Institut für Neuroinformatik [Publications


	Internal Reports 2005

Home

Publications

IRINI

2005

The Institute
•	News
•	People
•	Teaching
•	Graduate study
•	Study and thesis projects
•	Jobs
•	Directions
•	Search
•	FTP Server
•	Contact
Chairs
•	Prof. Dr. Laurenz Wiskott
•	Prof. Dr. Gregor Schöner
Research groups
•	Theory of embodied cognition Prof. Dr. Gregor Schöner
•	Theory of Neural Systems Prof. Dr. Laurenz Wiskott
•	Neural Plasticity Lab PD Dr. Hubert Dinse
•	Real-Time Optical Imaging Lab Dr. Dirk Jancke
•	Organic Computing Dr. Rolf Würtz
•	Optimization of adaptive systems Jun.-Prof. Dr. Christian Igel
•	Autonomous robotics Dr. Ioannis Iossifidis
•	Medical Image Processing Dr. Susanne Winter
•	Real-time computer vision Jan Salmen
•	Multi-sensory fusion Dr. Andrey Bogdanov

IRINI 2005-01
- Title: "Gradient-based Optimization of Kernel-Target Alignment for Sequence Kernels"
- Author:Britta Mersch, Nico Pfeifer, Tobias Glasmachers, Peter Meinicke, and Christian Igel
- Date:September 2005
- Keywords:support vector machines, translation initiation sites, oligo kernel, kernel target alignment
- Abstract: Oligo kernels for biological sequence analysis have a high discriminative power and yield classifiers that are easy to interpret and to visualize. We propose gradient-based optimization of the kernel-target alignment to adapt multiple hyperparameters of combined oligo kernels. Experimental results show a significant improvement in detection of bacterial gene starts.
  full paper : gzipped postscript, pdf
IRINI 2005-02
- Title: "Reducing the Number of Fitness Evaluations in Graph Genetic Programming Using a Canonical Graph Indexed Database"
- Author:Jens Niehaus, Christian Igel, and Wolfgang Banzhaf
- Date:September 2005
- Keywords:graph representation, genetic programming, graph isomorphism, neutrality, fitness database
- Abstract:We describe the genetic programming system GGP that operates on graphs. It is shown empirically that such a representation is also suitable for problems where the solutions are usually encoded by trees. When a search space consists of graphs, graph isomorphisms influence the dynamics of optimization algorithms. We review methods for dealing with isomorphic graphs and discuss them in the context of genetic programming (GP), where isomorphic topologies are one source of phenotypic neutrality. Fitness databases can be helpful to improve the scaling of GP for problems with expensive fitness evaluations. We argue that databases considering graph isomorphisms save a significant amount of evaluations, which is supported by experiments using the GGP system.
  full paper : gzipped postscript, pdf
IRINI 2005-03
- Title: "Maximum-Gain Working Set Selection for SVMs"
- Author: Tobias Glasmachers and Christian Igel
- Date:September 2005
- Keywords:working set selection, sequential minimal optimization, quadratic programming, support vector machines, large scale optimization
- Abstract:Support vector machines are trained by solving constrained quadratic optimization problems. This is usually done with an iterative decomposition algorithm operating on a small working set of variables in every iteration. The training time strongly depends on the selection of these variables. We introduce the maximum-gain working set selection algorithm for large scale quadratic programming. It is based on the idea to greedily maximize the progress in each single iteration. The algorithm takes second order information from cached kernel matrix entries into account. We prove the convergence to an optimal solution of a variant termed hybrid maximum-gain working set selection. This method is empirically compared to the prominent most violating pair selection and the latest algorithm using second order information. For large training sets our new selection scheme is significantly faster.
  full paper : gzipped postscript, pdf
IRINI 2005-04
- Title: "The Multi-objective Variable Metric Evolution Strategy, Part I"
- Author: Christian Igel, Nikolaus Hansen, and Stefan Roth
- Date:October 2005
- Keywords:multi-objective optimization, evolution strategy, covariance matrix adaptation
- Abstract:The covariance matrix adaptation evolution strategy (CMA-ES) is one of the most powerful evolutionary algorithms for real-valued single-objective optimization. Here a variant of the CMA-ES for multi-objective optimization (MOO) is developed. First a single-objective, elitist CMA-ES using plus-selection and step size control based on a success rule is introduced. This algorithm is compared to the standard CMA-ES. The elitist CMA-ES turns out to be slightly faster on unimodal functions, but is more prone to getting stuck in sub-optimal local minima. In the new multi-objective CMA-ES (MO-CMA-ES) a population of individuals that adapt their search strategy as in the elitist CMA-ES is maintained. These are subject to multi-objective selection. The selection is based on non-dominated sorting using either the crowding-distance or the contributing hypervolume as second sorting criterion. Both the elitist single-objective CMA-ES and the MO-CMA-ES inherit important invariance properties, in particular invariance against rotation of the search space, from the original CMA-ES. The benefits of the new MO-CMA-ES in comparison to the well-known NSGA-II and NSDE, a multi-objective differential evolution algorithm, are experimentally shown.
  full paper : gzipped postscript, pdf

Internal Reports 2005