Ebook: Artificial Intelligence Research and Development

Artificial Intelligence (AI) has been from its beginning the wave of evolution in computer science. It is in good health and a proof of it is the fact that many companies qualify its novelties as ‘smart’ or ‘intelligent’ independently of the features included in them; the term ‘society of knowledge’ has been imposed to draw society nearer to the future and a symbol of breakthrough. From this perspective, AI has reached its maturity and it has exploded into an endless set of sub-areas, getting in touch with all other disciplines to assist situation assessment, analysis and interpretation of music, management of environmental and biological systems, planning trains, routing of communication networks, assisting medical diagnosis or powering auctions.

The wide variety of Artificial Intelligence application areas has meant that AI researchers often become scattered in different micro specialized conferences. There are few occasions where the AI research community joins together, while computer scientists and engineers can find a lot of interesting ideas from the cross fertilization of results coming from all of these application areas. The Catalan Association for Artificial Intelligence (ACIA

ACIA, the Catalan Association for Artificial Intelligence, is member of the European Coordinating Committee for Artificial Intelligence (ECCAI). http://www.acia.org.

The book is organized according to the different sessions in which the papers were presented at the Eighth Catalan Conference on Artificial Intelligence, held in Alguer (Italy) on October 26–28^th, 2005. Namely: Neural Networks, Computer Vision, Applications, Machine Learning, Reasoning, Planning and Robotics, and Multi-Agent Systems. Papers have been selected after a double blind review process in which distinguished AI researchers from all over Europe participated. Among the 77 papers received, 26 were selected as oral presentations and 25 as posters. The quality of the papers was high on average, and the selection between an oral or a poster session was based on the degree of discussion that a paper could generate more than on its quality. All of the papers collected in this volume would be of interest to any computer scientist or engineer interested in AI.

We would like to express our sincere gratitude to all the authors and members of the scientific and organizing committees that have made this conference a success. Our special thanks also to the plenary speakers for their effort in preparing the lectures.

Alghero, October 2005, Beatriz López (University of Girona), Joaquim Meléndez (University of Girona), Petia Radeva (Computer Vision Center, UAB), Jordi Vitrià (Computer Vision Center, UAB)

The first part of the presentation surveys fundamental insights in the ‘science of the artificial’ as coined by the Herbert Simon. From these insights a number of design requirements and principles for complex adaptive man-invented systems are derived.

The second part discusses a multi-agent coordination and control system answering the above requirements for a specific application domain. The system design is inspired by the behavior of social insects (ant colonies) and includes some ‘social engineering’ of the agent society. The discussion focuses on the society of the agents: its architecture, structure and interaction mechanisms; the internal architecture of a single agent is not discussed. This sample MAS design is computationally efficient and gives its user full control over the computation and communication effort where the result is ‘best effort’ only. These coordination and control systems handle a going concern; they do not limit themselves to one-shot problem solving.

In this paper, we present Policy Methods as an alternative to Value Methods to solve Reinforcement Learning problems. The paper proposes a Direct Policy Search algorithm that uses a Neural Network to represent the control policies. Details about the algorithm and the update rules are given. The main application of the proposed algorithm is to implement robot control systems, in which the generalization problem usually arises. In this paper, we point out the suitability of our algorithm in a RL benchmark, that was specially designed to test the generalization capability of RL algorithms. Results check out that policy methods obtain better results than value methods in these situations.

Given a set of available assets, the portfolio selection problem consists in finding out the best way of investing a particular amount of money in the assets. Some heuristic methods based on evolutionary algorithms, tabu search and simulated annealing have been developed in the past. Here we present a Hopfield neural network model to solve the portfolio selection problem, comparing the new results to those obtained with the other heuristic methods.

This work is devoted to the study of the Canet-en-Roussillon (south of France) activated sludge wastewater treatment plant (WWTP) process and to the estimation of chemical parameters (influent and effluent chemical oxygen demand and suspended solids concentration) not easily on-line measurable. Their knowledge makes it possible to estimate both process efficiency and impact on natural environment.

A tool based on neural networks, including an Elman recurrent network, Kohonen's self-organzing maps and a multi level perceptron has been developed. The Elman network is used for the prediction of the incoming WWTP influent flow, specially in case of rain events increasing the quantity of water to be treated. The Kohonen'self-organizing maps neural network is applied to analyse the multi-dimensional Canet-en-Roussillon process data, and to diagnose the inter-relationship of the process variables in an activated sludge WWTP. The multi level perceptron is used as COD and SS estimation tool.

The final purpose of Automated Vehicle Guidance Systems (AVGSs) is to obtain fully automatic driven vehicles to optimize transport systems, minimizing delays, increasing safety and comfort. In order to achieve these goals, lots of Artificial Intelligence techniques must be improved and merged. In this article we focus on learning and simulating the Human-Level decisions involved in driving a racing car. To achieve this, we have studied the convenience of using Neuroevolution of Augmenting Topologies (NEAT). To experiment and obtain comparative results we have also developed an online videogame prototype called Screaming Racers, which is used as test-bed environment.

This paper presents a new feature selection method and an outliers detection algorithm. The presented method is based on using a genetic algorithm combined with a problem-specific-designed neural network. The dimensional reduction and the outliers detection makes the resulting dataset more suitable for training neural networks. A comparative analysis between different kind of proposed criteria to select the features is reported. A number of experimental results have been carried out to demonstrate the usefulness of the presented technique.

An unsupervised segmentation algorithm based on a multiresolution method is presented. This method uses variational functions as a segmentation criterion. The algorithm has been applied to multispectral images of fruits as a quality assessment application. In addition, due to the unsupervised nature of the procedure, it can be applied to real images in order to test the quality of these results.

One of the applications of image registration is to assess object differences from various images that have been spatially correlated. This paper discusses the use of features extracted from subtracted registered images in a classification framework with an aim to detect abnormal mammograms. Both quantitative and qualitative results are provided, which show that although non-optimal classification is obtained, region features extracted after registration can be used to discriminate between normal and abnormal mammograms.

IntraVascular UltraSound (IVUS) imaging is a useful tool in diagnosis of cardiac diseases since sequences completely show the morphology of coronary vessels. Vessel borders detection, especially the external adventitia layer, plays a central role in morphological measures and, thus, their segmentation feeds development of medical imaging techniques. Deterministic approaches fail to yield optimal results due to the large amount of IVUS artifacts and vessel borders descriptors. We propose using classification techniques to learn the set of descriptors and parameters that best detect vessel borders. Statistical hypothesis test on the error between automated detections and manually traced borders by 4 experts show that our detections keep within inter-observer variability.

In this paper we present an unsupervised colour-texture segmentation method. The novelty of this new approach relies on the definition of the interfeature distance map, which combines different textural properties, either those due to chromaticity variations or those due to intensity changes. The proposed framework works on a multi-channel representation that directly depends on the basic colours perceived in an image, trying to avoid the spatial correlation of usual 3D representations of colour. The textural information derived from intensity changes is added to the multi-channel representation and finally, chromaticity and intensity are combined with the inter-feature distance maps. The behaviour of the approach is tested on a set of natural and synthetic images to demonstrate the capabilities to deal with some representative types of colour-textures which are suggested in this work.

In document analysis field, Optical Musical Recognition is a mature area in printed scores, whereas few research works have been done in handwritten ones. The difficulties in handwritten scores are increased if we work in old documents, because of paper degradation and the lack of a standard in musical notation. In this paper we propose a method to segment staff and graphical primitives in old handwritten scores. The extraction of staff lines has been performed using Hough Transform, skeletization, median filters and a contour tracking process. The segmentation of lines and head notes has been done using morphological operations and median filters. Our method has been tested with several scores of XIX century with high performance rates.

The use of context can be a very relevant cue for computer vision-based systems, in order to eliminate a lot of ambiguity and uncertainty, otherwise inherent, in the human-computer interaction. Despite of the fact that its obvious importance is widely acknowledged, the great majority of the systems nowdays still lack of this capability. In this paper, we propose faces as a primary contextual information for person detection and present face tracking as a basic procedure for context-driven focus of attention applications. Our proposal was implemented in a combined system, by integrating a motion-based approach (Particle Filter) and a model-based approach (Ada-Boost).

In general face classification problems, only the internal features of the face are commonly used, rejecting the information located at head, chin, and ears, since due to their variability is not easy to extract this information. In this paper, a complete scheme based on a Top-Down reconstruction algorithm to extract External Features of face images is proposed. We use the Non negative Matrix Factorization (NMF) algorithm to obtain the final coefficients that encode the external information and using this codification the faces are classified. Our experimental results in different face classification problems show that the information contributed by the external features is significant and useful for classification purposes.

Tagged Magnetic Resonance Imaging (MRI) is a non-invasive technique used to examine cardiac deformation in vivo. An Angle Image is a representation of a Tagged MRI which recovers the relative position of the tissue respect to the distorted tags. Thus cardiac deformation can be estimated. This paper describes a new approach to generate Angle Images using a bank of Gabor filters in short axis cardiac Tagged MRI. Our method improves the Angle Images obtained by global techniques, like HARP, with a local frequency analysis. We propose to use the phase response of a combination of a Gabor filters bank, and use it to find a more precise deformation of the left ventricle. We demonstrate the accuracy of our method over HARP by several experimental results.

We propose an hybrid and probabilistic classification of image regions belonging to scenes primarily containing natural objects, e.g. sky, trees, etc. as a first step in solving the problem of scene context generation. Therefore, we will focus our work in the problem of image regions labeling to classify every pixel of a given image into one of several predefined classes. Our proposal begins with a top-down control to find the core of objects, which allow us to update the learned models. Moreover, they become the starting seeds for the growing of a set of concurrent active regions which, considering the own region model as well as region and boundary information, obtain an accurate recognition of known regions. Next, a general segmentation extracts the unknown regions by a bottom-up strategy. Finally, a last stage exploits the contextual information to classify initially unknown segmented objects. The result is both a segmentation of the image and a recognition of each segment as a given object class or as an unknown segmented object. Experimental results on a wide set of outdoor scene images are shown to evaluate and compare our proposal.

As an ongoing effort to develop a computer aided system for the detection of masses on mammograms, we propose on this work a new model-based segmentation algorithm. The algorithm is based on a template matching scheme by using the mutual information approach in the similarity metric. Thus, the system will be able to determine if it exists a true mass on the studying image. The proposal was developed and evaluated using a database of 120 mammograms, 40 mammograms with confirmed masses and 80 normal ones. CAD performance was assessed using Receiver Operating Characteristics (ROC) and Free Receiver Operating Characteristics (FROC) analysis. The results prove the validity of the proposed method.

This paper describes a feature selection method based on the quadratic mutual information. We describe the needed formulation to estimate the mutual information from the data. This paper is motivated for the high time cost of the training process using the classical boosting algorithms. This method allows to reuse part of the training time used in the first training process to speed up posterior training to update the detectors in front of samples changes.

The literal performance of the symbols contained in a traditional score is not enough to produce expressive music. Human interpreters use musical knowledge that is not explicitly represented in it. This paper presents a knowledge-based approach to introduce expressiveness to the performance of a score by calculating dynamics and tempo envelopes of the piece, combining implicit musical knowledge with the explicit one contained in the score.

Diagnosis, treatment and prognosis are three of the most frequent labours of physician in health care institutions. In decision making, these activities can be tackled from two approaches: decision and planning. Decision structures are designed to help physician in their task of taking atemporal decisions. Planning structures are designed to guide physicians in the time-dependant complex medical procedures. The goal of this article is to present a model that integrates several learning tools to develop decision and planning structures in the medical domain, specifically in the support of decision making.