Research activity in 1999
In 1999, the group continued to be active on several topics in collaboration with the group of Machine Learning at the University of Piemonte Orientale at Alessandria: for what concerns the integration of numerical and symbolic learning techniques in First Order Logic, the systems NTR and FONN previously developed have been further extended. Moreover, the techniques used in these systems have been integrated into a the new version of G-Net, a learning system based on genetic algorithms.
The studies on Phase Transitions in Matching and Learning, initiated in 1998, continued in 1999 with a thorough experimentation on artificial and real datasets.
The location of the phase transition divides the problem space in three regions: One in which the probability of the existence of a solution is almost zero, and then it is «easy» to prove unsolvability. Another one, where many alternative solutions exist, and then it is «easy» to find one, and, finally, one where the probability of solution abruptly changes from almost 1 to almost 0, potentially making very difficult to find one of the existing solutions or to prove unsolvability.
Interestingly, we found that this last region is an attractor for many relational learning systems, in the sense that even though the search space might be very large, the used heuristics guide the systems to search for hypotheses inside this region.
Another topic investigated in 1999 is the use of IFS (Iterated Function Systems) for isolated 2D-image classification. Automatic recognition of complex images is a hard and computationally expensive task. The main reason is that it is extremely difficult to encode the necessary discriminant information in a few features. In this work we have tackled the problem by using IFS, which allow the use of Machine Learning techniques for building image classification systems. In particular, an enhanced version of the XFF algorithm has been developed, by extending the ability to represent (code) a given image with a set of freely rotated self-affine copies of itself.
Finally, the research activity of the group is also centered around the extraction of statistical dependencies among data. New paradigms that allow to find the actual significant dependencies are currently studied in depth. The new concepts of data dependency find some applications not only for structured data and databases but also for WEB data and texts.
For what concerns national projects, a subgroup has been involved in a collaboration with CSELT (the research labs of Telecom Italia) aimed at developing a tool for pruning and visualize association rules. Moreover, the group is involved in the national project 'Intelligent Agents: Interaction and Knowledge Acquisition', whose objective is the realization of a multi-agent architecture based technology, that allows one to build applications for the extraction and synthesis of knowledge from structured and unstructured data accessible on the Web.
1999 Publications
M. Botta, A. Giordana, L. Saitta "An Experimental Study of Phase Transitions in Matching", in Proceedings of the 16th International Joint Conference on Artificial Intelligence, IJCAI 99, (Stockholm, Sweden, 1999), pp. 1198-1203.
M. Botta, A. Giordana, L. Saitta, M. Sebag "Relational Learning: Hard Probelms and Phase Transitions", in Proceedings of the 6th Congress of AI*IA, (Bologna, Italy, 1999), pp. 99-111.
M. Botta "Knowledge Discovery in Databases and Data Mining", AI*IA Notizie, pp. 5-7, June 1999.
M. Rossotto, M. Botta "S3: Sistema di Supporto alla Segmentazione, Manuale Utente", Tech. Report DPP 1999.00653, CSELT, 21 pag., May 1999
M. Botta, A. Giordana, G. Lo Bello, L. Saitta "Algoritmi di Ricerca di Regole di Associazione", Tech. Report DPP 1999.03274, CSELT, 26 pag., November 1999.
