Catálogo de publicaciones - libros

The main focus in algorithms has been on efficient construction of summary structures for data streams. This paper introduces the problem of construction of summary structures from sliding windows over data streams, and presents a new sampling-based summary structure and new techniques for its fast incremental maintenance. When a new data item v _ i arrives, a key k _ i is calculated and a random number X _ i is generated. The key k _ i is used to determine if v _ i will be selected to enter the sample, and X _ i is used to determine how many data items will be skipped over. The experiments show that the new algorithm is effective and efficient for construction of summary structures from sliding windows over data streams.

Machine learning approaches were employed for malignant breast tumour diagnosis and evaluation of the prognostic risk of recrudescence and metastasis by using age and ten cellular attributes of Fine Needle Aspirate of Breast (FNAB) and gene microarrays data of the breast cancer patient respectively. Feature ranking method was introduced to explore the salient elements for cancer identification and simultaneous improve the classification accuracy. In this paper, Support Vector Machine (SVM), K-Nearest Neighbor (K-NN) and Probabilistic Neural Network (PNN) combined with Signal-to-Noise Ratio (SNR) for feature ranking and filtering were applied to distinguish between the benign and malignant tumours of breast and evaluate the prognostic risk of recrudescence and metastasis. The results reveal that feature ranking method SNR can effectively pick out the informative and important features, which had significance for clinical assistant diagnosis and is useful for improving the performance of evaluation. The best overall accuracy for breast cancer diagnosis and evaluating the prognostic risk of recrudescence and metastasis achieved 96.24% and 88.81% respectively, by using SVM-Sigmoid and SVM-RBF combined with SNR under 5-fold cross validation. This study suggests that SVM may be further developed to be a practical methodology for clinical assistant differentiating between benign and malignant tumours and possible to help the inexperienced physicians avoid misdiagnosis. It also has benefit to the cured patients who are predicted as recrudescence and metastasis pay more attention to their diseases, and then reduce the mortality rate of breast cancer.

The high-resolution image reconstruction methods utilizing a super-resolution (SR) technique better can obtain higher quality image than the conventional interpolation methods when the input images are well registered onto a common high-resolution grid. Therefore, low-resolution input images should be carefully selected to take the minimized registration error. In this paper, we propose the input image evaluation algorithm to select the suitable input image with low registration error, by using the statistical feature of the motion-compensated low-resolution images. Maximum motion compensation error (MMCE) is estimated from the high-resolution image observation model and is used to evaluate the suitability of the low-resolution input image candidates. The low-resolution input image is selected when its motion compensation error (MCE) is in the range 0 < MCE < MMCE. The reference input image (RII) is selected by counting the number of selected low-resolution input images (SLRII) from all low-resolution input images (LRII). A good high-resolution image is reconstructed efficiently from a optimal reference input image (ORII) and the selected low-resolution input images (SLRII) by using the Hardie’s SR reconstruction algorithm.

A new chaotic neural network model called Morlet Wavelet chaotic neural network with chaotic noise was presented, the chaotic noise was produced by the Logistic map multiplied by an exponentially decreased variable factor in order to verify the ability of anti-disturbance, and the transiently chaotic mechanism was introduced by the attenuation of the self-feedback connection weight. In this paper, first, the figures of the reversed bifurcation and the maximal Lyapunov exponents of single neural unit were given. Second, the new model was applied to solve function optimizations. Finally, 10-city traveling salesman problem was given and simultaneously the effects of the proportional value of the non-monotonous-function coefficient to the monotonous-function coefficient in the model were discussed. As seen from the simulation results, the new model is more powerful than common chaotic neural network and has some certain ability of anti-disturbance.

In this paper, we propose an approach to offline definition extraction using machine learning. We introduced a new framework of knowledge-based question answering (QA) system. In this framework, the answers are extracted and saved in answer base beforehand. For adapting to large scale application, the answer extraction should be independent with question. We call this task as offline answer extraction. We propose an approach to offline definition extraction using machine learning. We manually label the definition in documents and take them as training data to train the definition extraction model. We employ three classification models: Decision tree, Naïve Bayesian and SVM. The experiment results show that SVM has best performance in definition extraction. Our approach outperforms the baseline in experiment. The experiment results indicate that our approach is effective.

This paper presents a shortest-path gradients setup algorithm for directed diffusion in wireless sensor networks. Based on the research of one basic gradients setup algorithm implemented in ns2.29 for directed diffusion protocol, this paper analyzes the algorithm’s key point: intermediate node’s interest re-send adopts random time delay and happens only one time in one interest propagation round. Then it points out that this basic algorithm can establish some inefficient gradients which construct longer paths. This paper improves this gradients setup algorithm utilizing interest packet’s forwarded times. The improved algorithm greatly decreases the probability to build “parallel gradients” and “reverse gradients”, and almost set up a best gradient topology which can construct multiple shortest paths from source node to sink node. The simulation results show, using the shortest-path gradients setup algorithm, data packets will be delivered along shorter paths, and directed diffusion for wireless sensor networks is more energy efficient.

This paper presented a novel passivity-based sliding-mode speed control system for vector-controlled permanent magnet synchronous motor(PMSM) drive. According to the strictly passivity of PMSM system, the vector control principle was derived. The sliding-mode control method was adopted here to design the speed controller, the stability of the controller was approved by the passivity theory, the parameters of the controller was chosen appropriately based on the Lyapunov stability theory. The fuzzy control method was adopted here to design a fuzzy switching function to replace the sliding mode signum function, and the chattering phenomenon was restrained greatly. The Simulink simulations and dSPACE experiments show the good performance of the control system, and that this system is robust to the parameter variations and load break.

Acoustic reverberation signal generated by experimental explosive source is analyzed by dynamic methods. Two dynamic multi-step prediction models, including adding weight one-order local region model and adaptive prediction model based on Volterra series have been introduced to model the oceanic reverberation signal in the reconstructed phase space. The results indicate that the reverberation time series can be predicted in short term with a minor prediction error by these two methods. An elementary conclusion can be obtained that the dynamic model is more suitable for modeling the oceanic reverberation than the classical random AR model.

In most contexts, learning is essential for the long-term autonomy of an agent. We describes here some essential and fundamental learning mechanisms implemented in a cognitive autonomous agent, CTS (Conscious Tutoring System), we suggest a model that maintains “conscious-” and at the same time "unconscious-" learning as means to increase the agent’s autonomy in unknown or changing environments, and a way to improve its fitness. The two mentioned mechanisms occur in parallel in CTS and are inspired by phenomena believed to exist in humans.

Based on the fusion of the immune evolutionary algorithm and immune feedback mechanism for the conventional PID control technique, a new immune PID controller with the immune tuning is proposed in this paper. Computer simulation results demonstrate that the performance of our immune controller is better than the regular PID controller. It is an appropriate candidate for the industrial production processes.