We present a theory of boosting probabilistic classifiers. We place ourselves in the situation of a user who only provides a stopping parameter and a probabilistic weak learner/classifier and compare three types of boosting algorithms: probabilistic Adaboost, decision tree, and tree of trees of ... of trees, which we call matryoshka. "Nested tree," "embedded tree" and "recursive tree" are also appropriate names for this algorithm, which is one of our contributions. Our other contribution is the theoretical analysis of the algorithms, in which we give training error bounds. This analysis suggests that the matryoshka leverages probabilistic weak classifiers more efficiently than simple decision trees.

I ns ti t ut Gal i lé e - U niv. P ar is - Nord 93430 V il l et ane us e - F RA NC E noui ouaf @l ipn.uni v-pa ri s 13.fr G eneral ly s peaking, form al NL s em antic s i s re ferenti al i .e. it as sum es t hat i t is pos si ble t o c reate a s tati c dis course uni verse and to equat e t he obj ect s of t his uni verse t o the (s tat ic) mea nings of w ords . The me aning of a sent ence is then buil t from t he me anings of the w ords in a c ompos iti onal proces s and the se mant ic inte rpretat ion of a s entenc e i s reduce d to it s logic al i nterpret ati on bas ed on t he t ruth condit ions . The very diffic ult tas k of ada pting the mea ning of a s ent ence to its c ontext is often left to the pragm ati c l evel, and this tas k re quires t o us e a huge a mount of com mon s ens e know ledge a bout the domai n. It has bee n s howe d t hat the above tri-pa rtit ion i s very arti fici al becaus e l inguis ti c a s we ll as e xtra-li nguis tic know ledge i nterac t i n t he s am e gl obal proces s to provide the ne ces sa ry elem ents for unders ta nding. But what kind of rea soni ng is needed for na tural language se manti cs? T he ans we r to thi s que st ion is bas ed on the remark t hat t exts s eldom provide norma l det ail s t hat are a ss umed to be known to the reader.

In this note we introduce the notion of islands for restricting local search. We show how we can construct islands for CNF SAT problems, and how much search space can be eliminated by restricting search to the island.

This paper belongs to the area of learning theory that has been variously referred to as prediction with expert advice, competitive on-line prediction, p rediction of individual sequences, and universal on-line learning; see [7] for a re view. There are many proof techniques known in this field; this paper is based on K alnishkan and Vyugin's Weak Aggregating Algorithm [16], but it is possible that som e of the numerous other techniques could be used instead. In Section 2 we give the main definitions and state our main results, Th e-orems 1-4; their proofs are given in Sections 3-6. In Section 7 we inf ormally discuss the notion of stationarity, and Section 8 concludes.

We present in this paper a general algorithm for solving first-order formulas in particular theories called "decomposable theories". First of all, using special quantifiers, we give a formal characterization of decomposable theories and show some of their properties. Then, we present a general algorithm for solving first-order formulas in any decomposable theory "T". The algorithm is given in the form of five rewriting rules. It transforms a first-order formula "P", which can possibly contain free variables, into a conjunction "Q" of solved formulas easily transformable into a Boolean combination of existentially quantified conjunctions of atomic formulas. In particular, if "P" has no free variables then "Q" is either the formula "true" or "false". The correctness of our algorithm proves the completeness of the decomposable theories. Finally, we show that the theory "Tr" of finite or infinite trees is a decomposable theory and give some benchmarks realized by an implementation of our algorithm, solving formulas on two-partner games in "Tr" with more than 160 nested alternated quantifiers.

In this paper, we address the class of distributed systems wh ere computing units are autonomous mobile robots (also sometimes referred to sensors or agents), i.e., devices equipped with sensors which do not depend on a central scheduler and designed to mov e in a two-dimensional plane. Also, we assume that the robots cannot remember any previous obser vation nor computation performed in any previous step. Such robots are said to be oblivious (or memoryless). The robots are also uniform and anonymous, i.e, they all have the same program using no local parameter (such that an identity) allowing to differentiate any of them. Moreover, no ne of them share any kind of common coordinate mechanism or common sense of direction, and they communicate only by observing the position of the others.

The formalizations of periods of time inside a linear model of Time are usually based on the notion of intervals, that may contain or may not their endpoints. This is not enought when the periods are written in terms of coarse granularities with respect to the event taken into account. For instance, how to express the inter-war period in terms of a {\em years} interval? This paper presents a new type of intervals, neither open, nor closed or open-closed and the extension of operations on intervals of this new type, in order to reduce the gap between the discourse related to temporal relationship and its translation into a discretized model of Time.

Questions related to the evolution of language have recently known an impressive increase of interest (Briscoe, 2002). This short paper aims at questioning the scientific status of these models and their relations to attested data. We show that one cannot directly model non-linguistic factors (exogenous factors) even if they play a crucial role in language evolution. We then examine the relation between linguistic models and attested language data, as well as their contribution to cognitive linguistics.

Diagrammatic, analogical or iconic representations are often contrasted with linguistic or logical representations, in which the shape of the symbols is arbitrary. The aim of this paper is to make a case for the usefulness of diagrams in inferential knowledge representation systems. Although commonly used, diagrams have for a long time suffered from the reputation of being only a heuristic tool or a mere support for intuition. The first part of this paper is an historical background paying tribute to the logicians, psychologists and computer scientists who put an end to this formal prejudice against diagrams. The second part is a discussion of their characteristics as opposed to those of linguistic forms. The last part is aimed at reviving the interest for heterogeneous representation systems including both linguistic and diagrammatic representations.

A standard approach in pattern classification is to estimate the distributions of the label classes, and then to apply the Bayes classifier to the estimates of the distributions in order to classify unlabeled examples. As one might expect, the better our estimates of the label class distributions, the better the resulting classifier will be. In this paper we make this observation precise by identifying risk bounds of a classifier in terms of the quality of the estimates of the label class distributions. We show how PAC learnability relates to estimates of the distributions that have a PAC guarantee on their $L_1$ distance from the true distribution, and we bound the increase in negative log likelihood risk in terms of PAC bounds on the KL-divergence. We give an inefficient but general-purpose smoothing method for converting an estimated distribution that is good under the $L_1$ metric into a distribution that is good under the KL-divergence.