Automatic Development and Adaptation of Concise Nonlinear Models for System Identification

Created by W.Langdon from gp-bibliography.bib Revision:1.3872

@PhdThesis{LaCava:thesis,
  author =       "William G. {La Cava}",
  title =        "Automatic Development and Adaptation of Concise
                 Nonlinear Models for System Identification",
  school =       "University of Massachusetts, Amherst",
  year =         "2016",
  month =        may,
  address =      "USA",
  keywords =     "genetic algorithms, genetic programming, Acoustics,
                 Dynamics, and Controls",
  URL =          "http://scholarworks.umass.edu/dissertations_2/731/",
  URL =          "http://scholarworks.umass.edu/cgi/viewcontent.cgi?article=1762&context=dissertations_2",
  size =         "218 pages",
  abstract =     "Mathematical descriptions of natural and man-made
                 processes are the bedrock of science, used by humans to
                 understand, estimate, predict and control the natural
                 and built world around them. The goal of system
                 identification is to enable the inference of
                 mathematical descriptions of the true behaviour and
                 dynamics of processes from their measured observations.
                 The crux of this task is the identification of the
                 dynamic model form (topology) in addition to its
                 parameters. Model structures must be concise to offer
                 insight to the user about the process in question. To
                 that end, this dissertation proposes three methods to
                 improve the ability of system identification to
                 identify succinct nonlinear model structures.

                 The first is a model structure adaptation method (MSAM)
                 that modifies first principles models to increase their
                 predictive ability while maintaining intelligibility.
                 Model structure identification is achieved by this
                 method despite the presence of parametric error through
                 a novel means of estimating the gradient of model
                 structure perturbations. I demonstrate MSAM's ability
                 to identify underlying nonlinear dynamic models
                 starting from linear models in the presence of
                 parametric uncertainty. The main contribution of this
                 method is the ability to adapt the structure of
                 existing models of processes such that they more
                 closely match the process observations.

                 The second method, known as epigenetic linear genetic
                 programming (ELGP), conducts symbolic regression
                 without a priori knowledge of the form of the model or
                 its parameters. ELGP incorporates a layer of genetic
                 regulation into genetic programming (GP) and adapts it
                 by local search to tune the resultant model structures
                 for accuracy and conciseness. The introduction of
                 epigenetics is made simple by the use of a stack-based
                 program representation. This method, tested on hundreds
                 of dynamics problems, demonstrates the ability of
                 epigenetic local search to improve GP by producing
                 simpler and more accurate models.

                 The third method relies on a multidimensional GP
                 approach (M4GP) for solving multiclass classification
                 problems. The proposed method uses stack-based GP to
                 conduct nonlinear feature transformations to optimize
                 the clustering of data according to their classes. In
                 comparison to several state-of-the-art methods, M4GP is
                 able to classify test data better on several real-world
                 problems. The main contribution of M4GP is its
                 demonstrated ability to combine the strengths of GP
                 (e.g. nonlinear feature transformations and feature
                 selection) with the strengths of distance-based
                 classification.

                 MSAM, ELGP and M4GP improve the identification of
                 succinct nonlinear model structures for continuous
                 dynamic processes with starting models, continuous
                 dynamic processes without starting models, and
                 multiclass dynamic processes without starting models,
                 respectively. A considerable portion of this
                 dissertation is devoted to the application of these
                 methods to these three classes of real-world dynamic
                 modelling problems. MSAM is applied to the
                 restructuring of controllers to improve the closed-loop
                 system response of nonlinear plants. ELGP is used to
                 identify the closed-loop dynamics of an industrial
                 scale wind turbine and to define a reduced-order model
                 of fluid-structure interaction. Lastly, M4GP is used to
                 identify a dynamic behavioral model of bald eagles from
                 collected data. The methods are analyzed alongside many
                 other state-of-the-art system identification methods in
                 the context of model accuracy and conciseness.",
  notes =        "Supervisor: Kourosh Danai",
}

Genetic Programming entries for William La Cava

Citations