Research:Levinson, Stephen E.
Doctoral Committee Chair(s):Levinson, Stephen E.
Doctoral Committee Member(s):Hasegawa-Johnson, Mark A.; Makela, Jonathan J.; Dell, Gary S.; Rothganger, Fred H
Department / Program:Electrical & Computer Eng
Discipline:Electrical & Computer Engr
Degree Granting Institution:University of Illinois at Urbana-Champaign
Abstract:Alan Turing teaches that cognition is symbol processing. Norbert Wiener teaches that intelligence rests on feedback control. Thus, there are discrete symbols and continuous sensory-motor signals. Sensorimotor dynamics are well-represented by nonlinear differential equations. A possible construction of symbols could be based on equilibria. Language is a symbol system and is one of the highest expressions of cognition. Much of this comes from spatial reasoning, which requires embodied cognition. Spatial reasoning derives from motor function. This thesis introduces a novel generalized non-heuristic method of linearizing nonlinear differential equations over a finite domain. It is used to engineer optimal convergence to target sets, a general form of spatial reasoning. Ordinary differential equations are ubiquitous models in physics and engineering that describe a wide range of phenomena including electromechanical systems. This thesis considers ordinary differential equations expressed in state-space form. For a given initial state, these equations generate signals that are continuous in both time and state. The control engineering objective is to find input functions that steer these states to desired target sets using only the measured output of the system. The state-space domain containing the target set, along with its cost, can be thought of as a symbol for high-level planning. Consider a basis state equal to a vector of nonlinear basis functions, computed from the state, where the state is generated from a multivariate nonlinear dynamic system. The basis state derivative can be expressed as a linear dynamic system with an additional error term. This thesis describes radial basis functions that minimize the error over the entire state-space domain, where the basis state equals zero if and only if the state is in a desired target set. This gives an approximate linear-dynamic system, and if the basis state goes to zero, then the state goes to the target set. This form of linear approximation is global over the domain. Careful selection of the basis gives a fully generalizable relationship between linear stability and nonlinear stability. This form of linearization can be applied to optimal state feedback and state estimation problems. This thesis carefully introduces optimal state feedback control with an emphasis on optimal infinite horizon solutions to linear-dynamic systems that have quadratic cost. A thorough introduction is also given to the optimal output feedback of linear-dynamics systems. The detectable and stabilizable subspaces of a linear-dynamic system are expressed in a generalized closed form. After introducing optimal control for linear systems, this thesis explores adaptive control from several different perspectives including: tuning, system identification, and reinforcement learning. Each of these approaches can be characterized as an optimal nonlinear output feedback problem. In each case, generalized representations can be found using a single layer of appropriately chosen nonlinear basis functions with linear parameterization. The primary focus of this thesis is to select these basis functions, in a fully generalized way, so that they have linear-dynamics. When this can be achieved, infinite horizon state feedback and state estimation can be computed using well-known closed-form solutions. This thesis demonstrates how multivariate nonlinear dynamic systems defined on a finite domain can be approximated by computationally equivalent high-dimensional linear-dynamic systems using a generalized basis state. This basis state is computed with a single layer of biologically inspired radial basis functions. The method of linearization is described as "global domain linearization" because it holds over a specified domain, and therefore provides a global linear approximation with respect to that domain. Any optimal state estimation or state feedback is globally optimal over the domain of linearization. The tools of optimal linear control theory can be applied. In particular, control and estimation problems involving under-actuated under-measured nonlinear systems with generalized nonlinear reward can be solved with closed-form infinite horizon linear-quadratic control and estimation. The controllable, uncontrollable, stabilizable, observable, unobservable, and detectable subspaces can all be described in a meaningful generalized way. State estimation and state feedback can then be implemented in computationally efficient low-dimensional highly nonlinear form. Generalized optimal state estimation and state feedback for continuous-time continuous-state systems is necessary machinery for any high-level symbolic planning that might involve unstable electromechanical systems. Symbols naturally form in the presence of more than one target state. This could provide a natural method of language acquisition. Given a state, all symbolic domains that intersect the state would have equilibrium and cost. These intersections define the legal grammar of symbol transition. An engineer or agent can design these symbols for high-level planning. Generalized infinite horizon state feedback and state estimation can then be computed for the continuous system that each symbol represents using traditional linear tools with domain linearization.
Issue Date:2017-07-09
Rights Information:Copyright 2017 Luke A. Wendt
Date Available in IDEALS:2017-09-29
Date Deposited:2017-08

This item appears in the following Collection(s)

Item Statistics