Catálogo de publicaciones - libros

A relaxed setting for Feature Selection is known as Feature Ranking in Machine Learning. The aim is to establish an order between the attributes that describe the entries of a learning task according to their utility. In this paper, we propose a method to establish these orders using Preference Learning by means of Support Vector Machines (SVM). We include an exhaustive experimental study that investigates the virtues and limitations of the method and discusses, simultaneously, the design options that we have adopted. The conclusion is that our method is very competitive, specially when it searchs for a ranking limiting the number of combinations of attributes explored; this supports that the method presented here could be successfully used in large data sets.

Real-time search methods allow an agent to move in unknown environments. We provide two enhancements to the real-time search algorithm HLRTA*(). First, we give a better way to perform bounded propagation, generating the HLRTA*() algorithm. Second, we consider the option of doing more than one action per planning step, by analyzing the quality of the heuristic found during lookahead, producing the HLRTA*(,) algorithm. We provide experimental evidence of the benefits of both algorithms, with respect to other real-time algorithms on existing benchmarks.

In this work a simulated wastewater treatment plant is controlled with a sliding mode control carried out with softcomputing techniques. The controller has two modules: the first one performs the plant control when its dynamics lies inside an optimal working region and is carried out by a neural network trained to reproduce the behavior of the technician who controls an actual plant, while the second one drives the system dynamics towards that region when it works outside it and is carried out by a corrective function whose parameters have been adjusted with a genetic algorithm. The controller so defined performs satisfactory even when extreme inputs are presented to the model.

Many recent successful planners use domain-independent heuristics to speed up the search for a valid plan. An orthogonal approach to accelerating search is to identify and remove redundant operators. We present a domain-independent algorithm for efficiently pruning redundant operators prior to search. The algorithm operates in the domain transition graphs of multi-valued state variables, so its complexity is polynomial in the size of the state variable domains. We prove that redundant operators can always be replaced in a valid plan with other operators. Experimental results in standard planning domains demonstrate that our algorithm can reduce the number of operators as well as speed up search.

We introduce a non-admissible heuristic for planning with action costs, called the , that combines the benefits of the used in the HSP planner and the used in FF. The set-additive heuristic is defined mathematically and handles non-uniform action costs like the additive heuristic , and yet like FF’s heuristic , it encodes the cost of a specific and is therefore compatible with FF’s helpful action pruning and its effective enforced hill climbing search. The definition of the set-additive heuristic is obtained from the definition of the additive heuristic, but rather than propagating the value of the best supports for a precondition or goal, it propagates the supports themselves, which are then combined by set-union rather than by addition. We report then empirical results on a planner that we call FF() that is like FF except that the relaxed plan is extracted from the set-additive heuristic. The results show that FF() adds only a slight time overhead over FF but results in much better plans when action costs are not uniform.

There exist various narrative systems, focused on different parts of the complex process of story generation. Some of them are oriented to , and some to , with different properties and characteristics. In this paper we propose a system based on BDI agents that generates stories (creating content, performing and simple ) with narrative parts and dialogs. The content for the story is generated in a multiagent social simulation system, and the is based on rules and a state space search algorithm based on the system representation of the reader’s perception of the story.

Models are used in science and engineering for experimentation, analysis, model-based diagnosis, design and planning/sheduling applications. Many of these models are overconstrained Numeric Constraint Satisfaction Problems (), where the numeric constraints could have linear or polynomial relations. In practical scenarios, it is very useful to know which parts of the overconstrained NCSP instances cause the unsolvability.

Although there are algorithms to find all optimal solutions for this problem, they are computationally expensive, and hence may not be applicable to large and real-world problems. Our objective is to improve the performance of these algorithms for numeric domains using structural analysis. We provide experimental results showing that the use of the different strategies proposed leads to a substantially improved performance and it facilitates the application of solving larger and more realistic problems.

In this paper we present several interest points detectors and we analyze their suitability when used as landmark extractors for vision-based simultaneous localization and mapping (vSLAM). For this purpose, we evaluate the detectors according to their repeatability under changes in viewpoint and scale. These are the desired requirements for visual landmarks. Several experiments were carried out using sequence of images captured with high precision. The sequences represent planar objects as well as 3D scenes.

Transformation Based Learning (TBL) is an intensively Machine Learning algorithm frequently used in Natural Language Processing. TBL uses rule templates to identify error-correcting patterns. A critical requirement in TBL is the availability of a problem domain expert to build these rule templates. In this work, we propose an evolutionary approach based on Genetic Algorithms to automatically implement the template selection process. We show some empirical evidence that our approach provides template sets with almost the same quality as human built templates.

We provide upper bounds for the Vapnik-Chervonenkis dimension of classes of subsets of that can be recognized by computer programs built from , (like -root extraction and, more generally, operations defined by algebraic series of fractional powers), and . This includes certain classes of GP-trees considered in Genetic Programming for symbolic regression and bi-classification. As a consequence we show explicit quantitative properties that can help to design the fitness function of a GP learning machine.