ALCUNI ARGOMENTI PER TESI

Alessandro Sperduti

The NAUTILUS

NAvigate AUtonomously and Target Interesting Links for USers

The Internet is a powerful medium to share information worldwide but its distributed and dynamical nature makes it difficult to organize complete and up-to-date indexes of its contents. Thus the Internet user has to navigate through many pages and links to discover the information he/she is interested in. The use of search engines is helpful but usually yields too many links most of which are loosely related to the actual interest of the user. The choice of appropriate queries to search engines is crucial but in any case the user is still required to browse directly through all the suggested pages in order to find the desired ones.

We propose a different approach that has been investigated recently by many researchers involved in machine learning systems. The idea is to learn the user's interests and provide him/her with suggestions of interesting web sites or pages automatically. The NAUTILUS is an intelligent robot that navigates the Web autonomously looking for pages that are potentially interesting for the registered users it manages. The NAUTILUS navigates in the background fetching pages and evaluating them with respect to the current user profiles stored in the system. The score obtained for each page is used to decide if the page should be suggested to the user and if the automatic navigation should continue through the links contained in that page. The learning module of NAUTILUS takes into account the users' feedbacks on the suggested links and modifies consequently the profiles.



Structured domains are characterized by complex patterns which are usually represented as lists, trees, and graphs of variable sizes and complexity. The ability to recognize and classify these patterns is fundamental for several applications that use, generate or manipulate structures. Examples of application domains are medical and technical diagnoses (discovery and manipulation of structured dependencies, constraints, explanations), molecular biology (DNA and protein analysis), chemistry (classification of chemical structures, quantitative structure-property relationship (QSPR), quantitative structure-activity relationship (QSAR)), automated reasoning (robust matching, manipulation of logical terms, proof plans, search space reduction), software engineering (quality testing, modularization of software), geometrical and spatial reasoning (robotics, structured representation of objects in space, figure animation, layouting of objects), speech and text processing (robust parsing, semantic disambiguation, organizing and finding structure in texts and speech). While algorithms that manipulate symbolic information are capable of dealing with highly structured data, adaptive neural networks are mostly known as learning models for domains in which instances are organized into {\em static} data structures, like records or fixed-size arrays. Recurrent neural networks, that generalize feedforward networks to sequences (a particular case of dynamically structured data) are perhaps the best known exception.
However, during the last few years, neural networks for the representation and processing of structures have been developed. In particular, recursive neural networks are a generalization of recurrent networks from sequences (i.e., linear chains from a graphical point of view) to directed acyclic graphs. These kind of networks are of paramount importance both for structural pattern recognition and for the development of hybrid systems, since they allow the treatment of structured information very naturally and, in several cases, very efficiently. Let consider, for example, the use of neural networks for classification of labeled graphs which represent chemical compounds. The standard approach with feedforward neural networks consists of encoding each graph as a fixed-size vector, which is then given as input to the network for classification. Unfortunately, the a priori definition of the encoding process has main drawbacks (for example, in chemistry encoding is performed through the definition of topological indexes which are designed by a very expensive trial and error approach) and the resulting vectorial representations of graphs may be very difficult to classify. The purpose of this tutorial is to examine the state of the art in the use of connectionist networks for data structures and to present a unified view of formalisms and tools for dealing with rich data representations, covering connectionist architectures for data structures, learning algorithms, and applications. In particular, we will show that using a generalization of a recurrent neuron --- the generalized recursive neuron --- it is possible to represent, classify, and store structured information very naturally. Moreover, it is possible to formalize several supervised models for classification of structures which stem very naturally from well known models, such as backpropagation through time networks, real-time recurrent networks, simple recurrent networks, recurrent cascade correlation networks, and neural trees. Because many concepts and formal tools are inherited from the theoretical framework of recurrent networks for sequence processing, the tutorial will begin with a review of recurrent neural networks, for those attendees which are not familiar with such a class of models.

Argomento 1:: Uso delle tecniche di apprendimento automatico durante la fase di modellazione di un problema reale come problema di vincoli con preferenze. Questa lavoro si potra' basare su una tesi precedente che ha gia' sviluppato un sistema di apprendimento automatico e ha effettuato degli esperimenti su problemi random. La nuova tesi dovra' valutare il lavoro precedente su altri problemi, confrontarlo ed eventualmente combinarlo con altre tecniche, e costruire un ambiente user-friendly (possibilmente basato su form su Internet) per la modellazione di problemi di vincoli tramite apprendimento automatico.
Argomento 2:: Definizione di un formalismo di vincoli temporali con preferenze, e studio delle sue proprieta' e delle classi trattabili. Possibile collaborazione con il centro NASA Ames di Moffett Field, CA, per usare questo sistema su problemi di scheduling di eventi in missioni NASA.

Neural Network Methodologies in Gas Turbine Diagnostic

Argomento :

Studio di un sistema di diagnosi e monitoraggio per turbine a gas con l'intento di identificare possibili situazioni di guasto o di danneggiamento presenti nell'impianto di generazione turbogas. In collaborazione con ENEL Pisa.

Alessandro Sperduti ( perso@di.unipi.it)