Research Projects

This page has information about the research projects in which we have worked (finished project) and the research projects in which we are working now (current projects).

CURRENT PROJECTS

Classification Methods Based on Belief Networks
Supported by:CICYT (Comisión Interministerial de Ciencia y Tecnología)
Project number: TIC96-0781
Team:
- Silvía Acid Carrillo
- Manuel Jorge Bolaños
- Luis Daniel Hernández Molinero
- Juan F. Huete Guadix
- Serafín Moral Callejón
Starting and ending dates: from 1-08-96 to 31-07-99
Manager: Luis Miguel de Campos Ibañez. (
lci@decsai.ugr.es)
Abstract:
Classification methods are very useful in a great number of different domains of knowledge (medical diagnosis is a typical example). The techniques for automatic classification start from a training set containing the atributes measured to a set of individuals, together with their correct classification. Then, they build a representation model (a tree, a set of rules, a statistical model,...), which is used to predict the classification of new individuals. On the other hand, belief networks are knowledge representation tools based on a graphical model, which contains qualitative information about relevance or cause-effect relationships, and a numerical model, that contains quantitative information about the strenghts of these relationships. Once a belief network has been built (either from expert knowledge, or automatically from a data set), it can be used to efficiently perform reasoning tasks. The main objective of this project is the development of new methods for automatic classification, based on belief This project can be considered as the second part of another, previously funded, research project, in which we developed learning methods for belief networks in non-purely probabilistic formalisms, based on the concept of conditional independence. Now, we intend to continue in this direction, but also using methods based on a fitting measure between the network and the data set and heuristic search. We want to focus on the application of these methods to classification problems. The combination of these methods and other, more classical, classification techniques, as for example, classification trees, could lead to the development of more powerful classifiers.
Keywords:Classification, Belief Networks, Learning.

FINISHED PROJECTS

Desarrollo de un Sistema de Representación del Conocimiento Evidencial para Visión Artificial.
Supported by: Junta de Andalucía (Spain).
Starting and ending dates: from October 1988 to October 1990.

Defeasible Reasoning and Uncertainty Management Systems (DRUMS).
Supported by: C.E.E. Esprit II Basic Research Action.
Project number: 3085
Starting and ending dates: from September 1989 to September 1992.

Comparative Testing of Statistical and Logical Learning Algorithms on Large-Scale Applications to Classification, Prediction and Control (STATLOG).
Supported by: C.E.E. Esprit Project.
Project number: 5170
Starting and ending dates: from October 1990 to October 1993.

Uso de Medidas Difusas para el Tratamiento de la Incertidumbre en Inteligencia Artificial.
Supported by: DGICYT (Spain).
Project number: PS89-0152
Starting and ending dates: from November 1990 to November 1993.

Automatic knowledge acquisition in causal networks
Supported by: DGICYT (Spain)
Project number: PB92-0939
Manager: Luis M. de Campos. (lci@decsai.ugr.es)
Team

Silvía Acid
Jorge Bolaños
Juan F. Huete
Serafín Moral
Juan F. Verdegay-López
Starting and ending dates:31-05-93/31-05-96
Abstract:
Causal networks constitute one of the most interesting uncertain knowledge representation formalisms. Once the network has been built, there exists efficient procedures to handle it and perform inferences with the knowledge codified in the network. However, the usual way to build the network from the experts is slow and hard.
This project focuses on the problem of learning causal networks from examples. Although some probabilistic-based methods for learning causal networks exist, no methods based on non probabilistic formalisms of uncertainty (as possibility theory, evidence theory, upper and lower probability theory) are available. When the inference problem to be solved consists in the classification of a new object as belonging to one of several categories, in the light of some observed characteristics, the causal networks learning techniques do not take advantage of the special structure of this kind of problem to construct more efficient classifiers.
So, the main interest of the project is the development of general models for learning causal networks, and their subsequent adaptation to the special characteristics displayed by the classification problems. The theoretical study of such models of learning, together with implementation and experimentation aspects will be considered.
Keywords: Causal networks, Inductive learning, Uncertainty management.

Defeasible Reasoning and Uncertainty Management Systems II (DRUMS II).
Supported by: C.E.E. Esprit III Basic Research Action.
Project number: 6156
Starting and ending dates: from October 1992 to October 1995.
Manager: Serafín Moral. (smc@decsai.ugr.es)
Abstract:
This is a large project involving 21 european research groups. The coordinator is:
Philippe Smets
IRIDIA
Universite Libre de Bruxelles

DRUMS II is a continuation of the DRUMS I project in which defeasible reasoning, uncertainty and incompleteness were studied. The overall objective is to investigate symbolic and quantitative methods of representation and reasoning able to deal with knowledge that is uncertain and defeasible.
Six topics of fundamental importance in the domain of defeasible and uncertain reasoning have been selected. The work programs are:

WP1. Belief Change.
WP2. Non-monotonic deduction
WP3. Inconsistency in reasoning
WP4. Abduction
WP5. Algorithmic problems
WP6. Dynamic reasoning with partial models
Our group is involved in WP4 and WP5 work programs. Our participation in WP4 is coordinated by Luis M. de Campos and the objective is to study learning procedures of dependence graphs under non-probabilistic representations of uncertainty, as possibility theory and upper and lower probabilities. WP5 is coordinated by Serafín Moral and the objective is to study abstract propagation algorithms in graphical structures and approximated algorithms based on Monte-Carlo techniques.

Decision and Linguistic Probabilities. Propation in Dependence Graphs.
Supported by: DGICYT (Spain).
Project number: PB92-0959
Starting and ending dates: from June 1993 to June 1996.
Manager: María Teresa Lamata. (mtl@decsai.ugr.es)
Team

Moral Callejón , Serafín
Cano Ocaña , José Enrique
Hernández Molinero, Luis Daniel
Verdegay López , Juan Francisco
Cano Utrera , Andrés
Abstract:
Firstly, we will try to make an study of different ways of representing lingüístic variables ( labels, probability intervals, fuzzy numbers ) and theirs management in Artificial Intelligence. We also going to study the linguistic variable sets.
Afterwards, we will make the theoretical study of the decision problem from the three points of view of the linguistic probability representation and the relation with the classic model. This work will be based on the valuation theory depthly studied in our work team in the last years.
In the third part we are going to study the linguistic variable propagation on dependence graphs. We will developpe the appropiated operators to solve the forward and backward propagation. The idea is to follow the previous line of ours team about conditioning.
In the last part, we are going to implement the results on a dependence graph.
Keywords:Linguistic probabilities, fuzzy number, intervals of probabilities, decision, valuations, dependence graphs.

Home UTAI Home DECSAI

Research Projects

CURRENT PROJECTS

Classification Methods Based on Belief Networks

FINISHED PROJECTS

Desarrollo de un Sistema de Representación del Conocimiento Evidencial para Visión Artificial.

Defeasible Reasoning and Uncertainty Management Systems (DRUMS).

Comparative Testing of Statistical and Logical Learning Algorithms on Large-Scale Applications to Classification, Prediction and Control (STATLOG).

Uso de Medidas Difusas para el Tratamiento de la Incertidumbre en Inteligencia Artificial.

Automatic knowledge acquisition in causal networks

Defeasible Reasoning and Uncertainty Management Systems II (DRUMS II).

Decision and Linguistic Probabilities. Propation in Dependence Graphs.