Yann Viémont’s research while affiliated with Versailles Saint-Quentin-en-Yvelines University and other places

This paper presents a new approach for query cost evaluation that may help or replace the known analytical approach. Our proposed approach is based on neural networks and the connectionist concept. A neural network is trained to learn the execution cost of the implementation algorithm(s) for a logical algebra operation (or query) with some predicates; after that, this network is used to estimate this operation (query) cost with other entries. The approach is based on a curve fitting like since neural networks have been proven to be "universal approximators." An additional advantage of this approach is its applicability to user defined methods where the user does not need to estimate the cost of its method since the system may apply this method several times, collects measurements, and capture its behavior with its curve fitting capacity. Keywords: Databases, Query Cost Evaluation, Performance Prediction, Neural Networks. 1 Introduction Cost models are still a problematic issue in data...

This paper describes how a geographical hypermedia model, called GeoAnchor, which integrates the hypermedia paradigm with geographical requirements, uses maps as a browsing device. With GeoAnchor, a map is built dynamically as a cartographic view of the hyperbase: each displayed geometry is an anchor to either a geographic node or to a related node. Hence, the map on the screen acts both as an index to the nodes and as a view to the hyperbase. We show that sucha map can be used as a metaphor for the representation of all the hyperbase elements. This metaphor relies on two main mechanisms: location inheritance for non-spatial nodes and a weighting scheme that provides for multi-scale management.

DBMS functionalities have been identified as a requirement for GIS (Geographical Information Systems). With the rise of G1S, several system architectures have been proposed for both commercial products and research prototypes, bringing G[S technology closer to database tecttnology. The approach described in this paper consists in adding spatial data capabilities to a data server while regrouping all graphical display functions into a user interface layer built as a client on the server. An SQLlike language. GQL, is proposed for the client-server interface, while a graphical language including drawings, display attributes and GQL queries is offered at the end-user interface level. An application exampte shows how this architecture can be used for fast protoo,ping with a high level of user friendliness and without writing any application code.

Managing and manipulating spatial data requires selective access to these data, performance at access level, and data protection. The need for a single spatial database system, integrating in a single model both alphanumeric and geometric data, is discussed. Such a database system, GéoSabrina, has been developed as an extension of an already existing relational database system. The system supports spatial data with corresponding spatial operators. Data manipulation is easily made through a spatial extension of SQL. Optimization techniques are used, providing for a strong integration of spatial operators based on spatial indexing and join-indices techniques. Built-in spatial functionalities may be further extended by application programmers through object oriented features.

Topological data structures are useful for reducing the cost of geometrical operations within a Geographic Information System (GIS). Unfortunately, manipulating such data structures can be quite complexespecially when supporting multiple, overlapping geographical maps. The GeoGraph storage model proposed in this paper solves this problem. It is implemented as a toolbox, and is used as a low-level system layer for support of a GIS. The Ge"oGraph storage model is based on a graph with the corresponding basic traversal primitives, and can be integrated within an extensible relational DBMS so that important spatial operations can be directly executed by graph traversals. Furthermore, the graph is decomposable so that only the useful subset of the database can be loaded from disk without format conversion.

In this paper, we investigate a novel method for database selectivity estimation. This method is based on neural networks and is similar to the curve fitting approach. A neural network is trained until it learns an attribute distribution over a given domain. After this learning phase, the network is used to estimate the selectivity of a predicate. The usefulness of neural networks in selectivity estimation relies on the proof that these are "universal approximators". A comprehensive experimental study is reported for several real, statistical, and highly irregular synthetic distributions. A comparison between the performance of neural networks, polynomials, and equi-height histograms for selectivity estimation is reported. The results show the proposed method to perform at least as well as the best of the other two for most tested distributions but the very irregular ones.