Catálogo de publicaciones - libros

In this work, we develop a multi-exchange heuristic based on an estimation improvement graph embedded in a simulated annealing to solve a problem arising in plant location planning where tariff exemptions apply. The method is shown to be effective in experiments since it provides good solutions for problems of realistic size. It is superior to CPLEX in terms of time, and is able to provide solutions for large test problems.

The car-sequencing problem arises from the manufacture of cars on an assembly line (based on [1]). A number of cars are to be produced; they are not identical, because different options are available as variants on the basic model. The assembly line has different stations (designed to handle at most a certain percentage of the cars passing along the assembly line) which install the various options. Furthermore, the cars requiring a certain option must not be bunched together, otherwise the station will not be able to cope. Consequently, the cars must be arranged in a sequence so that the capacity of each station is never exceeded. The solving methods for constraint satisfaction problems (CSPs) [2], [3], [4] represent good alternatives for certain instances of the problem. Constraint programming tools [5], [6] use a search algorithm based on Forward Checking (FC) [7] to solve CSPs, with different variable or value ordering heuristics. In this article, we undertake an experimental study for the instances of the car-sequencing problem in CSPLib, encoded as an n-ary CSP using an implementation with constraints of fixed arity 5. By applying value ordering heuristics based on fail-first principle, a great number of these instances can be solved in little time. Moreover, the parallel solving using a shared memory model based on OpenMP makes it possible to increase the number of solved problems.

With mixed feature data, problems are induced in modeling the gating network of normalized Gaussian (NG) networks as the assumption of multivariate Gaussian becomes invalid. In this paper, we propose an independence model to handle mixed feature data within the framework of NG networks. The method is illustrated using a real example of breast cancer data.

Text information processing depends critically on the proper text representation. A common and naïve way of representing a document is a bag of its component words [1], but the semantic relations between words are ignored, such as synonymy and hypernymy-hyponymy between nouns. This paper presents a model for representing a document in terms of the synonymy sets (synsets) in WordNet [2]. The synsets stand for concepts corresponding to the words of the document. The Vector Space Model describes a document as orthogonal term vectors. We replace terms with concepts to build Concept Vector Space Model (CVSM) for the training set. Our experiments on the Reuters Corpus Volume I (RCV1) dataset have shown that the result is satisfactory.

Bayesian learning techniques for MLPs are applied to the problem of forecasting the direction of change in daily close values of the Australian All Ordinaries Index. Predictions made over a 13 year out-of-sample period were tested against two null hypotheses—the null hypothesis of a mean accuracy of 0.5 (which is the expected accuracy if prices follow a random walk), and a null hypothesis which takes into account non-stationarity in the prices series. Results show that both null hypotheses can be rejected at the 0.005 level, but much more confidently in the case of the Bayesian approach as compared to an approach using conventional gradient descent based weight optimization.

Recently, there are several nonlinear dimensionality reduction algorithms that can discover the low-dimensional coordinates on a manifold based on training samples, such as ISOMAP, LLE, Laplacian eigenmaps. However, most of these algorithms work in batch mode. In this paper, we presented an incremental nonlinear dimensionality reduction algorithm to efficiently map new samples into the embedded space. The method permits one to select some landmark points and to only preserve geodesic distances between new data and landmark points. Self-organizing map algorithm is used to choose landmark points. Experiments demonstrate that the proposed algorithm is effective.

To minimize the outliers’ effects, in this paper, a new speaker identification scheme based on the t- distribution mixture model is proposed. Since the t- distribution provides a longer and heavier tailed alternative to the Gaussian distribution, the mixture model with multivariate t- distribution is expected to show more robust results than the Gaussian mixture model(GMM) in the cases where outliers exist. In experiments, we compared the performance of the proposed scheme with that of using the conventional GMM to show its robustness.

Inducing sequential patterns from time series data is an important data mining problem. While most of the current methods are generating sequential patterns within a single attribute, this paper proposes a new method, using Hellinger entropy measure, for generating multi-dimensional sequential patterns. A number of theorems are proposed to reduce the computational complexity of the proposed method.

This paper deals with relationship between BP learning for neural networks and numerical algorithm of differential equations. It is proposed that the iteration formula of BP algorithm is equivalent to Euler method of differential dynamic system under certain conditions, and the asymptotic solutions of the two formulas are consistent. It is also proved in theoretic that asymptotic solutions given by BP algorithm are equivalent to that computed by any numerical method for differential dynamic systems under certain conditions. Also, an example to train the BP network by modified numerical method is presented.

An effective feature selection scheme is proposed, which utilizes the combination of wrapper and filter: ant colony optimization (ACO) and mutual information (MI). By examining the modeling based on SVMs at the Australian Bureau of Meteorology, the simulation of three different methods of feature selection shows that the proposed method can reduce the dimensionality of inputs, speed up the training of the network and get better performance.