Catálogo de publicaciones - libros

We propose a speech-act analysis method for Korean dialogue using Support Vector Machines (SVM). We use a lexical word, its part of speech (POS) tags, and bigrams of POS tags as utterance feature and the contexts of the previous utterance as context feature. We select informative features by statistic. After training SVMs with the selected features, SVM classifiers determine the speech-act of each utterance. In experiment, we acquired overall 90.5% of accuracy with dialogue corpus for hotel reservation domain.

This paper presents a new algorithm to deal with nominal attributes in Support Vector Classification by modifying the most popular approach. For a nominal attribute with states, we translate it into points in – 1 dimensional space with flexible and adjustable position. Their final position is decided by minimizing the Leave-one-out error. This strategy overcomes the shortcoming in the most popular approach which assume that any two different attribute values have the same degree of dissimilarities. Preliminary experiments also show the superiority of our new algorithm.

This paper presents the comparableness of SVM method to artificial neural networks in the outlier detection problem of high dimensions. Experiments performed on real dataset show that the performance of this method is mostly superior to that of artificial neural networks. The proposed method, SVM served to exemplify that kernel-based learning algorithms can be employed as an efficient method for evaluating the revegetation potentiality of abandoned lands from coal mining activities.

T-cell activation is a pivotal process in immune response. A precondition for this activation is the recognition of antigenic epitopes by T-cell receptors. This recognition is antigen-specific. Therefore, identifying the pattern of a MHC restricted T-cell epitopes is of great importance for immunotherapies and vaccine design. In this paper, a new kernel is proposed to use together with support vector machine for the direct prediction of T-cell epitope. The experiment was carried on an MHC type I restricted T-cell clone LAU203-1.5. The results suggest that this approach is efficient and promising.

A Soft-magnetic ring cores (SMRC) inspection method using radial basis function support vector machine (RBFSVM) was developed. To gain the effective edge character of the SMRC, a sequence of image edge detection algorithms was developed. After edge was detected, feature vector was extracted. Subsequently, principal component analysis (PCA) is applied to reduce the dimension of the feature vector. Finally, RBFSVM is used for classification of SMRC, whose best accuracy in experiments is 97%.

The stochastic flow shop scheduling with uncertain processing time is a typical NP-hard combinatorial optimization problem and represents an important area in production scheduling, which is difficult because of inaccurate objective estimation, huge search space, and multiple local minima. As a novel evolutionary technique, particle swarm optimization (PSO) has gained much attention and wide applications for both function and combinatorial problems, but there is no research on PSO for stochastic scheduling cases. In this paper, a class of PSO approach with simulated annealing (SA) and hypothesis test (HT), namely PSOSAHT is proposed for stochastic flow shop scheduling with uncertain processing time with respect to the makespan criterion (i.e. minimizing the maximum completion time). Simulation results demonstrate the feasibility, effectiveness and robustness of the proposed hybrid algorithm. Meanwhile, the effects of noise magnitude and number of evaluation on searching performances are also investigated.

Precise prediction of algal booms is beneficial to fisheries and environmental management since it enables the fish farmers to gain more ample time to take appropriate precautionary measures. Since a variety of existing water quality models involve exogenous input and different assumptions, artificial neural networks have the potential to be a cost-effective solution. However, in order to accomplish this goal successfully, usual problems and drawbacks in the training with gradient algorithms, i.e., slow convergence and easy entrapment in a local minimum, should be overcome first. This paper presents the application of a particle swarm optimization model for training perceptrons to forecast real-time algal bloom dynamics in Tolo Harbour of Hong Kong, with different lead times on the basis of several input hydrodynamic and/or water quality variables. It is shown that, when compared with the benchmark backward propagation algorithm, its results can be attained both more accurately and speedily.

The particle swarm optimization algorithm presents a new way for finding an optimal solution of complex optimization problems, where each particle represents a solution to the problem. In this paper a modified particle swarm optimization algorithm is applied to the optimization of the antenna array. Adding an item of integral control and the contractive factor in the modified algorithm can improve its global search ability. Simulation results show that the optimal pattern of the antenna array is able to approach the desired pattern. The results also demonstrate that the modified algorithm is superior to the original algorithm and the nonlinear least-square method.

To improve the linearly varying inertia weigh particle swarm optimization method (LPSO), a new concept of Crowd Avoidance is introduced in this paper. In this newly developed LPSO (CA-LPSO), particles can avoid entering into a crowded space while collaborate with other particles searching for optimum. Four well-known benchmarks were used to evaluate the performance of CA-LPSO in comparison with LPSO. The simulation results show that, although CA-LPSO falls behind LPSO when optimizing simple unimodal problems, it is more effective than LPSO for most complex functions. The crowd avoidance strategy enables the particles to explore more areas in the search space and thus decreases the chance of premature convergence.

This paper proposes a Binary Ant System (BAS), a binary version of the hyper-cube frame for Ant Colony Optimization applied to unconstrained function optimization problem. In BAS, artificial ants construct the solutions by selecting either 0 or 1 at every bit stochastically biased by the pheromone level. For ease of implementation, the pheromone value is designed specially to directly represent the probability of selection. Principal settings of the parameters are analyzed and some methods to escape local optima, such as local search and pheromone re-initialization are incorporated into the proposed algorithm. Experimental results show that the BAS is able to find very good results for the unconstrained function optimization problems of different characteristics.