Ebook: Artificial Intelligence Research and Development

Although the benefit of scientific research belongs to the collective heritage of mankind, its roots are deeply entwined inside the many cultural backgrounds and the scattered local communities which are home to the researchers. Year after year, CCIA, the conference organized by the Catalan Association for Artificial Intelligence (ACIA), serves the twofold purpose of bringing together researchers from the Catalan speaking areas while at the same time broadcasting their work to the international scene by publishing full proceedings in English. This is the 18th edition of CCIA, and it is nowadays a mature conference that has succeeded to fulfill the goals decided 18 years ago when ACIA launched its first conference. At present, CCIA is a major reference for Catalan researchers working on Artificial Intelligence, but at the same time it is also open to scientists from other countries, who have also greatly contributed to the success of it (as members of the program committee, as reviewers, as authors, etc.).

Full versions of the communications presented at the 18th International Conference of the Catalan Association for Artificial Intelligence (CCIA 2015) are collected in this book. Additionally, shorter contributions displayed as posters are also included in this volume. All of them have been fully peer-reviewed.

The scope of CCIA conferences is AI in a broad sense. The topics of the papers provide an accurate indication of which are the dominant research trends around the Catalan speaking community. This year's contributions show a preference for Classification, Decision Support Systems and Data Mining, but many other topics are covered, ranging from theoretical aspects to description of real applications. An examination of the recent CCIA proceedings provides a measure of the dynamism and vigor which drives the pulse of the Catalan AI community.

The organizers want to publicly show their gratitude to the Escola Tècnica Superior d'Enginyeria, Universitat de València, which has hosted this year's edition of the conference.

València, October 2015

Eva Armengol, IIIA-CSIC

Dionís Boixader, ETSAV-UPC

Francisco Grimaldo, ETSE-UV

This work introduces new results on early-vocal development in infants and machines using artificial intelligent agents. It is addressed using the perspective of intrinsically-motivated learning algorithms for autonomous exploration. The agent autonomously selects goals to explore its own sensorimotor system in regions where a certain competence measure is maximized. Unlike previous experiments, we propose to include a somatosensory model to provide a proprioceptive feedback to reinforce learning. We argue that proprioceptive feedback will drive the learning process more efficiently than algorithms taking into account only auditory feedback. Considering the proprioceptive feedback to generate a constraint model, which is unknown beforehand to the learner, guarantees that the agent is less prone to selecting goals that violated the system constraints in previous experiments.

We develop a model for 2 agents to reach agreement over concept meaning in specific contexts. The model is based on an argumentation-based communication that engage the agents in a process of mutual adaptation using argumentation to reach an agreement over concept meaning. Our approach is to model concept meaning using the semiotic triangle and the notion of contrast sets. We implement and evaluate present three common sense scenarios where two agents argue and reach agreements over the contextual meaning of concepts.

Decentralized energy production is meant to reduce generation and distribution inefficiencies, leading to major economic and environmental benefits. This new model is meant to be supported by smart grids, electricity networks that can intelligently integrate the actions of all users connected to them —generators, consumers, and prosumers (those that do both)— to efficiently deliver sustainable, economic and secure electricity supplies. A major research challenge is the design of markets for prosumers in smart grids that consider distribution grid constraints. Recently, a discrete market model has been presented that allows prosumers to trade electricity while satisfying the constraints of the grid. However, most of the times energy flow problems possess a continuous nature, and that discrete market model can only provide approximate solutions. In this paper we extend the market model to deal with continuous (piecewise linear) utility functions. We also provide a mapping that shows that the clearing of such a market can be done by means of integer linear programming.

In a theoretical setting, the Representation Theorem is used to generate T-indistinguishability operators (fuzzy similarity relations) from a given set of fuzzy criteria. In applied domains, though, other ways of generation are often used which involve quasi-arithmetic means. In this paper we study when one single fuzzy subset, obtained by averaging multiple fuzzy criteria, is able to generate the same T-indistinguishability operator obtained by all of them, either exactly or approximately.

This paper is a preliminary study of the universal Horn fragment of Predicate Fuzzy Logics. We work in languages with a binary predicate symbol that is interpreted as a similarity. Using this similarity relation we define a term structure associated to a theory, and we prove that it is a free structure on the class of reduced models of the theory. Finally, we show that the substructure generated by the set of ground terms is a model of all universal Horn sentences that are logical consequences of the theory.

Preferences are part of every day life driving to choice and action. We consider that there is a gap between preferences expressed by people and those we can find in the repositories. In this paper we explore a small set of preferences in the domain of movies, given by humans, in order to understand the expressive possibilities of some languages appearing in the literature: conditional logics and reward logics. After some experiments we contribute with a proposal for reasoning about preferences.

Given a finite totally ordered set of linguistic terms, the operations intersection and connected union provide a lattice structure to the set of hesitant fuzzy linguistic term sets. In this framework, hesitant fuzzy linguistic descriptions of a given set of alternatives are considered. A distance between hesitant fuzzy linguistic descriptions is introduced based on the properties of the lattice. This metric structure allows distances between decision makers to be computed. A centroid or group representative is defined in group decision-making processes where decision makers are assessing different alternatives by means of hesitant fuzzy linguistic term sets.

Methodologies for pattern extraction and analysis from neural activity data captured by simultaneous sensors, are gaining major interest due to technological advances in sensor devices. From early Electroencephalography (EEG), able to capture a few simultaneous signals, to in-vivo spinal recording, Magneto Encephalograpy (MEG) or functional Magnetic Resonance Imaging (fMRI), that capture up to hundreds of signals, the amount of data from neurophysiological experimentation to be analyzed multiplies every year.

This paper proposes a methodology composed by several steps. It is targeted to extract qualitative and quantitative information from experiments recorded using such devices. The process selects patterns of interest for the hypothesis in the experiment, analyzes the characteristics of the patterns and its changes in temporal behavior from the perspective of individual sensors, identifies coocurrences of patterns from the simultaneous sensors and extracts frequent temporal patterns that show recurrent collective sensor activity. These patterns are expected to characterize and discriminate different states or conditions in the experiments.

An application to the analysis of recordings of spinal neural activity in cats is presented to show the use of the methodology.

Exploiting network data (i.e., graphs) is a rather particular case of data mining. The size and relevance of network domains justifies research on graph mining, but also brings forth severe complications. Computational aspects like scalability and parallelism have to be reevaluated, and well as certain aspects of the data mining process. One of those are the methodologies used to evaluate graph mining methods, particularly when processing large graphs. In this paper we focus on the evaluation of a graph mining task known as Link Prediction. First we explore the available solutions in traditional data mining for that purpose, discussing which methods are most appropriate. Once those are identified, we argue about their capabilities and limitations for producing a faithful and useful evaluation. Finally, we introduce a novel modification to a traditional evaluation methodology with the goal of adapting it to the problem of Link Prediction on large graphs.

When inducing Decision Trees, Windowing consists in selecting a random subset of the available training instances (the window) to induce a tree, and then enhance it by adding counter examples, i.e., instances not covered by the tree, to the window for inducing a new tree. The process iterates until all instances are well classified or no accuracy is gained. In favorable domains, the technique is known to speed up the induction process, and to enhance the accuracy of the induced tree; while reducing the number of training instances used. In this paper, a Windowing based strategy exploiting an optimized search of counter examples through the use of GPUs is introduced to cope with Distributed Data Mining (DDM) scenarios. The strategy is defined and implemented in JaCa-DDM, a novel system founded on the Agents & Artifacts paradigm. Our approach is well suited for DDM problems generating large amounts of training instances. Some experiments in diverse domains compare our strategy with the traditional centralized approach, including an exploratory case study on pixel-based segmentation for the detection of precancerous cervical lesions on colposcopic images.

Twitter has become one of the most popular Location-Based Social Networks (LBSNs) that enables bridging physical and virtual worlds. Tweets, 140-character-long messages published in Twitter, are aimed to provide basic responses to the What's happening? question. Occurrences and events in the real life are usually reported through geo-located tweets by users on site. Uncovering event-related tweets from the rest is a challenging problem that necessarily requires exploiting different tweet features. With that in mind, we propose Tweet-SCAN, a novel event discovery technique based on the density-based clustering algorithm called DB-SCAN. Tweet-SCAN takes into account four main features from a tweet, namely content, time, location and user to cluster homogeneously event-related tweets. This new technique models textual content through a probabilistic topic model called Hierarchical Dirichlet Process and introduces Jensen-Shannon distance for the task of neighborhood identification in the textual dimension. As a matter of fact, we show Tweet-SCAN performance in a real data set of geo-located tweets posted during Barcelona local festivities in 2014, for which some of the events were known beforehand. By means of this data set, we are able to assess Tweet-SCAN capabilities to discover events, justify using a textual component and highlight the effects of several parameters.

Nowadays ubiquitous connectivity, portable computing, pervasive sensing, novel interfaces, cheap and fast computing units, and advances in robotic devices and actuators are changing our lives, our living environments, and our social interaction. To truly benefit the elderly and fragile population, commodities based on these novel technologies need to be autonomous and interactive, and must be capable of anticipating user needs, managing complex and unforeseen situations on their own, seamlessly interfacing with casual end-users, and gracefully terminating their functioning when unrecoverable errors occur. Our main aim is to provide a model for developing a multi-agent system integrated into a medical social network. It must provide a tool for developing assistive services to support elderly patients with disabilities in their daily life.

In the Anchoring Problem actions and objects must be anchored to symbols; and movement primitives as DMPs seems a good option to describe actions. In the bottom-up approach to anchoring, the recognition of an action is done applying learning techniques as clustering. Although most work done about movement recognition with DMPs is focus on weights, we propose to use the shape-attractor function as feature vector. As several DMPs formulations exist, we have analyzed the two most known to check if using the shape-attractor instead of weights is feasible for both formulations. In addition, we propose to use distance-based kernels, as RBF and TrE, to classify DMPs in some predefined actions. Our experiments based on an existing dataset and using 1-NN and SVM techniques confirm that shape-attractor function is a better choice for movement recognition with DMPs.

We present a feature selection method for neuroimaging techniques to find objective criteria for diagnosis of schizophrenia. The method is based on kernel alignment with the ideal kernel using Support Vector Machines (SVM) in order to detect relevant features for the diagnostic task.

The method has been applied to a dataset obtained using multichannel MagnetoEncephalograpy (MEG), from a set individuals composed by patients with chronic schizophrenia stable compensate, patients with the same diagnosis but in an acute exacerbation state, and a control group. The diagnosis of the schizophrenia is characterized in this paper as differences of synchronism between different parts of the brain, so correlations among sensors readings for different brain areas are used as features. All signal frequency bands are also analyzed, from δ to high frequency γ, to find the best band for diagnosis.

One of the main advantages of the proposed method is that it is less prone to overfitting than other approaches. This requirement is essential in neuroimaging where the number of features representing recordings is usually very large compared with the number of recordings. Another advantage is the ablility to visualize brain areas showing different correlations in control individuals compared with correlations in patients. The proposed methodology can be easily applied to other pathologies.

Metric learning has been shown to outperform standard classification based similarity learning in a number of different contexts. In this paper, we show that the performance of classification similarity learning strongly depends on the sample format used to learn the model. We also propose an enriched classification based set-up that uses a set of standard distances to supplement the information provided by the feature vectors of the training samples. The method is compared to state-of-the-art metric learning methods, using a linear SVM for classification. Results obtained show comparable performances, slightly in favour of the method proposed.

The success of portfolio approaches in SAT solving relies on the observation that different SAT solving techniques perform better on different SAT instances. The Algorithm Selection Problem faces the problem of choosing, using a prediction model, the best algorithm from a predefined set, to solve a particular instance of a problem. Using Machine Learning techniques, this prediction is performed by analyzing some features of the instance and using an empirical hardness model, previously built, to select the expected fastest algorithm to solve such instance.

Recently, there have been some attempts to characterize the structure of industrial SAT instances. In this paper, we use some structural features of industrial SAT instances to build some classifiers of industrial SAT families of instances. Namely, they are the scale-free structure, the community structure and the self-similar structure. We measure the effectiveness of these classifiers by comparing them to other sets of SAT instances features commonly used in portfolio SAT solving approaches. Finally, we evaluate the performance of this set of structural features when used in a real portfolio SAT solver.