Research Activity

go back to my  home page

  list of publications 
  current research interests 
  curriculum 
 

During the first years of activity research focussed on Machine Learning topics and in particular: design and development of neural network learning algorithms, development of an algorithm for learning symbolic mutually-dependant relations, synthesis and refinement of control functions, reinforcement learning, integration of different techniques.

Reasoning and adaptation
After the year 2000 the activity in the Machine Learning field decreased and was substituted by new interests related to forms of adaptation obtained by applying inferencing techniques. In the beginning, the work focussed on the design of a virtual computer seller, accessible via the web, and able to assist a user who wanted to assemble a PC with the aim of a specific use. The main application, however, consisted in the development of a virtual tutor that supports university students in the creation of personalized curricula; the system can also verify the correctness of curricula proposed by the students, explaining them possible mistakes. The mentioned systems are examples of applications in which the user model is not sufficient to achieve adaptation, which can be obtained only by reasoning on the user's intentions and goals. In these works different reasoning techniques have been exploited: procedural planning, temporal projection, and temporal explanation. All the systems have been implemented as agents that could accomplish the described tasks. The main publications describing these works are [6] and [7]; other information can be retrieved by following this link.

Reasoning and interaction protocols
The most recent activity concerns the study of forms of adaptation and personalization based on reasoning about agent (in a broad sens) interaction protocols. The current context is given by the emerging area of web services. Both the personalization of the use of a web service and web service automatic composition can benefit from the application of reasoning about the communicative behavior of a resource. The main publications describing the results achieved are [8] and [9].

Acquisition of spatial maps for autonomous agents
Aside I have also been interested in the acquisition and manipulation of spatial maps, executed by autonomous agents that can explore their environment. Currently maps are bidimensional and agents are only simulated. The technique consists in exploiting neural networks with a dynamic topology, mainly Growing Neural Gas networks, that can discretize a continuous input space. The main ideas have been described in [10].

Duties

CURRICULUM

1992	Part-time job at CSELT (Centro Studi e Laboratori Telecomunicazioni), Torino, where I designed and implemented a single-epoch learning algorithm for RCE neural networks (Restricted Coulomb Energy). Afterwards, I have integrated the obtained system with Hidden Markov Models. The developed software has been applied to the classification of utterences of the ten digits. In parallel, cooperation with prof. Attilio Giordana, Computer Science Dept. of the University of Torino, to the development of a method for learning mututally recursive (symbolic) relations.
1993	With the beginning of the Ph.D. studies, I joined the ESPRIT project B-Learn II, aimed at studying the use of Artificial Intelligence (in particular, Machine Learning) techniques to the automatic synthesis of artificial agents controllers. I took care of learnt controllers refinement by means of Reinforcement Learning. Implementation of system RTL (Recursive Theory Learner) for learning mututally recursive relations.
1994	Guest at the Department of Mechanical Engeneering at the Katholieke Universiteit Leuven, Leuven, Belgium, (partner in the B-Learn II project) in february, march and september. From april through june, guest of the Adaptive Networks group at the University of Massachusetts (Amherst, MA, USA), directed by prof. Andy Barto. Here I developed the Shaping algorithm, that allows to bias the learning-by-interaction with the environmente process of an artificial agent by using a feedback produced by a qualitative (non-operational) knowledge of the world and of the task to be learned. Machine Learning Summer School in Dourdan, France. System RTL has been integrated with Smart+ (by Botta and Giordana).
1995	Implementation of a system for automatcally training artificial agents, in which a controller (implemented as a neural network that can alternatively be a RBFN, a Fuzzy Controller or a dynamic topology network) is trained by an artificial tutor obtained by compiling a qualitative knowledge base, containing non-operational knowledge about the goal to achieve. The system has been tested against the cart-pole and against an assembly task known as peg-in-hole. In june/july, guest of the Katholieke Universiteit Leuven. Invited speaker at AIT95, International Workshop on Artificial Intelligence Techniques, Brno, Tcheck Republic.
1996	member of the program commettee of AIT96. For two months, member of the European project MEDALUS-III aimed at building an automatic system for classifying meteorological states, in cooperation with ERSAL, Ente Regionale per lo Sviluppo Agricolo della Lombardia. I became dottore di ricerca presenting a thesis with title Soft Teaching Start of cooperation to a research project about the extraction of characteristic fractal features from 2D images, exploiting a coding based on Iterated Function System (IFS).
1997	I took part to the development of XFF, a system for the extraction of fractal features that capture the characteristic information of an isolated object starting from an image of it. The algorithm produces similar codings for similar images, which can be used in tasks of object recognition or classification. Between 1997 and 1998, research (founded by ABB research, S. S. Giovanni, MI, Italy) about the possibility of applying Machine Learning techniques to the automatic diagnosis of faults in electromechanical switches. Start of cooperation to the AUDIP project, aimed at building tools for self-evaluation for Computer Science students.
1998	The projects begun in the previous year have been continued. In particular I developed a more general version of XFF, that can handle transformations with any rotation. I also carried on tests for the ABB project. Post-doc at the University of Torino.
1999	Continuing the research on IFS condings, with a grant from the project Giovani Ricercatori; continuing the work to the AUDIP project. Researcher at the Department of Computer Science, Faculty of Sciences, from the Università degli Studi di Torino. Since then, I am part of the Logic Programming and Automated Reasoning group, directed by prof. Alberto Martelli.
2005	From november 1st I am an associate professor at the Department of Computer Science, Faculty of Mathematical Physical and Natural Sciences, University of Torino.