Catálogo de publicaciones - libros

We discuss a new design of group method of data handling (GMDH)-type neural network using evolutionary algorithm. The performances of the GMDH-type network depend strongly on the number of input variables and order of the polynomials to each node. They must be fixed by designer in advance before the architecture is constructed. So the trial and error method must go with heavy computation burden and low efficiency. To alleviate these problems we employed evolutionary algorithms. The order of the polynomial, the number of input variables, and the optimum input variables are encoded as a chromosome and fitness of each chromosome is computed. The appropriate information of each node are evolved accordingly and tuned gradually throughout the GA iterations. By the simulation results, we can show that the proposed networks have good performance.

Construction claims are normally affected by a large number of complex and interrelated factors. It is highly desirable for the parties to a dispute to know with some certainty how the case would be resolved if it were taken to court. The use of artificial neural networks can be a cost-effective technique to help to predict the outcome of construction claims, on the basis of characteristics of cases and the corresponding past court decisions. In this paper, a particle swarm optimization model is adopted to train perceptrons. The approach is demonstrated to be feasible and effective by predicting the outcome of construction claims in Hong Kong in the last 10 years. The results show faster and more accurate results than its counterparts of a benching back-propagation neural network and that the PSO-based network are able to give a successful prediction rate of up to 80%. With this, the parties would be more prudent in pursuing litigation and hence the number of disputes could be reduced significantly.

Intelligent and adaptive approach to model two links manipulator system with self-organizing radial basis function (RBF) network is presented in this paper. The self-organizing algorithm that enables the RBF neural network to be structured automatically and on-line is developed, and with this proposed scheme, the centers and widths of RBF neural network as well as the weights are to be adaptively determined. Based on the fact that a 3-layered RBF neural network has the capability that represents the nonlinear input-output map of any nonlinear function to a desired accuracy, the input output mapping of the two link manipulator using the proposed RBF neural network is shown analytically through experimental results without knowing the information of the system in advance.

In this paper, a neural network based approach to visualize performance data of a GSM network is presented. The proposed approach consists of several steps. First, a suitable proportion of measurement data is selected. Then, the selected set of multi-dimensional data is projected into two-dimensional space for visualization purposes with a neural network algorithm called Self-Organizing Map (SOM). Then, the data is clustered and additional visualizations for each data cluster are provided in order to infer the presence of various failure types, their sources and times of occurrence. We apply the proposed approach in the analysis of degradations in signaling and traffic channel capacity of a GSM network.

Tide tables are the method of choice for water level predictions in most coastal regions. In the United States, the National Ocean Service (NOS) uses harmonic analysis and time series of previous water levels to compute tide tables. This method is adequate for most locations along the US coast. However, for many locations along the coast of the Gulf of Mexico, tide tables do not meet NOS criteria. Wind forcing has been recognized as the main variable not included in harmonic analysis. The performance of the tide charts is particularly poor in shallow embayments along the coast of Texas. Recent research at Texas A&M University-Corpus Christi has shown that Artificial Neural Network (ANN) models including input variables such as previous water levels, tidal forecasts, wind speed, wind direction, wind forecasts and barometric pressure can greatly improve water level predictions at several coastal locations including open coast and deep embayment stations. In this paper, the ANN modeling technique was applied for the first time to a shallow embayment, the station of Rockport located near Corpus Christi, Texas. The ANN performance was compared to the NOS tide charts and the persistence model for the years 1997 to 2001. This site was ideal because it is located in a shallow embayment along the Texas coast and there is an 11-year historical record of water levels and meteorological data in the Texas Coastal Ocean Observation Network (TCOON) database. The performance of the ANN model was measured using NOS criteria such as Central Frequency (CF), Maximum Duration of Positive Outliers (MDPO), and Maximum Duration of Negative Outliers (MDNO). The ANN model compared favorably to existing models using these criteria and is the best predictor of future water levels tested.

This work proposes a decision model construction process by extracting the mapping function from the trained neural model. The construction process contains three tasks, namely, data preprocessing, hyperplane extraction, and decision model translation. The data preprocessing uses the correlation coefficient and canonical analysis for projecting the input vector into the canonical feature space. The hyperplane extraction uses the canonical feature space to train the neural networks and extracts the hyperplanes from the trained neural model. The genetic algorithm is used to adjust the slop and reduce the number of hyperplanes. The decision model translation uses the elliptical canonical model to formulate the preliminary decision model. Finally, the genetic algorithm is used again to optimize the canonical decision model.

Our work focuses on: the problem of detecting unusual changes of consumption in mobile phone users, the corresponding building of data structures which represent the recent and historic users’ behaviour bearing in mind the information included in a call, and the complexity of the construction of a function with so many variables where the parameterization is not always known.

The knowledge-driven causal models, implementing some inferential techniques, can prove useful in the assessment of effects of actions in contexts with complex probabilistic chains. Such exploratory tools can thus help in “forevisioning” of future scenarios, but frequently the inverse analysis is required, that is to say, given a desirable future scenario, to discover the “best” set of actions. This paper explores a case of such “future-retrovisioning”, coupling a causal model with a multi-objective genetic algorithm. We show how a genetic algorithm is able to solve the strategy-selection problem, assisting the decision-maker in choosing an adequate strategy within the possibilities offered by the decision space. The paper outlines the general framework underlying an effective knowledge-based decision support system engineered as a software tool.

In this article partitioning of finite element meshes is tackled using colonies of artificial ant–like agents. These agents must restructure the resources in their environment in a manner which corresponds to a good solution of the underlying problem. Standard approaches to these problems use recursive methods in which the final solution is dependent on solutions found at higher levels. For example partitioning into sets is done using recursive bisection which can often provide a partition which is far from optimal [15]. The inherently parallel, distributed nature of the swarm-based paradigm allows us to simultaneously partition into sets. Results show that this approach can be superior in quality when compared to standard methods. Whilst it is marginally slower, the reduced communication cost will greatly reduce the much longer simulation phase of the finite element method. Hence this will outweigh the initial cost of making the partition.

Local search, in either best or first admissible form, generally suffers from poor solution qualities as search cannot be continued beyond locally optimal points. Even multiple start local search strategies can suffer this problem. Meta-heuristic search algorithms, such as simulated annealing and tabu search, implement often computationally expensive optimisation strategies in which local search becomes a subordinate heuristic. To overcome this, a new form of local search is proposed. The Probabilistic Heuristic In Local (PHIL) search meta-strategy uses a recursive branching mechanism in order to overcome local optima. This strategy imposes only a small computational load over and above classical local search. A comparison between PHIL search and ant colony system on benchmark travelling salesman problem instances suggests that the new meta-strategy provides competitive performance. Extensions and improvements to the paradigm are also given.