ArticlePDF Available

Genetische Programmierung einer algorithmischen Chemie

Authors:

Abstract and Figures

Der genetischen Programmierung (GP) liegt zumeist die Annahme zugrunde, dass die Individuen eine evolvierte, wohldefinierte Struktur haben und ihre Ausführung deterministisch erfolgt. Diese Annahme hat ihren Ursprung nicht beim methodischen Vorbild, der natürlichen Evolution, sondern ist ein bewusstes oder unbewusstes Erbe der Umgebung, in der die Evolution nachgebildet wird - der von-Neumann-Architektur. John von Neumann hat mit der nach ihm benannten von-Neumann-Architektur weit mehr in der Informatik beeinflusst als das Gebiet der Rechnerarchitekturen. Daher ist sein Einfluss auf die Evolution von Algorithmen mittels genetischer Programmierung nicht verwunderlich, auch wenn die von-Neumann-Architektur wenig gemein mit den in der Natur evolvierten Systemen hat. In den letzten Jahren entstanden eine ganze Reihe von Konzepten und theoretischen Modellen, die nur noch wenig Anleihen bei von Neumanns Rechnerarchitektur machen und die in ihren Eigenschaften stärker natürlichen Systemen ähneln. Die Fähigkeit dieser Systeme, Berechnungen durchzuführen, entsteht erst durch die Interaktion ihrer parallel agierenden, nichtdeterministischen und dezentral organisierten Komponenten. Die Fähigkeit emergiert. Über die Evolution von Algorithmen für solche Systeme jenseits der von-Neumann-Architektur weiß man noch vergleichsweise wenig. Die vorliegende Arbeit nimmt sich dieser Fragestellung an und bedient sich hierbei der algorithmischen Chemie, einer künstlichen Chemie, die bei vereinfachter Betrachtungsweise aus einem veränderten Programmzeigerverhalten in der von-Neumann-Architektur resultiert. Reaktionen, eine Variante einfacher Instruktionen, werden hierbei in zufälliger Reihenfolge gezogen und ausgeführt. Sie interagieren miteinander, indem sie Produkte anderer Reaktionen verwenden und das Ergebnis ihrer Transformation, gespeichert in sogenannten Molekülen, anderen Reaktionen zur Verfügung stellen. Zur experimentellen Auswertung dieses nichtdeterministischen Systems wird die sequenzielle Parameteroptimierung um ein Verfahren zur Verteilung eines Experimentbudgets erweitert. Das systematische Design der Experimente und ihre anschließende Analyse ermöglichen es, generalisierte Erkenntnisse über das Systemverhalten jenseits konkreter Parametrisierungen zu gewinnen. Im Fall der genetischen Programmierung einer algorithmischen Chemie führen die gewonnenen Erkenntnisse zu einer Neuentwicklung des Rekombinationsoperators nach dem Vorbild homologer Rekombinationsoperationen und damit zu einer weiteren Verbesserung der Systemperformance. Es zeigt sich, dass die für ein zielgerichtetes Verhalten einer algorithmischen Chemie notwendigen Reaktionsschemata mittels genetischer Programmierung erlernt werden können. Für gängige Problemstellungen aus dem Bereich der genetischen Programmierung werden Lösungen gefunden, die in ihrer Güte mit denen anderer GP-Varianten und maschineller Lernverfahren vergleichbar sind. Die evolvierten Lösungen fallen dabei deutlich kompakter bezüglich der Datenflusshöhe und der Anzahl benötigter Operationen aus, als in dem zum Vergleich herangezogenen linearen GP-System.
Content may be subject to copyright.
A preview of the PDF is not available
... SPOT integrates sharpening as a simple method, which guarantees a fair comparison of the obtain solutions. Lasarczyk [27] and Bartz-Beielstein et al. [3,4] analyzed the integration of a more sophisticated control-theoretic simulation technique called optimal computing budget allocation (OCBA) into SPOT. The OCBA approach can intelligently determine the most efficient replication numbers [12]. ...
... Chen and Lee present a comprehensive coverage of the OCBA methodology [12]. Lasarczyk was the first who combined SPOT and OCBA [27]. The OCBA implementation in this study is based on Lasarczyk's work. ...
... In Sutcliffe, G. C. J. and Goebel, R. G., editors, Proceedings of the 19th International Florida AI Research Society Conference (FLAIRS 2006), pages 451-456, Menlo Park, CA. AIII Press Lasarczyk [2007] successfully employed SPO to analyze and tune parameters of an algorithmic chemistry. He also tuned the parameters of a single-objective GP system using register machine programs as genotype representation (linear GP) to provide a baseline for his results. ...
... Lasarczyk, C. W. G. (2007). Genetische Programmierung einer algorithmischen Chemie. ...
Article
Genetic Programming (GP) is an evolutionary algorithm for the automatic discovery of symbolic expressions, e.g. computer programs or mathematical formulae, that encode solutions to a user-defined task. Recent advances in GP systems and computer performance made it possible to successfully apply this algorithm to real-world applications. This work offers three main contributions to the state-of-the art in GP systems: (I) The documentation of RGP, a state-of-the art GP software implemented as an extension package to the popular R environment for statistical computation and graphics. GP and RPG are introduced both formally and with a series of tutorial examples. As R itself, RGP is available under an open source license. (II) A comprehensive empirical analysis of modern GP heuristics based on the methodology of Sequential Parameter Optimization. The effects and interactions of the most important GP algorithm parameters are analyzed and recommendations for good parameter settings are given. (III) Two extensive case studies based on real-world industrial applications. The first application involves process control models in steel production, while the second is about meta-model-based optimization of cyclone dust separators. A comparison with traditional and modern regression methods reveals that GP offers equal or superior performance in both applications, with the additional benefit of understandable and easy to deploy models. Main motivation of this work is the advancement of GP in real-world application areas. The focus lies on a subset of application areas that are known to be practical for GP, first of all symbolic regression and classification. It has been written with practitioners from academia and industry in mind.
... SPOT integrates sharpening as a simple method, which guarantees a fair comparison of the obtain solutions. Lasarczyk [27] and Bartz-Beielstein et al. [3,4] analyzed the integration of a more sophisticated controltheoretic simulation technique called optimal computing budget allocation (OCBA) into SPOT. The OCBA approach can intelligently determine the most e cient replication numbers [12]. ...
... Chen and Lee present a comprehensive coverage of the OCBA methodology [12]. Lasarczyk was the first who combined SPOT and OCBA [27]. The OCBA implementation in this study is based on Lasarczyk's work. ...
Article
Full-text available
Sequential Parameter Optimization (SPO) is a meta-model based search heuristic that combines classical and modern statistical techniques. It was originally developed for the analysis of search heuristics such as simulated annealing, particle swarm optimization and evolutionary algorithms [6]. Here, SPO itself will be used as a search heuristic, i.e., SPO is applied to the objective function directly. An introduction to the state-of-the-art R implementation of SPO, the so-called sequential parameter optimization toolbox (SPOT), is presented in [5].
... As the model is not updated in this second stage, the quality of the best parameter vector found heavily depends on the correctness of the model in the first stage. Unlike the two-stage procedure of Coy, Sequential Parameter Optimization (SPO) [36, 45] performs a true multi-stage procedure where the model is constantly updated. Each iteration starts with generating a set of new vectors and predicting their utility using the model. ...
... Coy's Procedure [44] + - Sequential Parameter Optimization (SPO) [36] ++ -++ ++ + + SPO + OCBA [45] ++ -++ ++ + + ...
Article
Full-text available
In this paper we present a conceptual framework for parameter tuning, provide a survey of tuning methods, and discuss related methodological issues. The framework is based on a three-tier hierarchy of a problem, an evolutionary algorithm (EA), and a tuner. Furthermore, we distinguish problem instances, parameters, and EA performance measures as major factors, and discuss how tuning can be directed to algorithm performance and/or robustness. For the survey part we establish different taxonomies to categorize tuning methods and review existing work. Finally, we elaborate on how tuning can improve methodology by facilitating well-funded experimental comparisons and algorithm analysis.
... Furthermore, their study analyzed the interaction between global and local search in sequential tuning procedures. Lasarczyk [88] applied Chen's optimal computing budget allocation (OCBA) [89] to the SPO Toolbox (SPOT). The SPOT package is implemented in the R programming language and it can be achieved on the website [90]. ...
Article
Full-text available
Metaheuristic and heuristic methods have many tunable parameters, and choosing their values can increase their ability to deal with hard optimization problems. Automated approaches for finding good parameter settings have attracted significant research and development efforts in the last few years. Because parameter tuning became commonly utilized in industry and research and there is a significant advancement in this area, a comprehensive review is an important requirement. Although there is very wide literature about algorithm configuration problems, a detailed survey analysis has not been conducted yet to the best of our knowledge. In this paper, we will briefly explain the automatic algorithm configuration problem and then survey the automated methods developed to handle this problem. After explaining the logic of these methods, we also argued about their main advantages and disadvantages to help researchers or practitioners select the best possible method for their specific problem. Moreover, some recommendations and possible future directions for this topic are provided as a conclusion.
... zation process, leading to a final choice of parameter settings that had only been evaluated using very few ( " lucky " ) target algorithm runs and that performed poorly in 5 Another approach for allocating an appropriate number of target algorithm runs to each parameter setting is Chen et al (2003)'s optimal computational budget allocation (OCBA). Lasarczyk (2007) implemented OCBA as an explicit intensification method to improve SPO's selection procedure, especially in high-noise scenarios. This implementation was done in R (Ihaka and Gentleman, 1996), and forthcoming versions of SPOT, which will also be based on R, will include OCBA. ...
Chapter
Full-text available
This work experimentally investigates model-based approaches for optimizing the performance of parameterized randomized algorithms. Such approaches build a response surface model and use this model for finding good parameter settings of the given algorithm. We evaluated two methods from the literature that are based on Gaussian process models: sequential parameter optimization (SPO) (Bartz-Beielstein et al. 2005) and sequential Kriging optimization (SKO) (Huang et al. 2006). SPO performed better "out-of-the-box," whereas SKO was competitive when response values were log transformed. We then investigated key design decisions within the SPO paradigm, characterizing the performance consequences of each. Based on these findings, we propose a new version of SPO, dubbed SPO+, which extends SPO with a novel intensification procedure and a log-transformed objective function. In a domain for which performance results for other (modelfree) parameter optimization approaches are available, we demonstrate that SPO+ achieves state-of-the-art performance. Finally, we compare this automated parameter tuning approach to an interactive, manual process that makes use of classical
... Sequential approaches are generally more efficient, i.e., require fewer function evaluations, than approaches that evaluate the information in one step only. Extensions of this sequential framework are discussed in Bartz-Beielstein et al. [3] and Lasarczyk [14]. We used SPO to tune parameters of NARX to minimize the RMSE on the test dataset. ...
... In [7], random permutation tests have been employed to decide if a newly tested configuration shall be repeated as often as the current best one or if it can safely be regarded as inferior. Lasarczyk [21] implemented Chen's optimal computing budget allocation (OCBA) [12] into SPOT. ...
Chapter
Full-text available
In this chapter we discuss the notion of Evolutionary Algorithm (EAs) parameters and propose a distinction between EAs and EA instances, based on the type of parameters used to specify their details. Furthermore, we consider the most important aspects of the parameter tuning problem and give an overview of existing parameter tuning methods. Finally, we elaborate on the methodological issues involved here and provide recommendations for further development. © 2012 Springer-Verlag Berlin Heidelberg. All rights are reserved.
Conference Paper
Full-text available
We provide a comprehensive, effective and very efficient methodology for the design and experimental analysis of algorithms. We rely on modern statistical techniques for tuning and understanding algorithms from an experimental perspective. Therefore, we make use of the sequential parameter optimization (SPO) method that has been successfully applied as a tuning procedure to numerous heuristics for practical and theoretical optimization problems. Two case studies, which illustrate the applicability of SPO to algorithm tuning and model selection, are presented. @InProceedings{bartzbeielstein:DSP:2009:2115, author = {Thomas Bartz-Beielstein}, title = {Sequential Parameter Optimization}, booktitle = {Sampling-based Optimization in the Presence of Uncertainty }, year = {2009}, editor = {J{"u}rgen Branke and Barry L. Nelson and Warren Buckler Powell and Thomas J. Santner}, number = {09181}, series = {Dagstuhl Seminar Proceedings}, ISSN = {1862-4405}, publisher = {Schloss Dagstuhl - Leibniz-Zentrum fuer Informatik, Germany}, address = {Dagstuhl, Germany}, URL = {http://drops.dagstuhl.de/opus/volltexte/2009/2115}, annote = {Keywords: Optimization, evolutionary algorithms, design of experiments} }
Conference Paper
A methodology is developed for automatically tuning the main parameters of model predictive control (MPC) such as prediction horizon, control horizon and control interval. The tuning of parameters is done by means of sequential parameter optimization. In the process of optimization one of the major issues is the choice of an objective function. Several types of objective functions are tested in order to choose the one which solves the MPC tuning problem most adequate. In addition, different scenarios are analyzed if an exact model of the true plant does not exist.
Article
Full-text available
Many scientific phenomena are now investigated by complex computer models or codes. A computer experiment is a number of runs of the code with various inputs. A feature of many computer experiments is that the output is deterministic—rerunning the code with the same inputs gives identical observations. Often, the codes are computationally expensive to run, and a common objective of an experiment is to fit a cheaper predictor of the output to the data. Our approach is to model the deterministic output as the realization of a stochastic process, thereby providing a statistical basis for designing experiments (choosing the inputs) for efficient prediction. With this model, estimates of uncertainty of predictions are also available. Recent work in this area is reviewed, a number of applications are discussed, and we demonstrate our methodology with an example.
Conference Paper
Full-text available
Article
Two types of sampling plans are examined as alternatives to simple random sampling in Monte Carlo studies. These plans are shown to be improvements over simple random sampling with respect to variance for a class of estimators which includes the sample mean and the empirical distribution function.
Chapter
While crossover is generally accepted as an explorative operator in string based G.A.s (Goldberg, 1989), the benefit or otherwise of employing crossover in tree based Genetic Programming is often disputed. Work such as (Collins, 1992) went as far as to dismiss GP as a biological search method due to its use of trees, while (Angeline, 1997) presented results which suggested that crossover in GP can provide little benefit over randomly generating subtrees.
Article
The emergence of order in systems with many actors or agents is an interesting problem for sociology as well as for computer science. Starting the from sociological theory of the dyadic "situation of double contingency", our main focus is on large actor populations and their capability to produce order depending on different actors' constellations. Based on the theory for dyadic actor constellations we present our model of the actor. We do not want the actors to identify one another, so we do not need to modify this model if we scale up population size next and introduce constellations. Thereby we take regular, random and small-world constellations into account. After describing our measures of order we study emergence of order in different constellations for varying population sizes. By means of simulation experiments we show that systems with small-worlds exhibit highest order on large populations which gently decreases on increasing population sizes.
Article
Membrane Computing (MC) is part of the powerful trend in computer science known under the name of Natural Computing. Its goal is to abstract computing models from the structure and the functioning of the living cell. The present paper is a short and informal introduction to MC, presenting the basic ideas, the central (types of) results, and the main directions of research.
Article
Recently, biochemical systems have been shown to possess interesting computational properties. In a parallel development, the chemical computation metaphor is becoming more and more frequently used as part of the emergent computation paradigm in computer science. We review in this contribution the idea behind the chemical computational metaphor and outline its relevance for nanotechnology. We set up a simulated reaction system of mathematical objects and examine its dynamics by computer experiments. Typical problems of computer science, such as sorting, parity checking or prime number computation are placed within this context. The implications of this approach for nanotechnology, parallel computers based on molecular devices and DNA-RNA-protein information processing are discussed.