Michael Affenzeller

Offspring Selection: A New Self-Adaptive Selection Scheme for Genetic Algorithms

In terms of goal orientedness, selection is the driving force of Genetic Algorithms (GAs). In contrast to crossover and mutation, selection is completely generic, i.e. independent of the actually employed problem and its representation. GA-selection is usually implemented as selection for reproduction (parent selection). In this paper we propose a second selection step after reproduction which is also absolutely problem independent. This self-adaptive selection mechanism, which will be referred to as offspring selection, is closely related to the general selection model of population genetics. As the problem- and representation-specific implementation of reproduction in GAs (crossover) is often critical in terms of preservation of essential genetic information, offspring selection has proven to be very suited for improving the global solution quality and robustness concerning parameter settings and operators of GAs in various fields of applications. The experimental part of the paper discusses the potential of the new selection model exemplarily on the basis of standardized real-valued test functions in high dimensions.

A Comprehensive Survey on Fitness Landscape Analysis

In the past, the notion of fitness landscapes has found widespread adoption. Many different methods have been developed that provide a general and abstract framework applicable to any optimization problem. We formally define fitness landscapes, provide an in-depth look at basic properties and give detailed explanations and examples of existing fitness landscape analysis techniques. Moreover, several common test problems or model fitness landscapes that are frequently used to benchmark algorithms or analysis methods are examined and explained and previous results are consolidated and summarized. Finally, we point out current limitations and open problems pertaining to the subject of fitness landscape analysis.

HeuristicLab: A Generic and Extensible Optimization Environment

Today numerous variants of heuristic optimization algorithms are used to solve different kinds of optimization problems. This huge variety makes it very difficult to reuse already implemented algorithms or problems. In this paper the authors describe a generic, extensible, and paradigm-independent optimization environment that strongly abstracts the process of heuristic optimization. By providing a well organized and strictly separated class structure and by introducing a generic operator concept for the interaction between algorithms and problems, HeuristicLab makes it possible to reuse an algorithm implementation for the attacking of lots of different kinds of problems and vice versa. Consequently HeuristicLab is very well suited for rapid prototyping of new algorithms and is also useful for educational support due to its state-of-the-art user interface, its self-explanatory API and the use of modern programming concepts.

SASEGASA: A New Generic Parallel Evolutionary Algorithm for Achieving Highest Quality Results

This paper presents a new generic Evolutionary Algorithm (EA) for retarding the unwanted effects of premature convergence. This is accomplished by a combination of interacting generic methods. These generalizations of a Genetic Algorithm (GA) are inspired by population genetics and take advantage of the interactions between genetic drift and migration. In this regard a new selection scheme is introduced, which is designed to directedly control genetic drift within the population by advantageous self-adaptive selection pressure steering. Additionally this new selection model enables a quite intuitive heuristics to detect premature convergence. Based upon this newly postulated basic principle the new selection mechanism is combined with the already proposed Segregative Genetic Algorithm (SEGA), an advanced Genetic Algorithm (GA) that introduces parallelism mainly to improve global solution quality. As a whole, a new generic evolutionary algorithm (SASEGASA) is introduced. The performance of the algorithm is evaluated on a set of characteristic benchmark problems. Computational results show that the new method is capable of producing highest quality solutions without any problem-specific additions.

Architecture and Design of the HeuristicLab Optimization Environment

Many optimization problems cannot be solved by classical mathematical optimization techniques due to their complexity and the size of the solution space. In order to achieve solutions of high quality though, heuristic optimization algorithms are frequently used. These algorithms do not claim to find global optimal solutions, but offer a reasonable tradeoff between runtime and solution quality and are therefore especially suitable for practical applications. In the last decades the success of heuristic optimization techniques in many different problem domains encouraged the development of a broad variety of optimization paradigms which often use natural processes as a source of inspiration (as for example evolutionary algorithms, simulated annealing, or ant colony optimization). For the development and application of heuristic optimization algorithms in science and industry, mature, flexible and usable software systems are required. These systems have to support scientists in the development of new algorithms and should also enable users to apply different optimization methods on specific problems easily. The architecture and design of such heuristic optimization software systems impose many challenges on developers due to the diversity of algorithms and problems as well as the heterogeneous requirements of the different user groups. In this chapter the authors describe the architecture and design of their optimization environment HeuristicLab which aims to provide a comprehensive system for algorithm development, testing, analysis and generally the application of heuristic optimization methods on complex problems.

Using enhanced genetic programming techniques for evolving classifiers in the context of medical diagnosis

There are several data based methods in the field of artificial intelligence which are nowadays frequently used for analyzing classification problems in the context of medical applications. As we show in this paper, the application of enhanced evolutionary computation techniques to classification problems has the potential to evolve classifiers of even higher quality than those trained by standard machine learning methods. On the basis of five medical benchmark classification problems taken from the UCI repository as well as the Melanoma data set (prepared by members of the Department of Dermatology of the Medical University Vienna) we document that the enhanced genetic programming approach presented here is able to produce comparable or even better results than linear modeling methods, artificial neural networks, kNN classification, support vector machines and also various genetic programming approaches.

Effects of constant optimization by nonlinear least squares minimization in symbolic regression

In this publication a constant optimization approach for symbolic regression is introduced to separate the task of finding the correct model structure from the necessity to evolve the correct numerical constants. A gradient-based nonlinear least squares optimization algorithm, the Levenberg-Marquardt (LM) algorithm, is used for adjusting constant values in symbolic expression trees during their evolution. The LM algorithm depends on gradient information consisting of partial derivations of the trees, which are obtained by automatic differentiation. The presented constant optimization approach is tested on several benchmark problems and compared to a standard genetic programming algorithm to show its effectiveness. Although the constant optimization involves an overhead regarding the execution time, the achieved accuracy increases significantly as well as the ability of genetic programming to learn from provided data. As an example, the Pagie-1 problem could be solved in 37 out of 50 test runs, whereas without constant optimization it was solved in only 10 runs. Furthermore, different configurations of the constant optimization approach (number of iterations, probability of applying constant optimization) are evaluated and their impact is detailed in the results section.

Advanced Genetic Programming Based Machine Learning

A Genetic Programming based approach for solving classification problems is presented in this paper. Classification is understood as the act of placing an object into a set of categories, based on the object’s properties; classification algorithms are designed to learn a function which maps a vector of object features into one of several classes. This is done by analyzing a set of input-output examples (“training samples”) of the function. Here we present a method based on the theory of Genetic Algorithms and Genetic Programming that interprets classification problems as optimization problems: Each presented instance of the classification problem is interpreted as an instance of an optimization problem, and a solution is found by a heuristic optimization algorithm. The major new aspects presented in this paper are advanced algorithmic concepts as well as suitable genetic operators for this problem class (mainly the creation of new hypotheses by merging already existing ones and their detailed evaluation). The experimental part of the paper documents the results produced using new hybrid variants of Genetic Algorithms as well as investigated parameter settings. Graphical analysis is done using a novel multiclass classifier analysis concept based on the theory of Receiver Operating Characteristic curves.

NOx Virtual Sensor Based on Structure Identification and Global Optimization

On-line measurement of engine NOx emissions is the object of a substantial effort, as it would strongly improve the control of Cl engines. Many efforts have been directed towards hardware solutions, in particular to physical sensors, which have already reached a certain degree of maturity. In this paper, we are concerned with an alternative approach, a virtual sensor, which is essentially a software code able to estimate the correct value of an unmeasured variable, thus including in some sense an input/output model of the process. Most virtual sensors are either derived by fitting data to a generic structure (like an artificial neural network, ANN) or by physical principles. In both cases, the quality of the sensor tends to be poor outside the measured values. In this paper, we present a new approach: the data are screened for hidden analytical structures, combining structure identification and evolutionary algorithms, and these structures are then used to develop the sensor presented. While the computational time for the sensor design can be significant (e.g. 1 or more hours), the resulting formula is very compact and proves able to predict the behaviour of the system at other operating points. The method has been validated with NOx data from a production engine measured with a Horiba Mexa 7000. The approach is able to yield a good prediction behaviour over a whole cycle. The results are consistent with physical knowledge.

Using FE Calculations and Data-Based System Identification Techniques to Model the Nonlinear Behavior of PMSMs

This paper investigates the modeling of brushless permanent-magnet synchronous machines (PMSMs). The focus is on deriving an automatable process for obtaining dynamic motor models that take nonlinear effects, such as saturation, into account. The modeling is based on finite element (FE) simulations for different current vectors in the