Catálogo de publicaciones - libros

The increasing reliance on Computational Intelligence applications to predict stock market positions have resulted in numerous researches in financial forecasting and trading trend identifications. Stock market price prediction applications are required to be adaptive to new incoming data as well as have fast learning capabilities due to the volatility nature of market movements. This paper analyses stock market price prediction based on a Fuzzy Cerebellar Model Articulation Controller – Bayesian Ying Yang (FCMAC-BYY) neural network. The model is motivated from the Chinese ancient Ying-Yang philosophy which states that everything in the universe can be viewed as a product of a constant conflict between opposites, Ying and Yang. A perfect status is reached if Ying and Yang achieves harmony. The analyzed experiment on a set of real stock market data (Singapore Airlines Ltd – SIA) in the Singapore Stock Exchange (SGX) and Ibex35 stock index shows the effectiveness of the FCMAC-BYY in the universal approximation and prediction.

Rough sets and neural networks are two common techniques applied to rule extraction from data table. Integrating the advantages of two approaches, this paper presents a Hybrid Rule Extraction Method (HREM) using rough sets and neural networks. In the HREM, the rule extraction is mainly done based on rough sets, while neural networks are only served as a tool to reduce the decision table and filter its noises when the final knowledge (rule sets) is generated from the reduced decision table by rough sets. Therefore, the HREM avoids the difficult of extracting rules from a trained neural network and possesses the robustness which the rough sets based approaches are lacking. The effectiveness of HREM is verified by comparing the experiment results with the approaches of traditional rough sets and neural networks.

Genetic algorithms (GAs) are widely used in the parameter training of Neural Network (NN). In this paper, we investigate GAs based on our proposed novel genetic representation to train the parameters of NN. A splicing/decomposable (S/D) binary encoding is designed based on some theoretical guidance and existing recommendations. Our theoretical and empirical investigations reveal that the S/D binary representation is more proper than other existing binary encodings for GAs’ searching. Moreover, a new genotypic distance on the S/D binary space is equivalent to the Euclidean distance on the real-valued space during GAs convergence. Therefore, GAs can reliably and predictably solve problems of bounded complexity and the methods depended on the Euclidean distance for solving different kinds of optimization problems can be directly used on the S/D binary space. This investigation demonstrates that GAs based our proposed binary representation can efficiently and effectively train the parameters of NN.

Adaptive training of a classifier is necessary when feature selection and sparse representation are considered. Previously, we proposed a kernel-based nonlinear classifier for simultaneous representation and discrimination of pattern features. Its batch training has a closed-form solution. In this paper we implement an adaptive training algorithm using an incremental learning procedure that exactly retains the generalization ability of batch training. It naturally yields a sparse representation. The feasibility of the presented methods is illustrated by experimental results on handwritten digit classification.

Grey-box neural models mix differential equations, which act as white boxes, and neural networks, used as black boxes. The purpose of the present work is to show the training of a grey-box model by means of indirect backpropagation and Levenberg-Marquardt in Matlab, extending the black box neural model in order to fit the discretized equations of the phenomenological model. The obtained grey-box model is tested as an estimator of a state variable of a biotechnological batch fermentation process on solid substrate, with good results.

We present a robust recursive least squares algorithm for multilayer feed-forward neural network training. So far, recursive least squares (RLS) has been successfully applied to training multilayer feed-forward neural networks. However, RLS method has a tendency to become diverse due to the instability in the recursive inversion procedure. In this paper, we propose a numerically robust recursive least square type algorithm using prewhitening. The proposed algorithm improves the performance of RLS in infinite numerical precision as well as in finite numerical precision. The computer simulation results in the various precision cases show that the proposed algorithm improves the numerical robustness of RLS training.

At present, most of the algorithms for learning Bayesian Networks (BNs) use EM algorithm to deal with incomplete data. They are of low efficiency because EM algorithm has to perform iterative process of probability reasoning to complete the incomplete data. In this paper we present an efficient BN learning algorithm, which use the combination of EMI method and a scoring function based on mutual information theory. The algorithm first uses EMI method to estimate, from incomplete data, probability distributions over local structures of BNs, then evaluates BN structures with the scoring function and searches for the best one. The detailed procedure of the algorithm is depicted in the paper. The experimental results on Asia and Alarm networks show that when achieving high accuracy, the algorithm is much more efficient than two EM based algorithms, SEM and EM-EA algorithms.

An optimization algorithm for dynamic Bayesian networks (DBN) based on Bayesian optimization algorithm (BOA) is developed for learning and constructing the DBN structure. In this paper, we first introduce some basic theories and concepts of probability model evolutionary algorithm. Then we describe, the basic mode for constructing DBN diagram and the mechanism of DBN structure learning based on BOA. The DBN structure learning based on BOA consists of two parts. The first part is to obtain the structure and parameters of DBN in terms of a good solution, and the second part is to produce new groups according to the obtained DBN structure. In this paper, the DBN learning is achieved by genetics algorithm based on a greedy mechanism. The DBN inference is performed by a forward-simulation algorithm. Simulation results are provided to demonstrate the effectiveness of the proposed algorithm.

A tapped delay neural network (TDNN) with an adaptive learning and pruning algorithm is proposed to predict the nonlinear time serial stock indexes. The TDNN is trained by the recursive least square (RLS) in which the learning-rate parameter can be chosen automatically. This results in the network converging fast. Subsequently the architecture of the trained neural network is optimized by utilizing pruning algorithm to reduce the computational complexity and enhance the network’s generalization. And then the optimized network is retrained so that it has optimum parameters. At last the test samples are predicted by the ultimate network. The simulation and comparison show that this optimized neuron network model can not only reduce the calculating complexity greatly, but also improve the prediction precision. In our simulation, the computational complexity is reduced to 0.0556 and mean square error of test samples reaches 8.7961×10.

The Eleman Neural Network has been widely used in various fields ranging from temporal version of the Exclusive-OR function to the discovery of syntactic categories in natural language date. However, one of the problems often associated with this type of network is the local minima problem which usually occurs in the process of the learning. To solve this problem, we have proposed an error function which can harmonize the update weights connected to the hidden layer and those connected to the output layer by adding one term to the conventional error function. It can avoid the local minima problem caused by this disharmony. We applied this method to the Boolean Series Prediction Questions problems to demonstrate its validity. The result shows that the proposed method can avoid the local minima problem and largely accelerate the speed of the convergence and get good results for the prediction tasks.