Ebook: Artificial Intelligence Research and Development

This volume contains the original contributions accepted for presentation at the 12th International Conference of the Catalan Association for Artificial Intelligence (CCIA'09), held in the ancient salt mining town of Cardona, on October 21–23, 2009.

The Catalan Association for Artificial Intelligence (ACIA),

AClA is a member of ECCAI, the European Coordinating Committee for Artificial Intelligence (http://www.acia.org). Its name is an acronym of Associació Catalana d'Intel·ligència Artificial.

The collaboration between ACIA members and the world AI community at large is well-established for many decades now. The papers submitted to the conference reflect that long-standing collaboration; most of its authors come from Catalan-speaking regions of Spain, but also from neighboring countries such as France and Italy, and farther countries such as Mexico and Australia. Over the years, CCIA has become steadily more international as the conference settled itself. This international character can also be noted in the present edition in the scientific committee composition, with 20% of it constituted of reputed researchers from the international AI community.

Each of the 60 papers submitted to CCIA'09 was carefully reviewed by at least 2 program committee members and finally 47 papers have been accepted for presentation during the conference. Of those contributions, the 45 original ones are presented in this volume as well as a brief summary of the two invited talks. The accepted papers, from most of the areas of AI, were divided into 27 oral presentations and 20 posters, but no difference has been made in the proceedings, as the division was not based in quality but on the suitability of each paper for one or the other kind of presentation.

Of all the areas of Computer Science, AI is the one most intertwined with all sorts of disciplines dealt with in the human experience, often employing lessons from one discipline to achieve a task in another one. For example, algorithms that mimic the behavior of humble creatures can be applied to the search for celestial bodies in the depths of the universe. This diversity is also reflected in the invited talks in this conference: one involving music, Scientific and technological challenges of musical interest, presented by Prof. Xavier Serra, and the other involving animal cognition, How great apes think, presented by Prof. Josep Call.

We would like to express our sincere gratitude to all the authors, who chose this conference to send their work, and to the members of the organizing committee, who took care of the uncountable detailed tasks necessary to have a successful conference. We heartily thank all the members of the Scientific Committee and other reviewers for the hard task of judging the submitted papers. Special thanks go also to the plenary speakers for their effort in preparing very interesting lectures and, last but not least, to the president of the ACIA, Núria Agell, for her kind and constant support.

Cardona, October 2009

Sandra Sandri, Institut d'Investigació en Intel·ligència Artificial, CSIC

Miquel Sànchez-Marrè, Universitat Politècnica de Catalunya

Ulises Cortés, Universitat Politécnica de Catalunya

Generic object recognition in mobile robots is of primary importance in order to enhance the representation of the environment that robots will use for their reasoning processes. Towards this aim, the contribution of this paper is an evaluation of the SIFT Object Recognition method in a challenging dataset, focusing on issues relevant to mobile robotics. The method presented robustness to the typical problems of images acquired in the robotics domain, but its good performance was limited mainly to well-textured objects.

Safe mobility in rough terrain is important for high-risk missions. In order to achieve stability and mobility precision it is necessary to have a good knowledge of the terrain properties. This work treats the problem of outdoor classification terrain analyzing proprioceptives sensor data, current measure, wheel speeds and slippage. The use of principal component analysis (PCA) reduces the space dimension of acquired data to a lower dimension to classificate in which terrain the robot is moving. This paper presents experimental results to validate the proposed methodology.

In this paper we propose the construction of a visual content layer which describes the visual appearance of geographic locations in a city. We captured, by means of a Mobile Mapping system, a huge set of georeferenced images (>500K) which cover the whole city of Barcelona. For each image, hundreds of region descriptions are computed off-line and described as a hash code. All this information is extracted without an object of reference, which allows to search for any type of objects using their visual appearance. A new Visual Content layer is built over Google Maps, allowing the object recognition information to be organized and fused with other content, like satellite images, street maps, and business locations.

Text detection in urban scenes is a hard task due to the high variability of text appearance: different text fonts, changes in the point of view, or partial occlusion are just a few problems. Text detection can be specially suited for georeferencing business, navigation, tourist assistance, or to help visual impaired people. In this paper, we propose a general methodology to deal with the problem of text detection in outdoor scenes. The method is based on learning spatial information of gradient based features and Census Transform images using a cascade of classifiers. The method is applied in the context of Mobile Mapping systems, where a mobile vehicle captures urban image sequences. Moreover, a cover data set is presented and tested with the new methodology. The results show high accuracy when detecting multi-linear text regions with high variability of appearance, at same time that it preserves a low false alarm rate compared to classical approaches.

In this paper, we argue that only using behavioural motion information, we are able to predict the interest of observers when looking at face-to-face interactions. We propose a set of movement-related features from body, face, and mouth activity in order to define a set of higher level interaction features, such as stress, activity, speaking engagement, and corporal engagement. Error-Correcting Output Codes framework with an Adaboost base classifier is used to learn to rank the perceived observer's interest in face-to-face interactions. The automatic system shows good correlation between the automatic categorization results and the manual ranking made by the observers. In particular, the learning system shows that stress features have a high predictive power for ranking interest of observers when looking at of face-to-face interactions.

This paper reviews new challenges in the area of long-term navigation, and new approaches to environment representation and robots capable of coping with dynamic environments. As a result of this review, we propose an appearancebased simultaneous localization and mapping (SLAM) solution which represents the robot environment using an appearance-based topological map. Dynamic environment changes are dealt with using human memory and fixed action pattern concepts. The former is used to build a histogram to register local feature stability, the latter for robot navigation purposes. We take omnidirectional vision and laser range data to extract textured 2D scans as global features, and textured-vertical edges as local features for map updating and robot localization. From the navigational point of view, we consider visual potential field-based behavior to adjust high level motion commands.

Object segmentation is a challenging and important problem in computer vision. The difficulties to obtain accurate segmentations using only the traditional Top-down or Bottom-up approaches have introduced new proposals based on the idea of combining them in order to obtain better results. In this paper we present a novel approach for object segmentation based on the following two steps: 1) oversegment the image in homogeneous regions using a Region Growing algorithm (Bottom-up), and 2) use prior knowledge about the object appearence (local patches and spatial coherence) from annotated images to validate and merge the regions that belong to the object (Top-down). Our experiments using different object classes from the well-known TUD and the Weizmann databases show that we are able to obtain good object segmentations from a generalistic segmentation method.

In this work, we present a first approach to activity patterns discovery by mean of topic models. Using motion data collected with a wearable device we prototype, TheBadge, we analyse raw accelerometer data using Latent Dirichlet Allocation (LDA), a particular instantiation of topic models. Results show that for particular values of the parameters necessary for applying LDA to a countinous dataset, good accuracies in activity classification can be achieved.

This paper presents a methodology for generating coordination motions task performed by articulated mobile robots. It is considered for the case that both, states defining the task dynamics are not completely observable and model-based solutions fail to scale up due to the complexity of unstructured problems. The novel approach try to optimize primitive trajectories to achieve joint coordination emerging as a consequence of the dynamic followed by the actuators. A simple dynamical system is proposed as basic trajectory generator, becoming a nonlinear solution in the task space due to the actual nonlinearities and constraints of the robot. Joints share information while act along its common natural environment, the robot body. Policies gathering and distributing signals to the actuators are optimized based on a measure of the task's performance. Trajectories are generated through the experiences of the robot, rather than computed from a mathematical model of its body. The simulated version of the AIBO robot completing a ball throwing task is used to illustrate the effectiveness of the proposed scheme.

Shout and Act (S&A) is an evolution of Bar Systems, a family of algorithms for different classes of complex optimization problems in static and dynamic environments by reactive multi agent systems. We adapt these systems to RoboRescue, where robots explore land looking for victims. When they find someone they “shout” so that robot mates can hear it. The louder the shout, the most important or urgent the finding. Louder shouts can also refer to closeness. Several experiments show that this system works very scalably, and how heterogeneous teams of robots outperform homogeneous ones over a range of task complexity. Finally, our results impact the design of RoboRescue teams: a properly designed combination of robots is cheaper and more scalable when confronted with uncertain maps of victims.

Often, dynamic patterns are constructed from the concatenation of the patterns identified on each state of the process. In this work it is shown that the interpretation of dynamic patterns can not be constructed as a simple concatenation of the concepts associated to the components of the dynamic process. This phenomenon is defined as a Paradox of the Mitigant Trajectories (PTM). A method to identify PTM's is proposed as well as a method to correct them. Also, the proposal is applied to a Wastewater Treatment Plant (WWTP). The correct interpretation of dynamic patterns is crucial as inducted knowledge to be considered for an input to Intelligent Environmental Decision Support Systems (IEDSS).

One of the most widely used assumptions in supervised learning is that data is independent and identically distributed. This assumption does not hold true in many real cases. Sequential learning is the discipline of machine learning that deals with dependent data such that neighboring examples exhibit some kind of relationship. In the literature, there are different approaches that try to capture and exploit this correlation, by means of different methodologies. In this paper we focus on meta-learning strategies and, in particular, the stacked sequential learning approach. The main contribution of this work is two-fold: first, we generalize the stacked sequential learning. This generalization reflects the key role of neighboring interactions modeling. Second, we propose two different ways of capturing and exploiting sequential correlations that takes into account long-range interactions by means of a multi-scale decompositions of the predicted labels: one using a gaussian function and another using a frequency function. Moreover, this new method subsumes the standard stacked sequential learning approach. We tested the proposed method on image classification task. Results on this task clearly show that our approach outperforms the standard stacked sequential learning. Moreover, we show that the proposed method allows to control the trade-off between the detail and the desired range of the interactions.

This paper introduces a new approach to enhance learning in adjustment processes by using a support vector machine (SVM) algorithm as discriminant function jointly with an action generator module. The method trains a SVM with state-action patterns and uses trained SVM to select an appropriate action given a certain state in order to reach the target state. The system incorporates a simulated annealing technique to increase the exploration capacity and improve the ability to avoid local minima. The methodology has been tested in an example with artificial data.

This paper presents the collaborative agent-based learning subsystem of HealthAgents, a multiagent distributed decision support system for brain tumour diagnosis. The subsystem aims to boost the performance of the independent and heterogeneous classifiers in spite of the limited data transfer conditions prevailing in the system. The susbsystem is composed by local autonomous agents which are interacting among them, following an existing collaborative learning model. The different aspects and decisions dodged during the adaptation of this model are described in addition to the results of its initial evaluation with the data of HealthAgents. Significant increments of classification performance attained by the learning agents demonstrate the potential benefits of this subsystem.

This paper presents results of a project from the Master of Information and Communication Technologies at the University of the Balearic Islands. The aim of the work is the application of intelligent approaches to data analysis. These techniques have been applied to the area of social assistance in order to construct an efficient structure that classifies persons depending on the profiles available at the organization's data bases. This classification is later used to assign the professional expert that best fits the user's characteristics.

The design and the validation of an automatic plaque characterization technique based on Intravascular Ultrasound (IVUS) usually requires a data ground-truth. The histological analysis of post-mortem coronary arteries is commonly assumed as the state-of-the-art process for the extraction of a reliable data-set of atherosclerotic plaques. Unfortunately, the amount of data provided by this technique is usually few, due to the difficulties in collecting post-mortem cases and phenomena of tissue spoiling during histological analysis. In this paper we tackle the process of fusing in-vivo and in-vitro IVUS data starting with the analysis of recently proposed approaches for the creation of an enhanced IVUS data-set; furthermore, we propose a new approach, named pLDS, based on semi-supervised learning with a data selection criterion. The enhanced data-set obtained by each one of the analyzed approaches is used to train a classifier for tissue characterization purposes. Finally, the discriminative power of each classifier is quantitatively assessed and compared by classifying a data-set of validated in-vitro IVUS data.

Continuous domains are domains where cases are generated from a continuous data stream. In these domains, a lot of cases are continuously solved and learned by a CBR system. This means that many cases could be stored in the case library. Thus the efficiency of the CBR system both in size and time could be deeply worsened. In this research work a dynamic adaptive case library (DACL) is proposed. It is able to adapt itself to dynamic environments by means of a set of dynamic clusters of cases and a discriminant tree associated to each cluster. The prototype of a cluster is called a Meta-Case. The aim is to get an optimal and competent case library that works efficiently in a continuous domain. In this paper, the improvement of time efficiency in the retrieval step has been evaluated by means of testing several data bases. The result shows a good improvement using the proposed DACL approach.

We present a mathematical framework for communicating about trust in terms of interactions. We argue that sharing an ontology about trust is not enough and that interactions are the building blocks that all trust- and reputation models use to form their evaluations. Thus, a way of talking about these interactions is essential to gossiping in open heterogeneous environments. We give an overview of the formal framework we propose for aligning trust, based on the gossip agents send.