Catálogo de publicaciones - libros

The problem of mechanized deduction requires carrying out research and comparison of different methods for inference search in first-order classical logic, such as resolution-type methods, the model elimination method, the SLD-resolution, and so on. In this connection, it is desired to give a way for investigating their common and distinct features in order to use them better in theory and practice. This paper is devoted to such an investigation. Interconnection between a complete extension of the SLD-type resolution having the form of literal trees calculus and a ertain resolution technique is established. The interconnection permits to obtain some results on soundness and completeness for different resolution-type methods in the case of the weakest requirements to the factorization. In addition, when classical logic with equality is considered, it gives a possibility to make an original way for complete incorporation of the paramodulation into resolution with weak factorization as well as into the model elimination method.

Learning scenario determination is one of the key tasks of every Intelligent Learning Systems (ILS). This paper presents a method for learner classification in ILS based on rough classification methods proposed by Pawlak. The goal of rough learner classification is based on the selection of such a minimal set of learner profile attributes and their values that can be used for determination of optimal learning scenario. For this aim the problems of rough classification are defined and their solutions are presented.

This article introduces a new form of ethogram that provides a basis for studying reinforcement learning in biologically inspired collective robotics systems. In general, an ethogram is a record of behavior patterns, which has grown out of ethology (ways to explain agent behavior). The rough set approach introduced by Zdzisław Pawlak in 1982 provides a ground for deriving pattern-based rewards in the context of an approximation space. The framework provided by an approximation space makes it possible to derive pattern-based reference rewards used to compute action rewards as well as action preferences. A brief description of a prototype of an ecosystem testbed used to record ethograms in a dynamically changing system of agents is presented. The contribution of this article is an introduction to an ethological approach to the study of action preferences and action rewards during reinforcement learning in intelligent systems considered in the context of approximation spaces.

This paper presents the method of automatic knowledge retrieval from the web. The aim of the system that implements it, is to automatically create entries to a knowledge database, similar to the ones that are being provided by the volunteer contributors. As only a small fraction of the statements accessible on the web can be treated as valid knowledge concepts we considered the method for their filtering and verification, based on the similarity measurements with the concepts found in the manually created knowledge database. The results demonstrate that the system can retrieve valid knowledge concepts both for topics that are described in the manually created database, as well as the ones that are not covered there.

Knowledge Visualizer (KV) uses a General Logic Diagram (GLD) to display examples and/or various forms of knowledge learned from them in a planar model of a multi-dimensional discrete space. Knowledge can be in different forms, for example, decision rules, decision trees, logical expressions, clusters, classifiers, and neural nets with discrete input variables. KV is implemented as a module of the inductive database system VINLEN, which integrates a conventional database system with a range of inductive inference and data mining capabilities. This paper describes briefly the KV module and then focuses on the problem of arranging attributes that span the diagram in a way that leads to the most readable rule visualization in the diagram. This problem has been solved by applying a simulated annealing.

One of the classic data mining problems is discovery of frequent item-sets. Frequent itemset discovery tasks can be regarded as advanced database queries specifying the source dataset, the minimum support threshold, and optional constraints on itemsets. We consider a data mining system which supports storing of results of previous queries in the form of materialized data mining views. Previous work on materialized data mining views addressed the issue of reusing results of one of the previous frequent itemset queries to efficiently answer the new query. In this paper we present a new approach to frequent itemset query processing in which a collection of materialized views can be used for that purpose.

In this article we present two grammars (GramCat and GramEsp) for chunking of unrestricted Catalan and Spanish texts. With these grammars we extend the classical notion of as it is defined by Abney, taking advantage of Catalan and Spanish morphosyntactic features: Catalan and Spanish rich inflectional morphology and the high frequency of some prepositional patterns allow us to include both pre- and post-nominal modifiers in the noun phrase.

Minimal perfect hashing provides a mapping between a set of unique words and consecutive numbers. When implemented with minimal finite-state automata, the mapping is determined only by the (usually alphabetical) order of words in the set. Addition of new words would change the order of words already in the language of the automaton, changing the whole mapping, and making it useless in many domains. Therefore, we call it static. Dynamic minimal perfect hashing assigns consecutive numbers to consecutive words as they are added to the language of the automaton. Dynamic perfect hashing is important in many domains, including text retrieval and databases. We investigate three methods for its implementation.

A set of 70 lexical symbols is defined which covers written Polish. Under the assumption of the second order Markov process, a dictionary-based method of tagging parts of speech is presented. Instead of having to be trained on earlier tagged text, parameters of the method are estimated on frequencies of unambiguously classifiable unigrams, bigrams and trigrams of lexical symbols found in untagged literary text. The method is being proved to tag correctly 88% of all words in text.

Support Vector Machines have been applied to text classification with great success. In this paper, we apply and evaluate the impact of using part-of-speech tags (nouns, proper nouns, adjectives and verbs) as a feature selection procedure in a European Portuguese written dataset — the Portuguese Attorney General’s Office documents.

From the results, we can conclude that verbs alone don’t have enough information to produce good learners. On the other hand, we obtain learners with equivalent performance and a reduced number of features (at least half) if we use specific part-of-speech tags instead of all words.