There are several methods that are frequently used for solving data based system identification problems; genetic programming (GP) has already been used successfully for solving data mining problems in the context of several scientific domains. Extended functional bases, additional optimization phases and further developed selection mechanisms essentially contribute to the method's ability to generate high quality results for various kinds of data based identification scenarios. Even though there has already been a lot of investigation regarding the optimization of the method and its parameter settings, there is still rather little systematic analysis of internal processes regarding genetic dynamics and the progress of genetic diversity during the execution of genetic programming based identification using these algorithmic extensions. In this paper, we report on results of investigations regarding exactly these aspects: We have developed methods and statistical features that are able to describe genetic diversity and dynamics of GP-based structure identification algorithms; here, we introduce statistic analysis of genetic diversity regarding variables and time offset settings within GP populations. Genetic diversity is (amongst other aspects) characterized by the occurrence of variables for the models in which they are used; statistical methods for estimating respective impact features are also presented here. Data sets representing two different kinds of systems (complex mechatronical systems as well as medical benchmark data) have been used for empirical tests; furthermore, standard implementations of genetic programming are compared to extended techniques including offspring selection as well as sliding window techniques.
Identifying nonlinear model structures as a part of analyzing a physical system means trying to generate an algebraic expression as a part of an equation that describes the physical representation of a dynamic system. Many existing system identification methods are based on parameter identification. In this paper, we describe a method using genetic programming to evolve an algebraic representation of measured input-output response data. The main advantage of the presented approach is that unlike many other identification methods, it does not restrict the set of models that can be identified but can be applied to any kind of data sets representing a system's observed or simulated input and output signals. This paper describes research that was done for the project "Specification, Design and Implementation of a Genetic Programming Approach for Identifying Nonlinear Models of Mechatronic Systems". The goal of the project is to find models for mechatronic systems; our task was to examine whether the methods of Genetic Programming are suitable for determining the structures of physical systems by analyzing a system's measured behaviour or not.
This contribution considers recent results of population genetics in order to present generic extensions to the general concept of a Genetic Algorithm (GA). Consequently a new model for self-adaptive selection pressure steering is presented (Offspring Selection), taking advantage of the interplay between directed genetic drift and selection, resulting in a new class of Genetic Algorithms. As a result, we introduce and empirically analyze the generic extensions to the general GA concept, which make genetic search more stable in terms of operators, and allows steering and scaling up of global solution quality to the highest quality regions without using problem specific information or local searches.
Heuristic optimization techniques turned out to be very well suited for attacking various kinds of problems. However, when it comes to practical applications like scheduling problems, route planning, etc., also these algorithms still suffer from a very long running time mainly due to the rather large problem instances relevant in real world applications. Consequently, parallel optimization methods like parallel Genetic Algorithms are widely used to overcome this handicap. In this paper, the authors present a new environment for parallel heuristic optimization based upon the already proposed HeuristicLab. In contrast to other existing grid computing or parallel optimization projects, HeuristicLab Grid offers the possibility of rapid and easy use of existing optimization algorithms and problems in a parallel way without the need of complex installation and maintenance.
JavaScript jest wyłączony w Twojej przeglądarce internetowej. Włącz go, a następnie odśwież stronę, aby móc w pełni z niej korzystać.