Gilles Venturini

On how pachycondyla apicalis ants suggest a new search algorithm

In this paper we present a new optimization algorithm based on a model of the foraging behavior of a population of primitive ants (Pachycondyla apicalis). These ants are characterized by a relatively simple but efficient strategy for prey search in which individuals hunt alone and try to cover a given area around their nest. The ant colony search behavior consists of a set of parallel local searches on hunting sites with a sensitivity to successful sites. Also, their nest is periodically moved. Accordingly, the proposed algorithm performs parallel random searches in the neighborhood of points called hunting sites. Hunting sites are created in the neighborhood of a point called nest. At constant intervals of time the nest is moved, which corresponds to a restart operator which re-initializes the parallel searches. We have applied this algorithm, called API, to numerical optimization problems with encouraging results.

On Improving Clustering in Numerical Databases with Artificial Ants

We present in this paper a new hybrid algorithm for data clustering. This algorithm discovers automatically clusters in numerical data without prior knowledge of a possible number of cleisses, without any initial partition, and without complex parameter settings. It uses the stochastic eind exploratory principles of an ant colony with the deterministic and heuristic principles of the K-means cJgorithm. Ants move on a 2D bosird and may load or drop objects. Dropping aa object on an existing heap of objects depends on the similarity between this object and the heap. The K-means algorithm improves the convergence of the ant colony clustering. We repeat two stochastic/deterministic steps and introduce hierarchical clustering on heaps of objects and not just objects. We also use other refinements such as aji heterogeneous population of ants to avoid complex parameters settings, and a local memory in each ant. We have applied this algorithm on standard databases cind we get very good results compared to the K-means and ISODATA algorithms.

A hierarchical ant based clustering algorithm and its use in three real-world applications

In this paper is presented a new model for data clustering, which is inspired from the self-assembly behavior of real ants. Real ants can build complex structures by connecting themselves to each others. It is shown is this paper that this behavior can be used to build a hierarchical tree-structured partitioning of the data according to the similarities between those data. Several algorithms have been detailed using this model (called AntTree): deterministic or stochastic algorithms that may use or not global or local thresholds. Those algorithms have been evaluated using artificial and real databases. Our algorithms obtain competitive results when compared to the Kmeans, to ANTCLASS, and to Ascending Hierarchical Clustering. AntTree has been applied to three real world applications: the analysis of human healthy skin, the on-line mining of web sites usage, and the automatic construction of portal sites.

A Critical and Empirical Study of Epistasis Measures for Predicting GA Performances: A Summary

Epistasis measures have been developed for measuring statistical information about the difficulty of a function to be optimized by a genetic algorithm (GA). We give first a review of the work on these measures such as the epistasis correlation. Then we try to relate the epistasis correlation to the overall performances of a binary GA on a set of 14 functions. The only relation that seems to hold strongly is that a high epistasis correlation implies GA-easy, as indicated by the GA theory of deceptiveness. Then, we show that changing the representation of the search space with transformations that improve epistasis measures does not imply the same increasing in the genetic algorithm performances. These both empirical studies seem to indicate that the generality of epistasis measures is limited.

Data and Text Mining with Hierarchical Clustering Ants

Summary. In this paper is presented a new model for data clustering, which is inspired from the self-assembly behavior of real ants. Real ants can build complex structures by connecting themselves to each others. It is shown is this paper that this behavior can be used to build a hierarchical tree-structured partitioning of the data according to the similarities between those data. Several algorithms have been detailed using this model : deterministic or stochastic algorithms that may use or not global or local thresholds. We have also introduce an incremental version of our artificial ants algorithm. Those algorithms have been evaluated using artificial and real databases. Our algorithms obtain competitive results when compared to the Kmeans, Ascending Hierarchical Clustering, AntClass and AntClust (two biomimetic methods). Our methods have been applied to three real world applications: the analysis of human healthy skin, the on-line mining of web sites usage, and the automatic construction of portal sites. Finally, we have developed two possibilities to explore the portal site. The first possibility consists in representing the tree in HTML pages in order to explore the portal site with a conventional browser. The second possibility we have studied is to integrate the results of our algorithms in our virtual reality data mining tool VRMiner (39).

Optimizing Hidden Markov Models with a Genetic Algorithm

In this paper is presented the application of genetic algorithms (GAs) to the learning of hidden Markov models (HMMs). The Baum-Welch algorithm (BW), which optimizes the coefficients of a HMM, is improved by the use of a GA. The GA is able to find rapidly a good initial model compared to random generation, and this initial model is optimized further with BW. A representation and adapted genetic operators have been introduced in order to evolve matrix of probabilities. Several tests on artificial data show the interest in using a GA with BW.

FIBRASCAN: a novel method for 3D white matter tract reconstruction in MR space from cadaveric dissection.

INTRODUCTION Diffusion tractography relies on complex mathematical models that provide anatomical information indirectly, and it needs to be validated. In humans, up to now, tractography has mainly been validated by qualitative comparison with data obtained from dissection. No quantitative comparison was possible because Magnetic Resonance Imaging (MRI) and dissection data are obtained in different reference spaces, and because fiber tracts are progressively destroyed by dissection. Here, we propose a novel method and software (FIBRASCAN) that allow accurate reconstruction of fiber tracts from dissection in MRI reference space. METHOD Five human hemispheres, obtained from four formalin-fixed brains were prepared for Klinglers dissection, placed on a holder with fiducial markers, MR scanned, and then dissected to expose the main association tracts. During dissection, we performed iterative acquisitions of the surface and texture of the specimens using a laser scanner and two digital cameras. Each texture was projected onto the corresponding surface and the resulting set of textured surfaces was coregistered thanks to the fiducial holders. The identified association tracts were then interactively segmented on each textured surface and reconstructed from the pile of surface segments. Finally, the reconstructed tracts were coregistered onto ex vivo MRI space thanks to the fiducials. Each critical step of the process was assessed to measure the precision of the method. RESULTS We reconstructed six fiber tracts (long, anterior and posterior segments of the superior longitudinal fasciculus; Inferior fronto-occipital, Inferior longitudinal and uncinate fasciculi) from cadaveric dissection and ported them into ex vivo MRI reference space. The overall accuracy of the method was of the order of 1mm: surface-to-surface registration=0.138mm (standard deviation (SD)=0.058mm), deformation of the specimen during dissection=0.356mm (SD=0.231mm), and coregistration surface-MRI=0.6mm (SD=0.274mm). The spatial resolution of the method (distance between two consecutive surface acquisitions) was 0.345mm (SD=0.115mm). CONCLUSION This paper presents the robustness of a novel method, FIBRASCAN, for accurate reconstruction of fiber tracts from dissection in the ex vivo MR reference space. This is a major step toward quantitative comparison of MR tractography with dissection results.

A New Approach of Data Clustering Using a Flock of Agents

This paper presents a new bio-inspired algorithm (FClust) that dynamically creates and visualizes groups of data. This algorithm uses the concepts of a flock of agents that move together in a complex manner with simple local rules. Each agent represents one data. The agents move together in a 2D environment with the aim of creating homogeneous groups of data. These groups are visualized in real time, and help the domain expert to understand the underlying structure of the data set, like for example a realistic number of classes, clusters of similar data, isolated data. We also present several extensions of this algorithm, which reduce its computational cost, and make use of a 3D display. This algorithm is then tested on artificial and real-world data, and a heuristic algorithm is used to evaluate the relevance of the obtained partitioning.

Fast Unsupervised Clustering with Artificial Ants

AntClust is a clustering algorithm that is inspired by the chemical recognition system of real ants. It associates the genome of each artificial ant to an object of the initial data set and simulates meetings between ants to create nests of individuals that share a similar genome. Thus, the nests realize a partition of the original data set with no hypothesis concerning the output clusters (number, shape, size ...) and with minimum input parameters. Due to an internal mechanism of nest selection and finalization, AntClust runs in the worst case in quadratic time complexity with the number of ants. In this paper, we evaluate new heuristics for nest selection and finalization that allows AntClust to run on linear time complexity with the number of ants.

Epistasis for real encoding in genetic algorithms

Epistasis is a well known tool introduced by Davidor (1991) to understand and predict the performance of a genetic algorithm using binary encoding. The meaning of variance of epistasis is analyzed using Walsh basis; it is established that this variance can be viewed as a measure of the quality of a linear approximation. Variance of epistasis is therefore easier to estimate using the well known context of polynomial approximation regarding Euclidian distance. This approach also leads to the extension of variance of epistasis to real encoding without considering the notion of schema. Following this new definition, numerical analysis tools can also be applied to compute efficiently this generalized variance of epistasis.