Ebook: Artificial Intelligence Research and Development

In this year of the first centenary of Alan Turing's birth, the Fifteenth International Conference of the Catalan Association of Artificial Intelligence (CCIA 2012) took place on October 24–26, 2012 at the Universitat d'Alacant (Spain). Alan Turing made a unique impact on the history of computer science, artificial intelligence, developmental biology and the mathematical theory of computability during his relatively brief life. In its own short life since 1994, the Catalan Association of Artificial Intelligence (ACIA) has, as a member of the European Coordinating Committee for Artificial Intelligence (ECCAI), fostered cooperation between researchers in Artificial intelligence within the Catalan-speaking community. One of the instruments that ACIA has used for this purpose has been the yearly organization of CCIA, an international conference where not only researchers from Catalan-speaking countries, but also from other countries in Europe and worldwide, have found a place to show, discuss and publish the results of their researches and developments.

In its fifteenth edition, CCIA 2012 received 27 original contributions that were carefully reviewed by three members of a steering committee. Of these 27 submissions, 23 were accepted for inclusion in this book for their relevance to CCIA 2012, originality, technical validity, and relevance of the conclusions. Despite the fact that the majority of papers were from Catalonia and Spain, CCIA also attracted submissions from Colombia, France, Italy, Mexico, The Netherlands and Venezuela.

Submissions correspond to the topics of KDD, DM and machine learning; natural language processing and recommenders; computer vision; robotics; AI for optimization problems, and AI applications in the real world.

CCIA 2012 also hosted two invited keynote speakers: Oscar Cordón and Eduardo Nebot.

Oscar Cordón has been professor at the Department of Computer Science and Artificial Intelligence of the University of Granada since 1995, and is a member of the Soft Computing and Intelligent Information Systems research group at this university. He was the head of the Virtual Learning Centre of the University of Granada (CEVUG) between 2001 and 2005. He has published more than 50 international journal papers, advised on nine PhD dissertations, participated in 15 research projects and contracts supported by the European Commission, the Spanish Government, the Andalusian Government, the University of Granada and Puleva Food S.A. (being the main researcher in five of them). He is co-editor of five special issues in different international journals and three books, and co-author of the book Genetic Fuzzy Systems: Evolutionary Tuning and Learning of Fuzzy Knowledge Bases. He also created, and since 2004 has chaired, the Genetic Fuzzy Systems Task Force in the Fuzzy Systems Technical Committee (IEEE Computational Intelligence Society) and is Area Editor of the International Journal of Approximate Reasoning and has been treasurer of the EUSFLAT Society since 2005. His current main research interests are in the fields of: fuzzy rule-based systems, genetic fuzzy systems, soft computing for forensic anthropology and medical imaging, evolutionary algorithms, ant colony optimization and other metaheuristics, and their applications to different topics (information retrieval, bioinformatics, etc.).

Eduardo Nebot is the Executive Director of the Australian Centre for Field Robotics, at the University of Sydney, Australia. This centre numbers more than 70 researchers in four core areas: sensors, fusion & perception, actuators, control & decision, modelling, learning & adaptation, and architectures, systems & cooperation. These four core research areas define the science of field robotics and intelligent systems and represent the main focus of the ACFR. Eduardo's main research interest is the field of robotics, focusing on the automation of open cast mining. He is the author of more than 100 journal and conference papers and has several patents in the area described above. His talk at the conference was about a variety of research in the area of open cast mining, involving intelligent vehicles, computer vision, pattern recognition, sensors, etc.

We would like to express our sincere gratitude to all the authors and members of the scientific and organizing committees who have made this conference a success. Our special thanks also go to the plenary speakers, Óscar Cordón and Eduardo Nebot for their efforts in preparing very interesting lectures, and to the president of the ACIA, Vicenç Torra, for his kind support.

Alacant, October 2012

David Riaño, Universitat Rovira i Virgili

Eva Onaindia, Universitat Politècnica de València

Miguel Cazorla, Universitat d'Alacant

Controlled clinical trials are studies done by scientists on a concrete domain with the objective of execute a prospective analysis that attempts to compare the effect and value of one or more interventions in humans under specific conditions. Our case study is focused on applying artificial intelligence techniques to support scientists working on bio-nutritional trials, which rely on the study of the effect of consuming food with biomarkers on human health. This work is developed under the HENUFOOD² project. In order to simplify the domain, two trial stages are considered: the trial experimental execution and the trial data analysis. It is after the trial has been executed and all the information collected, when we try to discover non obvious knowledge among the trial data and support the scientists on their everyday analytical tasks. On the one hand, this analysis includes generic statistical manipulation and data discovery empowered by data mining techniques. On the other hand, a scientist support engine has been developed in order to allow non data manipulation and analysis experts to access the information and ask for new analysis processes. The approach, thanks to a user friendly interface enabled with query answering techniques, is able to interact with the user in a conversational mode and run analytical or data discovery processes, do a semantic search on previous indexed works or, finally, infer new rules to feed the expert system ontology. The expert system is not fully covered on this paper, as it will focus on the explanation of the usage of data mining and query answering techniques on that bio-nutritional focused expert system.

At the very early stage of finding homogeneous and distinguishable patterns from data, many different clustering algorithms can be used, providing each of them different results. Further analysis is required to select the definite clustering. In this research, we are introducing Traffic Light Panel (TLP) for the first time in this discovery process, with the aim of using it to interpret the clusters and establishing relationships between the quality of a certain clustering and the aspect of the TLP and consider the TLP a goodness-of-clustering indicator itself. In a totally non-supervised context, were no previously validated references are available, new quality criteria are defined for the TLP and used to select among several possible clusterings of the same. Our methodological approach is validated with a financial data set from Venezuelan Stock Exchange providing good results, which have been confronted and supported with posterior feedback from expert financial analysts

In this paper we propose to use Gabriel graphs on standard Borderline-SMOTE, in order to improve its performance on severe two-class imbalance problem in the artificial neural networks context. The standard Borderline-SMOTE shows two drawbacks: 1) it only takes into account the number of neighbors, so information about prototypes distribution is lost. The global classifiers as neural networks need more information to define the borderline decision. 2) The standard Borderline-SMOTE requires a free parameter to find the borderline samples. The advantage of using Gabriel graphs is that it avoids setting free parameters. Empirical results obtained from experiments on real data sets show that the use of Gabriel graphs in Borderline-SMOTE improve the standard Borderline-SMOTE performance on neural networks.

Action learning is a methodology based on a machine learning system that makes it possible to select a suitable action or sequence of actions given a state. The main drawback of this methodology is the difficulty of assigning a class to the state-action pair to be included in the training set. This paper proposes an active learning methodology in the learning phase of an action learning process. With the help of an artificial example, the active methodology is compared with a passive methodology consisting of randomly selecting the training set from the pool of unlabelled patterns.

Re-identification methods are in common use in data privacy for disclosure risk assessment. Record linkage is one of these re-identification methods. Given two data files, record linkage establishes links between records that correspond to the same individual but are found in the two different files. In this paper we review our definition of re-identification, which is based on belief functions. The use of these functions permits us to model in an appropriate way the uncertainty inherent in the re-identification process. Then, we discuss record linkage for data masked using approaches that ensure k-anonymity, and for data masked using rank swapping p-bucket.

An approach is presented aimed to produce narrative descriptions of objects within digital images which are understandable by human beings. A context-free grammar is defined based on qualitative models for colour and shape description. This approach was tested using images of the MPEG-7 CE Shape-1 library and images of tile geometric pieces. The results show that a general or detailed narrative description can be obtained from different kinds of images containing one or more objects and they also support the further application of our approach to (i) teach children about shapes by giving them instructions on how play the game known as join-the-dots and (ii) help children with visual impairments to carry out visual oddity tasks in psychological tests.

Recommendation models based on collaborative filtering have good prediction accuracy; however, they have very poor performance in cold-start scenarios, when no or few rating data are associated to certain items or users. An approach for dealing with cold-start problem is to build hybrid models that include content-based filtering. This paper focuses on knowledge-based hybrids that exploit semantic relations among item attributes to enhance traditional models. We investigate novel methods for exploiting this semantics during the recommendation process, in domains where explicit semantic relations among attributes in the form of ontologies are not available. We carry out an extensive performance comparison in which the proposed semantically-enhanced models are evaluated in normal and cold-start scenarios. The experimental results with a publicly available dataset demonstrate superior better performance of the proposed semantically-enhanced models, above all in cold-start scenarios, with respect to a state-of-the-art collaborative-filtering model based on matrix factorization. The proposed models, additionally, provide for interpretable recommendations, in contrast to black-box models.

Gabor Filters have been extensively used to solve the texture-based image segmentation problem, following the filter bank and filter design approaches. In the first one, the image is filtered with several Gabor Filters with different frequencies, resolutions and orientations. The parameters of these filters are fixed and can be suboptimal for a particular processing task. The techniques based on filter design, on which this work is focused, permit to “tune” the parameters of the filter. This work proposes the use of two optimization algorithms (Guided Random Search and Particle Swarm) in this tuning process, showing good results in texture classification tests.

It has been shown that the temporal evolution of the face during an expression is important to its correct interpretation. In recent years some approaches have attempt to model facial expressions dynamics by different methods such as HMMs, Dynamic Bayesian Networks or Bayesian temporal manifold model. In this paper, we present a novel approach to model the expression temporal dynamics using Gaussian Process Regression. Specifically, we use it to map the geometric features in any sequence frame into continuous values related with the intensity of the different types of expressions. Using the evolution of these values along the sequence, we are able to determine if the subject has performed any expression. In tests with the Cohn-Kanade database we show that this method can reach a higher performance than other existing methods with a 97.4% of accuracy.

Human Pose Recovery approaches have been studied in the field of Computer Vision for the last 40 years. Several approaches have been reported, and significant improvements have been obtained in both data representation and model design. However, the problem of Human Pose Recovery in uncontrolled environments is far from being solved. In this paper, we define a global taxonomy to group the model based methods and discuss their main advantages and drawbacks.

The task of registering three dimensional data sets with rigid motions is a fundamental problem in many areas as computer vision, medical images, mobile robotic, arising whenever two or more 3D data sets must be aligned in a common coordinate system. In this paper we make a study of registration methods for egomotion calculation on robotics. We are interested in determining which is the best method to apply depending on execution time and accuracy. We have selected two of the most used methods: Iterative Closest Point (ICP) and Ransac (with visual features), and we have developed a fast implementation of ICP in the GPU. Several experiments have been done in order to determine which method must be select depending of the input data.

In this work we address the problem of object detection for the purpose of object manipulation in a service robotics scenario. Several implementations of state-of-the-art object detection methods were tested, and the one with the best performance was selected. During the evaluation, three main practical limitations of current methods were identified in relation with long-range object detection, grasping point detection and automatic learning of new objects; and practical solutions are proposed for the last two. Finally, the complete pipeline is evaluated in a real grasping experiment.

In nowadays aging society, many people require assistance for pedestrian mobility. In some cases, assistive devices require a certain degree of autonomy when the persons' disabilities difficult manual control. Our aim is to develop an open and innovative reference architecture, based upon ontologies and agent services, that will allow plug and play and cost-effective interconnection of existing and new services in all domains required for the independent and autonomous living of the elderly and their enhanced Quality of Life (QoL). We show how the use of a robotic platform with some embedded intelligence the i-Walker, helps to improve and speed-up the performance of the post-stroke individuals' rehabilitation.

This paper presents a method for fast calculation of the egomotion done by a robot using visual features. The method is part of a complete system for automatic map building and Simultaneous Localization and Mapping (SLAM). The method uses optical flow in order to determine if the robot has done a movement. If so, some visual features which do not accomplish several criteria (like intersection, unicity, etc,) are deleted, and then the egomotion is calculated. We use a state-of-the-art algorithm (TORO) in order to rectify the map and solve the SLAM problem. The proposed method provides better efficiency that other current methods.

The map building from data collected form the environment is an important field in robotics. The map could be used in different tasks, as localization, place recognition, obstacle avoidance, SLAM, etc. A topological map does not seek accurate measures, but the classification of the real environment in different areas. The use of learning techniques can help us to define areas which the robot is able to recognize in subsequent steps. In this paper, we propose the adaptation of the Viola-Jones supervised learning method based on AdaBoost to learn what visual features are good to classify an image into a given area. In our case, AdaBoost will select the best MSER features that best define each node of the map.

In this paper we introduce a new kind of immobile Location-Allocation problem that consists in determining the service each facility has to offer in order to maximize the covered demand given the positions of the customers and their service requirements and the positions of the immobile facilities. First, we provide a formalization of the problem and then we tackle the problem using two heuristic methods, genetic algorithms and simulated annealing, comparing the performance of both algorithms.