Catálogo de publicaciones - libros

Kernel based methods (such as -nearest neighbors classifiers) for AI tasks translate the classification problem into a proximity search problem, in a space that is usually very high dimensional. Unfortunately, no proximity search algorithm does well in high dimensions. An alternative to overcome this problem is the use of approximate and probabilistic algorithms, which trade time for accuracy.

In this paper we present a new probabilistic proximity search algorithm. Its main idea is to order a set of samples based on their distance to each element. It turns out that the closeness between the order produced by an element and that produced by the query is an excellent predictor of the relevance of the element to answer the query.

The performance of our method is unparalleled. For example, for a 128-dimensional dataset, it is enough to review 10% of the database to obtain 90% of the answers, and to review less than 1% to get 80% of the correct answers. The result is more impressive if we realize that a full 128-dimensional dataset may span thousands of dimensions of clustered data. Furthermore, the concept of proximity preserving order opens a totally new approach for both exact and approximated proximity searching.

We present a neurobiologically motivated model for an agent which generates a representation of its spacial environment by an active exploration. Our main objectives is the introduction of an action-selection mechanism based on the principle of self-reinforcement learning. We introduce the action-selection mechanism under the constraint that the agent receives only information an animal could receive too. Hence, we have to avoid all supervised learning methods which require a teacher. To solve this problem, we define a self-reinforcement signal as qualitative comparison between predicted an perceived stimulus of the agent. The self-reinforcement signal is used to construct internally a self-punishment function and the agent chooses its actions to minimize this function during learning. As a result it turns out that an active action-selection mechanism can improve the performance significantly if the problem to be learned becomes more difficult.

The accuracy of the rules produced by a concept learning system can be hindered by the presence of errors in the data. Although these errors are most commonly attributed to random noise, there also exist “ill-defined” attributes that are too general or too specific that can produce systematic classification errors. We present a computer program called Newton which uses the fact that ill-defined attributes create an ordered error pattern among the instances to compute hypotheses explaining the classification errors of a concept in terms of too general or too specific attributes. Extensive empirical testing shows that Newton identifies such attributes with a prediction rate over 95%.

Reduced Support Vector Machines (RSVM) was proposed as the alternate of standard support vector machines (SVM) in order to resolve the difficulty in the learning of nonlinear SVM for large data set problems. RSVM preselects a subset as support vectors and solves a smaller optimization problem, and it performs well with remarkable efficiency on training of SVM for large problem. All the training points of the subset will be support vectors, and more training points are selected into this subset results in high possibility to obtain RSVM with better generalization ability. So we first obtain the RSVM with more support vectors, and selects out training examples near classification hyper plane. Then only these training examples are used as training set to obtain a standard SVM with less support vector than that of RSVM. Computational results show that standard SVMs on the basis of RSVM have much less support vectors and perform equal generalization ability to that of RSVM.

Learning systems need to face several problems: incrementality, tracking concept drift, robustness to noise and recurring contexts in order to operate continuously. A method for on-line induction of decision trees motivated by the above requirements is presented. It uses the following strategy: creating a delayed window in every node for applying forgetting mechanisms; automatic modification of the delayed window; and constructive induction for identifying recurring contexts. The default configuration of the proposed approach has shown to be globally efficient, reactive, robust and problem-independent, which is suitable for problems with unknown dynamics. Notable results have been obtained when noise and concept drift are present.

When dealing with multi-class classification tasks, a popular and applicable way is to decompose the original problem into a set of binary subproblems. The most well-known decomposition strategy is and the corresponding widely-used method to recombine the outputs of all binary classifiers is (PWC). However PWC has an intrinsic shortcoming; many meaningless partial classification results contribute to the global prediction result. Moreira and Mayoraz suggested to tackle this problem by using [4]. Though much better performance was obtained, their algorithm is simple and has some disadvantages. In this paper, we propose a novel algorithm which works in two steps: First the original are converted into a new set of , then is employed to construct the global posterior probabilities. Employing support vector machines as binary classifiers, we perform investigation on several benchmark datasets. Experimental results show that our algorithm is effective and efficient.

The kernel function of support vector machine (SVM) is an important factor for the learning result of SVM. Based on the wavelet decomposition and conditions of the support vector kernel function, Littlewood-Paley wavelet kernel function for SVM is proposed. This function is a kind of orthonormal function, and it can simulate almost any curve in quadratic continuous integral space, thus it enhances the generalization ability of the SVM. According to the wavelet kernel function and the regularization theory, Least squares Littlewood-Paley wavelet support vector machine (LS-LPWSVM) is proposed to simplify the process of LPWSVM. The LS-LPWSVM is then applied to the regression analysis and classifying. Experiment results show that the precision is improved by LS-LPWSVM, compared with LS-SVM whose kernel function is Gauss function.

This paper proposes a state transition (ST) model as a classifier and its generalization by the minimization. Different from previous works using statistical methods, tree-based classifiers and neural networks, we use a ST model which determines classes of strings. Though an initial ST model only accepts given strings, the minimum ST model can accepts various strings by the generalization. We use a minimization algorithm by Mixed-Integer Linear Programming (MILP) approach. The MILP approach guarantees a minimum solution. Experiment was done for the classification of pseudo-strings. Experimental results showed that the reduction ratio from an initial ST model to the minimal ST model becomes small, as the number of examples increases. However, a current MILP solver was not feasible for large scale ST models in our formalization.

Compilers are nowadays fundamental tools for the development of any kind of application. However, their task gets increasingly difficult due to the constant increase in the complexity of modern computer architecture, as well as to the increased requirements imposed upon programming languages by the great diversity of applications handled at present. In the compilation process several optimization problems must be solved, some of them belonging to the NP-Hard class. The quality of the solution found for these problems has direct impact over the quality of the generated object code. To solve them, compilers do it locally through naive heuristics which might consequently lead to solutions that are far from optimal. Knowing that metaheuristics methods have recently been used massively and successfully to solve combinatorial optimization problems, similar performance in the problems found in the compilation process can be expected beforehand. Following this line of reasoning, such problems are presented in this paper and the potential use of metaheuristics techniques to find their solutions is analyzed. A review is also made of the work that has been done in this field, and finally a proposal is made of the road that this development should follow.

In recent years, the importance of the construction of fuzzy models from measured data has increased. Nevertheless, the complexity of real-life process is characterized by nonlinear and non-stationary dynamics, leaving so much classical identification techniques out of choice. In this paper, we present a comparison of Support Vector Machines (SVMs) for density estimation (SVDE) and for regression (SVR), versus traditional techniques as Fuzzy C-means and Gustafson-Kessel (for clustering) and Least Mean Squares (for regression), in order to find the parameters of Takagi-Sugeno (TS) fuzzy models. We show the properties of the identification procedure in a waste-water treatment database.