Catálogo de publicaciones - libros

Modern control systems and other monitoring systems require the acquisition of values of most of the parameters involved in the process. Examples of processes are industrial procedures or medical treatments or financial forecasts. However, sometimes some parameters are inaccessible through the use of traditional instrumentation. One example is the blades temperature in a gas turbine during operation. Other parameters require costly instrumentation difficult to install, operate and calibrate. For example, the contaminant emissions of power plant chimney. One solution of this problem is the use of analytical estimation of the parameter using complex differential equations. However, these models sometimes are very difficult to obtain and to maintain according the changes in the processes. Other solution is to borrow an instrument and measure a data set with the value of the difficult variable and its related variables at all the operation range. Then, use an automatic learning algorithm that allows inferring the difficult measure, given the related variables. This paper presents the use of Bayesian networks that represents the probabilistic relations of all the variables in a process, in the design of a virtual sensor. Experiments are presented with the temperature sensors of a gas turbine.

Markov decision processes (MDPs) have developed as a standard for representing uncertainty in decision-theoretic planning. However, MDPs require an explicit representation of the state space and the probabilistic transition model which, in continuous or hybrid continuous-discrete domains, are not always easy to define. Even when this representation is available, the size of the state space and the number of state variables to consider in the transition function may be such that the resulting MDP cannot be solved using traditional techniques. In this paper a reward-based abstraction for solving hybrid MDPs is presented. In the proposed method, we gather information about the rewards and the dynamics of the system by exploring the environment. This information is used to build a decision tree (C4.5) representing a small set of abstract states with equivalent rewards, and then is used to learn a probabilistic transition function using a Bayesian networks learning algorithm (K2). The system output is a problem specification ready for its solution with traditional dynamic programming algorithms. We have tested our abstract MDP model approximation in real-world problem domains. We present the results in terms of the models learned and their solutions for different configurations showing that our approach produces fast solutions with satisfying policies.

Learning and making decisions in a complex uncertain multiagent environment like RoboCup Soccer Simulation 3D is a non-trivial task. In this paper, a probabilistic approach to handle such uncertainty in RoboCup 3D is proposed, specifically a Naive Bayes classifier. Although its conditional independence assumption is not always accomplished, it has proved to be successful in a whole range of applications. Typically, the Naive Bayes model assumes discrete attributes, but in RoboCup 3D the attributes are continuous. In literature, Naive Bayes has been adapted to handle continuous attributes mainly using Gaussian distributions or discretizing the domain, both of which present certain disadvantages. In the former, the probability density of attributes is not always well-fitted by a normal distribution. In the latter, there can be loss of information. Instead of discretizing, the use of a Fuzzy Naive Bayes classifier is proposed in which attributes do not take a single value, but a set of values with a certain membership degree. Gaussian and Fuzzy Naive Bayes classifiers are implemented for the pass evaluation skill of 3D agents. The classifiers are trained with different number of training examples and different number of attributes. Each generated classifier is tested in a scenario with three teammates and four opponents. Additionally, Gaussian and Fuzzy approaches are compared versus a random pass selector. Finally, it is shown that the Fuzzy Naive Bayes approach offers very promising results in the RoboCup 3D domain.

This paper presents the preliminary results of the Student Model based on beliefs of Self-Efficacy aiming to improve the effectiveness of Intelligent Tutoring Systems. The Self-efficacy construct means the student’s belief on his own capacity of performing a task. This belief affects his behavior, motivation, affectivity and the choices he makes. We design an e-Learning System, called InteliWeb, this environment is composed by the Self-Efficacy Mediator Agent and offers instruction material on Biological sciences. We use fuzzy theory for dealing with uncertainty in the assessment of the students and the incomplete knowledge about his Self-Efficacy.

A bifurcation occurs in a dynamic system when the structure of the system itself and therefore also its qualitative behavior change as a result of changes in one of the system’s parameters. In most cases, an infinitesimal change in one of the parameters make the dynamic system exhibit dramatic changes. In this paper, we present a framework (QRBD) for performing qualitative analysis of dynamic systems exhibiting bifurcations. QRBD performs a simulation of the system with bifurcations, in the presence of perturbations, producing accounts for all events in the system, given a qualitative description of the changes it undergoes. In such a sequence of events, we include catastrophic changes due to perturbations and bifurcations, and hysteresis. QRBD currently works with first-order systems with only one varying parameter. We propose the qualitative representations and algorithm that enable us to reason about the changes a dynamic system undergoes when exhibiting bifurcations, in the presence of perturbations.

In this paper, to conquer the scalability issue of evolvable hardware (EHW), we introduce a novel system-decomposition-strategy which realizes training set partition in the intrinsic evolution of a non-truth table based 32 characters classification system. The new method is expected to improve the convergence speed of the proposed evolvable system by compressing fitness value evaluation period which is often the most time-consuming part in an evolutionary algorithm (EA) run and reducing computational complexity of EA. By evolving target characters classification system in a complete FPGA-based experiment platform, this research investigates the influence of introducing partitioning training set technique to non-truth table based circuit evolution. The experimental results conclude that it is possible to evolve characters classification systems larger and faster than those evolved earlier, by employing our proposed scheme.

We propose a new perspective for solving systems of nonlinear equations by viewing them as a multiobjective optimization problem where every equation represents an objective function whose goal is to minimize the difference between the right- and left-hand side of the corresponding equation of the system. An evolutionary computation technique is suggested to solve the problem obtained by transforming the system into a multiobjective optimization problem. Results obtained are compared with some of the well-established techniques used for solving nonlinear equation systems.

This paper presents a new multi-objective evolutionary algorithm which consists of a hybrid between a particle swarm optimization (PSO) approach and scatter search. The main idea of the approach is to combine the high convergence rate of the particle swarm optimization algorithm with a local search approach based on scatter search. We propose a new leader selection scheme for PSO, which aims to accelerate convergence. Upon applying PSO, scatter search acts as a local search scheme, improving the spread of the nondominated solutions found so far. Thus, the hybrid constitutes an efficient multi-objective evolutionary algorithm, which can produce reasonably good approximations of the Pareto fronts of multi-objective problems of high dimensionality, while only performing 4,000 fitness function evaluations. Our proposed approach is validated using ten standard test functions commonly adopted in the specialized literature. Our results are compared with respect to a multi-objective evolutionary algorithm that is representative of the state-of-the-art in the area: the NSGA-II.

This work addresses an important problem in Feedforward Neural Networks (FNN) training, i.e. finding the pseudo-global minimum of the cost function, assuring good generalization properties to the trained architecture. Firstly, pseudo-global optimization is achieved by employing a combined parametric updating algorithm which is supported by the transformation of network parameters into interval numbers. It solves the network weight initialization problem, performing an exhaustive search for minimums by means of Interval Arithmetic (IA). Then, the global minimum is obtained once the search has been limited to the region of convergence (ROC). IA allows representing variables and parameters as compact-closed sets, then, a training procedure using interval weights can be done. The methodology developed is exemplified by an approximation of a known non-linear function in last section.

A gradient-based algorithm for ensemble weights modification is presented and applied on the regression tasks. Simulation results show that this method can produce an estimator ensemble with better generalization than those of bagging and single neural network. The method can not only have a similar function to GASEN of selecting many subnets from all trained networks, but also be of better performance than GASEN, bagging and best individual of regressive estimators.